Upload 123 files

README.md

@@ -1,13 +1,81 @@
----
-title: Pavement Damage Yolo9tr
-emoji: 💻
-colorFrom: pink
-colorTo: gray
-sdk: streamlit
-sdk_version: 1.37.0
-app_file: app.py
-pinned: false
-license: mit
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# YOLO9tr: Yolo9 with partial self attention
+This is the repo for using yolov9 with partial self attention (PSA) \
+This model was developed to be used in pavement damage detection based on YOLO9s Model. 
+### From paper
+YOLO9tr: A Lightweight Model for Pavement Damage Detection Utilizing a Generalized Efficient Layer Aggregation Network and Attention Mechanism [Access](https://arxiv.org/abs/2406.11254)
+
+## Authors
+
+Authors: Dr. Sompote Youwai, Achitaphon Chaiyaphat and Pawarotorn Chaipetch
+
+AI research Group \
+Department of Civil Engineering\
+King Mongkut's University of Technology Thonburi\
+Thailand
+
+
+
+
+
+<p align="center">
+  <img src="https://github.com/Sompote/YOLO9tr/assets/62241733/40d64fae-23ac-46a9-a62b-5f5eb99553a0" alt="Picture11223"/>
+</p>
+<p align="center">
+  <img src="https://github.com/Sompote/YOLO9tr/assets/62241733/851ad8f3-f92a-43af-a481-c7c83b6e6269" alt="Picture11"/>
+</p>
+<p align="center">
+  <img src="https://github.com/Sompote/YOLO9tr/assets/62241733/902aa180-73fd-422e-985f-28a09166f52f" alt="detect_result"/>
+</p>
+
+
+
+
+## Deployment
+
+To deploy this project run
+
+```bash
+git clone https://github.com/Sompote/YOLO9tr
+pip install -r  requirements.txt
+```
+
+
+Reccomend to use weight for [YOLO9s](https://github.com/WongKinYiu/yolov9/releases/download/v0.1/yolov9-s.pt) as initial training
+
+
+### Train with Single GPU
+ ```bash
+ python train_dual.py --workers 8 --device 0 --batch 4 --data '/workspace/6400 images/data.yaml' --img 640 \
+ --cfg models/detect/yolov9tr.yaml  --weights '../yolov9s' --name yolov9-tr --hyp hyp.scratch-high.yaml\
+  --min-items 0 --epochs 200 --close-mosaic 15
+
+```
+
+
+### Train with Dual GPU
+ ```bash
+ torchrun  --nproc_per_node 2 --master_port 9527 train_dual.py  \
+--workers 8 --device 0,1 --sync-bn --batch 30 --data '/workspace/road damage/data.yaml'  \
+--img 640 --cfg models/detect/yolov9tr.yaml --weights '../yolov9s' --name yolov9-c --hyp hyp.scratch-high.yaml \
+--min-items 0 --epochs 200 --close-mosaic 15
+```
+
+
+
+### Evaluation
+[YOLO9tr.pt](https://drive.google.com/file/d/1DtXXICCulTPN8DP4HbVLP3T3sk5BP5HI/view?usp=share_link)
+```
+python val_dual.py --data data/coco.yaml --img 640 --batch 32 --conf 0.001\
+ --iou 0.7 --device 0 --weights './yolov9tr.pt' \
+--save-json --name yolov9_c_640_val
+```
+### Inference
+```
+python detect_dual.py --source './data/images/horses.jpg' --img 640 --device 0 \
+--weights './yolov9tr.pt' --name yolov9_c_640_detect
+```
+The file format of data can be used the same as YOLOv8 in Roboflow
+
+
+
+

YOLO9_run.ipynb

@@ -0,0 +1,1123 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "metadata": {
+    "colab": {
+     "base_uri": "https://localhost:8080/"
+    },
+    "id": "Mw7RoF7eRJEG",
+    "outputId": "1aa46197-e525-4227-b78a-5068873ebd65"
+   },
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Cloning into 'YOLO9_KMUTT'...\n",
+      "remote: Enumerating objects: 182, done.\u001b[K\n",
+      "remote: Counting objects: 100% (182/182), done.\u001b[K\n",
+      "remote: Compressing objects: 100% (132/132), done.\u001b[K\n",
+      "remote: Total 182 (delta 49), reused 179 (delta 46), pack-reused 0\u001b[K\n",
+      "Receiving objects: 100% (182/182), 2.27 MiB | 8.47 MiB/s, done.\n",
+      "Resolving deltas: 100% (49/49), done.\n"
+     ]
+    }
+   ],
+   "source": [
+    "!git clone https://github.com/Sompote/YOLO9_KMUTT"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 9,
+   "metadata": {
+    "colab": {
+     "base_uri": "https://localhost:8080/"
+    },
+    "id": "6KgJYQKfryz2",
+    "outputId": "f359b42d-f9d5-41ac-bbc6-d4c623672e92"
+   },
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "/workspace/YOLO9_KMUTT/yolov9\n"
+     ]
+    }
+   ],
+   "source": [
+    "%cd /workspace/YOLO9_KMUTT/yolov9/"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 10,
+   "metadata": {
+    "colab": {
+     "base_uri": "https://localhost:8080/"
+    },
+    "id": "AnohaOgqq5rQ",
+    "outputId": "d4a78b77-33f3-43dd-af6e-c1ef5e32a8f9"
+   },
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Collecting gitpython (from -r requirements.txt (line 5))\n",
+      "  Downloading GitPython-3.1.43-py3-none-any.whl.metadata (13 kB)\n",
+      "Requirement already satisfied: ipython in /opt/conda/lib/python3.10/site-packages (from -r requirements.txt (line 6)) (8.20.0)\n",
+      "Collecting matplotlib>=3.2.2 (from -r requirements.txt (line 7))\n",
+      "  Downloading matplotlib-3.9.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (11 kB)\n",
+      "Requirement already satisfied: numpy>=1.18.5 in /opt/conda/lib/python3.10/site-packages (from -r requirements.txt (line 8)) (1.26.3)\n",
+      "Collecting opencv-python>=4.1.1 (from -r requirements.txt (line 9))\n",
+      "  Downloading opencv_python-4.10.0.82-cp37-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (20 kB)\n",
+      "Requirement already satisfied: Pillow>=7.1.2 in /opt/conda/lib/python3.10/site-packages (from -r requirements.txt (line 10)) (10.0.1)\n",
+      "Requirement already satisfied: psutil in /opt/conda/lib/python3.10/site-packages (from -r requirements.txt (line 11)) (5.9.0)\n",
+      "Requirement already satisfied: PyYAML>=5.3.1 in /opt/conda/lib/python3.10/site-packages (from -r requirements.txt (line 12)) (6.0.1)\n",
+      "Requirement already satisfied: requests>=2.23.0 in /opt/conda/lib/python3.10/site-packages (from -r requirements.txt (line 13)) (2.31.0)\n",
+      "Collecting scipy>=1.4.1 (from -r requirements.txt (line 14))\n",
+      "  Downloading scipy-1.13.1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (60 kB)\n",
+      "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m60.6/60.6 kB\u001b[0m \u001b[31m9.3 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
+      "\u001b[?25hCollecting thop>=0.1.1 (from -r requirements.txt (line 15))\n",
+      "  Downloading thop-0.1.1.post2209072238-py3-none-any.whl.metadata (2.7 kB)\n",
+      "Requirement already satisfied: torch>=1.7.0 in /opt/conda/lib/python3.10/site-packages (from -r requirements.txt (line 16)) (2.2.0)\n",
+      "Requirement already satisfied: torchvision>=0.8.1 in /opt/conda/lib/python3.10/site-packages (from -r requirements.txt (line 17)) (0.17.0)\n",
+      "Requirement already satisfied: tqdm>=4.64.0 in /opt/conda/lib/python3.10/site-packages (from -r requirements.txt (line 18)) (4.65.0)\n",
+      "Collecting tensorboard>=2.4.1 (from -r requirements.txt (line 22))\n",
+      "  Downloading tensorboard-2.17.0-py3-none-any.whl.metadata (1.6 kB)\n",
+      "Collecting pandas>=1.1.4 (from -r requirements.txt (line 27))\n",
+      "  Downloading pandas-2.2.2-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (19 kB)\n",
+      "Collecting seaborn>=0.11.0 (from -r requirements.txt (line 28))\n",
+      "  Downloading seaborn-0.13.2-py3-none-any.whl.metadata (5.4 kB)\n",
+      "Collecting albumentations>=1.0.3 (from -r requirements.txt (line 46))\n",
+      "  Downloading albumentations-1.4.8-py3-none-any.whl.metadata (37 kB)\n",
+      "Collecting pycocotools>=2.0 (from -r requirements.txt (line 47))\n",
+      "  Downloading pycocotools-2.0.7-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (1.1 kB)\n",
+      "Collecting gitdb<5,>=4.0.1 (from gitpython->-r requirements.txt (line 5))\n",
+      "  Downloading gitdb-4.0.11-py3-none-any.whl.metadata (1.2 kB)\n",
+      "Requirement already satisfied: decorator in /opt/conda/lib/python3.10/site-packages (from ipython->-r requirements.txt (line 6)) (5.1.1)\n",
+      "Requirement already satisfied: jedi>=0.16 in /opt/conda/lib/python3.10/site-packages (from ipython->-r requirements.txt (line 6)) (0.18.1)\n",
+      "Requirement already satisfied: matplotlib-inline in /opt/conda/lib/python3.10/site-packages (from ipython->-r requirements.txt (line 6)) (0.1.6)\n",
+      "Requirement already satisfied: prompt-toolkit<3.1.0,>=3.0.41 in /opt/conda/lib/python3.10/site-packages (from ipython->-r requirements.txt (line 6)) (3.0.43)\n",
+      "Requirement already satisfied: pygments>=2.4.0 in /opt/conda/lib/python3.10/site-packages (from ipython->-r requirements.txt (line 6)) (2.15.1)\n",
+      "Requirement already satisfied: stack-data in /opt/conda/lib/python3.10/site-packages (from ipython->-r requirements.txt (line 6)) (0.2.0)\n",
+      "Requirement already satisfied: traitlets>=5 in /opt/conda/lib/python3.10/site-packages (from ipython->-r requirements.txt (line 6)) (5.7.1)\n",
+      "Requirement already satisfied: exceptiongroup in /opt/conda/lib/python3.10/site-packages (from ipython->-r requirements.txt (line 6)) (1.2.0)\n",
+      "Requirement already satisfied: pexpect>4.3 in /opt/conda/lib/python3.10/site-packages (from ipython->-r requirements.txt (line 6)) (4.8.0)\n",
+      "Collecting contourpy>=1.0.1 (from matplotlib>=3.2.2->-r requirements.txt (line 7))\n",
+      "  Downloading contourpy-1.2.1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (5.8 kB)\n",
+      "Collecting cycler>=0.10 (from matplotlib>=3.2.2->-r requirements.txt (line 7))\n",
+      "  Downloading cycler-0.12.1-py3-none-any.whl.metadata (3.8 kB)\n",
+      "Collecting fonttools>=4.22.0 (from matplotlib>=3.2.2->-r requirements.txt (line 7))\n",
+      "  Downloading fonttools-4.53.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (162 kB)\n",
+      "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m162.2/162.2 kB\u001b[0m \u001b[31m22.0 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
+      "\u001b[?25hCollecting kiwisolver>=1.3.1 (from matplotlib>=3.2.2->-r requirements.txt (line 7))\n",
+      "  Downloading kiwisolver-1.4.5-cp310-cp310-manylinux_2_12_x86_64.manylinux2010_x86_64.whl.metadata (6.4 kB)\n",
+      "Requirement already satisfied: packaging>=20.0 in /opt/conda/lib/python3.10/site-packages (from matplotlib>=3.2.2->-r requirements.txt (line 7)) (23.1)\n",
+      "Collecting pyparsing>=2.3.1 (from matplotlib>=3.2.2->-r requirements.txt (line 7))\n",
+      "  Downloading pyparsing-3.1.2-py3-none-any.whl.metadata (5.1 kB)\n",
+      "Requirement already satisfied: python-dateutil>=2.7 in /opt/conda/lib/python3.10/site-packages (from matplotlib>=3.2.2->-r requirements.txt (line 7)) (2.9.0.post0)\n",
+      "Requirement already satisfied: charset-normalizer<4,>=2 in /opt/conda/lib/python3.10/site-packages (from requests>=2.23.0->-r requirements.txt (line 13)) (2.0.4)\n",
+      "Requirement already satisfied: idna<4,>=2.5 in /opt/conda/lib/python3.10/site-packages (from requests>=2.23.0->-r requirements.txt (line 13)) (3.4)\n",
+      "Requirement already satisfied: urllib3<3,>=1.21.1 in /opt/conda/lib/python3.10/site-packages (from requests>=2.23.0->-r requirements.txt (line 13)) (1.26.18)\n",
+      "Requirement already satisfied: certifi>=2017.4.17 in /opt/conda/lib/python3.10/site-packages (from requests>=2.23.0->-r requirements.txt (line 13)) (2023.11.17)\n",
+      "Requirement already satisfied: filelock in /opt/conda/lib/python3.10/site-packages (from torch>=1.7.0->-r requirements.txt (line 16)) (3.13.1)\n",
+      "Requirement already satisfied: typing-extensions>=4.8.0 in /opt/conda/lib/python3.10/site-packages (from torch>=1.7.0->-r requirements.txt (line 16)) (4.9.0)\n",
+      "Requirement already satisfied: sympy in /opt/conda/lib/python3.10/site-packages (from torch>=1.7.0->-r requirements.txt (line 16)) (1.12)\n",
+      "Requirement already satisfied: networkx in /opt/conda/lib/python3.10/site-packages (from torch>=1.7.0->-r requirements.txt (line 16)) (3.1)\n",
+      "Requirement already satisfied: jinja2 in /opt/conda/lib/python3.10/site-packages (from torch>=1.7.0->-r requirements.txt (line 16)) (3.1.2)\n",
+      "Requirement already satisfied: fsspec in /opt/conda/lib/python3.10/site-packages (from torch>=1.7.0->-r requirements.txt (line 16)) (2023.12.2)\n",
+      "Collecting absl-py>=0.4 (from tensorboard>=2.4.1->-r requirements.txt (line 22))\n",
+      "  Downloading absl_py-2.1.0-py3-none-any.whl.metadata (2.3 kB)\n",
+      "Collecting grpcio>=1.48.2 (from tensorboard>=2.4.1->-r requirements.txt (line 22))\n",
+      "  Downloading grpcio-1.64.1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (3.3 kB)\n",
+      "Collecting markdown>=2.6.8 (from tensorboard>=2.4.1->-r requirements.txt (line 22))\n",
+      "  Downloading Markdown-3.6-py3-none-any.whl.metadata (7.0 kB)\n",
+      "Collecting protobuf!=4.24.0,<5.0.0,>=3.19.6 (from tensorboard>=2.4.1->-r requirements.txt (line 22))\n",
+      "  Downloading protobuf-4.25.3-cp37-abi3-manylinux2014_x86_64.whl.metadata (541 bytes)\n",
+      "Requirement already satisfied: setuptools>=41.0.0 in /opt/conda/lib/python3.10/site-packages (from tensorboard>=2.4.1->-r requirements.txt (line 22)) (68.2.2)\n",
+      "Requirement already satisfied: six>1.9 in /opt/conda/lib/python3.10/site-packages (from tensorboard>=2.4.1->-r requirements.txt (line 22)) (1.16.0)\n",
+      "Collecting tensorboard-data-server<0.8.0,>=0.7.0 (from tensorboard>=2.4.1->-r requirements.txt (line 22))\n",
+      "  Downloading tensorboard_data_server-0.7.2-py3-none-manylinux_2_31_x86_64.whl.metadata (1.1 kB)\n",
+      "Collecting werkzeug>=1.0.1 (from tensorboard>=2.4.1->-r requirements.txt (line 22))\n",
+      "  Downloading werkzeug-3.0.3-py3-none-any.whl.metadata (3.7 kB)\n",
+      "Requirement already satisfied: pytz>=2020.1 in /opt/conda/lib/python3.10/site-packages (from pandas>=1.1.4->-r requirements.txt (line 27)) (2023.3.post1)\n",
+      "Collecting tzdata>=2022.7 (from pandas>=1.1.4->-r requirements.txt (line 27))\n",
+      "  Downloading tzdata-2024.1-py2.py3-none-any.whl.metadata (1.4 kB)\n",
+      "Collecting scikit-image>=0.21.0 (from albumentations>=1.0.3->-r requirements.txt (line 46))\n",
+      "  Downloading scikit_image-0.23.2-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (14 kB)\n",
+      "Collecting scikit-learn>=1.3.2 (from albumentations>=1.0.3->-r requirements.txt (line 46))\n",
+      "  Downloading scikit_learn-1.5.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (11 kB)\n",
+      "Collecting pydantic>=2.7.0 (from albumentations>=1.0.3->-r requirements.txt (line 46))\n",
+      "  Downloading pydantic-2.7.3-py3-none-any.whl.metadata (108 kB)\n",
+      "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m109.0/109.0 kB\u001b[0m \u001b[31m31.4 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
+      "\u001b[?25hCollecting albucore>=0.0.4 (from albumentations>=1.0.3->-r requirements.txt (line 46))\n",
+      "  Downloading albucore-0.0.9-py3-none-any.whl.metadata (3.1 kB)\n",
+      "Collecting opencv-python-headless>=4.9.0.80 (from albumentations>=1.0.3->-r requirements.txt (line 46))\n",
+      "  Downloading opencv_python_headless-4.10.0.82-cp37-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (20 kB)\n",
+      "Requirement already satisfied: tomli>=2.0.1 in /opt/conda/lib/python3.10/site-packages (from albucore>=0.0.4->albumentations>=1.0.3->-r requirements.txt (line 46)) (2.0.1)\n",
+      "Collecting smmap<6,>=3.0.1 (from gitdb<5,>=4.0.1->gitpython->-r requirements.txt (line 5))\n",
+      "  Downloading smmap-5.0.1-py3-none-any.whl.metadata (4.3 kB)\n",
+      "Requirement already satisfied: parso<0.9.0,>=0.8.0 in /opt/conda/lib/python3.10/site-packages (from jedi>=0.16->ipython->-r requirements.txt (line 6)) (0.8.3)\n",
+      "Requirement already satisfied: ptyprocess>=0.5 in /opt/conda/lib/python3.10/site-packages (from pexpect>4.3->ipython->-r requirements.txt (line 6)) (0.7.0)\n",
+      "Requirement already satisfied: wcwidth in /opt/conda/lib/python3.10/site-packages (from prompt-toolkit<3.1.0,>=3.0.41->ipython->-r requirements.txt (line 6)) (0.2.5)\n",
+      "Collecting annotated-types>=0.4.0 (from pydantic>=2.7.0->albumentations>=1.0.3->-r requirements.txt (line 46))\n",
+      "  Downloading annotated_types-0.7.0-py3-none-any.whl.metadata (15 kB)\n",
+      "Collecting pydantic-core==2.18.4 (from pydantic>=2.7.0->albumentations>=1.0.3->-r requirements.txt (line 46))\n",
+      "  Downloading pydantic_core-2.18.4-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (6.5 kB)\n",
+      "Collecting imageio>=2.33 (from scikit-image>=0.21.0->albumentations>=1.0.3->-r requirements.txt (line 46))\n",
+      "  Downloading imageio-2.34.1-py3-none-any.whl.metadata (4.9 kB)\n",
+      "Collecting tifffile>=2022.8.12 (from scikit-image>=0.21.0->albumentations>=1.0.3->-r requirements.txt (line 46))\n",
+      "  Downloading tifffile-2024.5.22-py3-none-any.whl.metadata (30 kB)\n",
+      "Collecting lazy-loader>=0.4 (from scikit-image>=0.21.0->albumentations>=1.0.3->-r requirements.txt (line 46))\n",
+      "  Downloading lazy_loader-0.4-py3-none-any.whl.metadata (7.6 kB)\n",
+      "Collecting joblib>=1.2.0 (from scikit-learn>=1.3.2->albumentations>=1.0.3->-r requirements.txt (line 46))\n",
+      "  Downloading joblib-1.4.2-py3-none-any.whl.metadata (5.4 kB)\n",
+      "Collecting threadpoolctl>=3.1.0 (from scikit-learn>=1.3.2->albumentations>=1.0.3->-r requirements.txt (line 46))\n",
+      "  Downloading threadpoolctl-3.5.0-py3-none-any.whl.metadata (13 kB)\n",
+      "Requirement already satisfied: MarkupSafe>=2.1.1 in /opt/conda/lib/python3.10/site-packages (from werkzeug>=1.0.1->tensorboard>=2.4.1->-r requirements.txt (line 22)) (2.1.3)\n",
+      "Requirement already satisfied: executing in /opt/conda/lib/python3.10/site-packages (from stack-data->ipython->-r requirements.txt (line 6)) (0.8.3)\n",
+      "Requirement already satisfied: asttokens in /opt/conda/lib/python3.10/site-packages (from stack-data->ipython->-r requirements.txt (line 6)) (2.0.5)\n",
+      "Requirement already satisfied: pure-eval in /opt/conda/lib/python3.10/site-packages (from stack-data->ipython->-r requirements.txt (line 6)) (0.2.2)\n",
+      "Requirement already satisfied: mpmath>=0.19 in /opt/conda/lib/python3.10/site-packages (from sympy->torch>=1.7.0->-r requirements.txt (line 16)) (1.3.0)\n",
+      "Downloading GitPython-3.1.43-py3-none-any.whl (207 kB)\n",
+      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m207.3/207.3 kB\u001b[0m \u001b[31m35.9 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
+      "\u001b[?25hDownloading matplotlib-3.9.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (8.3 MB)\n",
+      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m8.3/8.3 MB\u001b[0m \u001b[31m127.3 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0ma \u001b[36m0:00:01\u001b[0m\n",
+      "\u001b[?25hDownloading opencv_python-4.10.0.82-cp37-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (62.5 MB)\n",
+      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m62.5/62.5 MB\u001b[0m \u001b[31m89.1 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m:00:01\u001b[0m00:01\u001b[0m\n",
+      "\u001b[?25hDownloading scipy-1.13.1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (38.6 MB)\n",
+      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m38.6/38.6 MB\u001b[0m \u001b[31m100.9 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m00:01\u001b[0m00:01\u001b[0m\n",
+      "\u001b[?25hDownloading thop-0.1.1.post2209072238-py3-none-any.whl (15 kB)\n",
+      "Downloading tensorboard-2.17.0-py3-none-any.whl (5.5 MB)\n",
+      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m5.5/5.5 MB\u001b[0m \u001b[31m148.0 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m00:01\u001b[0m\n",
+      "\u001b[?25hDownloading pandas-2.2.2-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (13.0 MB)\n",
+      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m13.0/13.0 MB\u001b[0m \u001b[31m151.9 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m00:01\u001b[0m0:01\u001b[0m\n",
+      "\u001b[?25hDownloading seaborn-0.13.2-py3-none-any.whl (294 kB)\n",
+      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m294.9/294.9 kB\u001b[0m \u001b[31m50.5 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
+      "\u001b[?25hDownloading albumentations-1.4.8-py3-none-any.whl (156 kB)\n",
+      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m156.8/156.8 kB\u001b[0m \u001b[31m35.6 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
+      "\u001b[?25hDownloading pycocotools-2.0.7-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (426 kB)\n",
+      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m426.2/426.2 kB\u001b[0m \u001b[31m61.1 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
+      "\u001b[?25hDownloading absl_py-2.1.0-py3-none-any.whl (133 kB)\n",
+      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m133.7/133.7 kB\u001b[0m \u001b[31m33.2 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
+      "\u001b[?25hDownloading albucore-0.0.9-py3-none-any.whl (7.9 kB)\n",
+      "Downloading contourpy-1.2.1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (305 kB)\n",
+      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m305.2/305.2 kB\u001b[0m \u001b[31m50.4 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
+      "\u001b[?25hDownloading cycler-0.12.1-py3-none-any.whl (8.3 kB)\n",
+      "Downloading fonttools-4.53.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (4.6 MB)\n",
+      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m4.6/4.6 MB\u001b[0m \u001b[31m141.5 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
+      "\u001b[?25hDownloading gitdb-4.0.11-py3-none-any.whl (62 kB)\n",
+      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m62.7/62.7 kB\u001b[0m \u001b[31m13.0 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
+      "\u001b[?25hDownloading grpcio-1.64.1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (5.6 MB)\n",
+      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m5.6/5.6 MB\u001b[0m \u001b[31m140.6 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m00:01\u001b[0m\n",
+      "\u001b[?25hDownloading kiwisolver-1.4.5-cp310-cp310-manylinux_2_12_x86_64.manylinux2010_x86_64.whl (1.6 MB)\n",
+      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m1.6/1.6 MB\u001b[0m \u001b[31m115.4 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
+      "\u001b[?25hDownloading Markdown-3.6-py3-none-any.whl (105 kB)\n",
+      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m105.4/105.4 kB\u001b[0m \u001b[31m20.2 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
+      "\u001b[?25hDownloading opencv_python_headless-4.10.0.82-cp37-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (49.9 MB)\n",
+      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m49.9/49.9 MB\u001b[0m \u001b[31m93.7 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m:00:01\u001b[0m00:01\u001b[0m\n",
+      "\u001b[?25hDownloading protobuf-4.25.3-cp37-abi3-manylinux2014_x86_64.whl (294 kB)\n",
+      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m294.6/294.6 kB\u001b[0m \u001b[31m49.4 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
+      "\u001b[?25hDownloading pydantic-2.7.3-py3-none-any.whl (409 kB)\n",
+      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m409.6/409.6 kB\u001b[0m \u001b[31m70.2 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
+      "\u001b[?25hDownloading pydantic_core-2.18.4-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (2.0 MB)\n",
+      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m2.0/2.0 MB\u001b[0m \u001b[31m117.8 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
+      "\u001b[?25hDownloading pyparsing-3.1.2-py3-none-any.whl (103 kB)\n",
+      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m103.2/103.2 kB\u001b[0m \u001b[31m21.2 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
+      "\u001b[?25hDownloading scikit_image-0.23.2-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (14.7 MB)\n",
+      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m14.7/14.7 MB\u001b[0m \u001b[31m145.4 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0ma \u001b[36m0:00:01\u001b[0m\n",
+      "\u001b[?25hDownloading scikit_learn-1.5.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (13.3 MB)\n",
+      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m13.3/13.3 MB\u001b[0m \u001b[31m137.8 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m \u001b[36m0:00:01\u001b[0m\n",
+      "\u001b[?25hDownloading tensorboard_data_server-0.7.2-py3-none-manylinux_2_31_x86_64.whl (6.6 MB)\n",
+      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m6.6/6.6 MB\u001b[0m \u001b[31m96.1 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0mta \u001b[36m0:00:01\u001b[0m\n",
+      "\u001b[?25hDownloading tzdata-2024.1-py2.py3-none-any.whl (345 kB)\n",
+      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m345.4/345.4 kB\u001b[0m \u001b[31m51.1 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
+      "\u001b[?25hDownloading werkzeug-3.0.3-py3-none-any.whl (227 kB)\n",
+      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m227.3/227.3 kB\u001b[0m \u001b[31m42.1 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
+      "\u001b[?25hDownloading annotated_types-0.7.0-py3-none-any.whl (13 kB)\n",
+      "Downloading imageio-2.34.1-py3-none-any.whl (313 kB)\n",
+      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m313.5/313.5 kB\u001b[0m \u001b[31m56.0 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
+      "\u001b[?25hDownloading joblib-1.4.2-py3-none-any.whl (301 kB)\n",
+      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m301.8/301.8 kB\u001b[0m \u001b[31m49.3 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
+      "\u001b[?25hDownloading lazy_loader-0.4-py3-none-any.whl (12 kB)\n",
+      "Downloading smmap-5.0.1-py3-none-any.whl (24 kB)\n",
+      "Downloading threadpoolctl-3.5.0-py3-none-any.whl (18 kB)\n",
+      "Downloading tifffile-2024.5.22-py3-none-any.whl (225 kB)\n",
+      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m225.5/225.5 kB\u001b[0m \u001b[31m49.4 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
+      "\u001b[?25hInstalling collected packages: werkzeug, tzdata, tifffile, threadpoolctl, tensorboard-data-server, smmap, scipy, pyparsing, pydantic-core, protobuf, opencv-python-headless, opencv-python, markdown, lazy-loader, kiwisolver, joblib, imageio, grpcio, fonttools, cycler, contourpy, annotated-types, absl-py, tensorboard, scikit-learn, scikit-image, pydantic, pandas, matplotlib, gitdb, albucore, thop, seaborn, pycocotools, gitpython, albumentations\n",
+      "Successfully installed absl-py-2.1.0 albucore-0.0.9 albumentations-1.4.8 annotated-types-0.7.0 contourpy-1.2.1 cycler-0.12.1 fonttools-4.53.0 gitdb-4.0.11 gitpython-3.1.43 grpcio-1.64.1 imageio-2.34.1 joblib-1.4.2 kiwisolver-1.4.5 lazy-loader-0.4 markdown-3.6 matplotlib-3.9.0 opencv-python-4.10.0.82 opencv-python-headless-4.10.0.82 pandas-2.2.2 protobuf-4.25.3 pycocotools-2.0.7 pydantic-2.7.3 pydantic-core-2.18.4 pyparsing-3.1.2 scikit-image-0.23.2 scikit-learn-1.5.0 scipy-1.13.1 seaborn-0.13.2 smmap-5.0.1 tensorboard-2.17.0 tensorboard-data-server-0.7.2 thop-0.1.1.post2209072238 threadpoolctl-3.5.0 tifffile-2024.5.22 tzdata-2024.1 werkzeug-3.0.3\n",
+      "\u001b[33mWARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv\u001b[0m\u001b[33m\n",
+      "\u001b[0m"
+     ]
+    }
+   ],
+   "source": [
+    "!pip install -r  requirements.txt"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 11,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Hit:1 https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64  InRelease\n",
+      "Hit:2 http://archive.ubuntu.com/ubuntu jammy InRelease                     \n",
+      "Hit:3 http://security.ubuntu.com/ubuntu jammy-security InRelease\n",
+      "Hit:4 http://archive.ubuntu.com/ubuntu jammy-updates InRelease\n",
+      "Hit:5 http://archive.ubuntu.com/ubuntu jammy-backports InRelease\n",
+      "Reading package lists... Done\n",
+      "Reading package lists... Done\n",
+      "Building dependency tree... Done\n",
+      "Reading state information... Done\n",
+      "The following additional packages will be installed:\n",
+      "  libdrm-amdgpu1 libdrm-common libdrm-intel1 libdrm-nouveau2 libdrm-radeon1\n",
+      "  libdrm2 libgl1 libgl1-amber-dri libgl1-mesa-dri libglapi-mesa libglvnd0\n",
+      "  libglx-mesa0 libglx0 libllvm15 libpciaccess0 libsensors-config libsensors5\n",
+      "  libx11-6 libx11-data libx11-xcb1 libxau6 libxcb-dri2-0 libxcb-dri3-0\n",
+      "  libxcb-glx0 libxcb-present0 libxcb-randr0 libxcb-shm0 libxcb-sync1\n",
+      "  libxcb-xfixes0 libxcb1 libxdmcp6 libxext6 libxfixes3 libxshmfence1\n",
+      "  libxxf86vm1\n",
+      "Suggested packages:\n",
+      "  pciutils lm-sensors\n",
+      "The following NEW packages will be installed:\n",
+      "  libdrm-amdgpu1 libdrm-common libdrm-intel1 libdrm-nouveau2 libdrm-radeon1\n",
+      "  libdrm2 libgl1 libgl1-amber-dri libgl1-mesa-dri libgl1-mesa-glx\n",
+      "  libglapi-mesa libglvnd0 libglx-mesa0 libglx0 libllvm15 libpciaccess0\n",
+      "  libsensors-config libsensors5 libx11-6 libx11-data libx11-xcb1 libxau6\n",
+      "  libxcb-dri2-0 libxcb-dri3-0 libxcb-glx0 libxcb-present0 libxcb-randr0\n",
+      "  libxcb-shm0 libxcb-sync1 libxcb-xfixes0 libxcb1 libxdmcp6 libxext6\n",
+      "  libxfixes3 libxshmfence1 libxxf86vm1\n",
+      "0 upgraded, 36 newly installed, 0 to remove and 56 not upgraded.\n",
+      "Need to get 40.1 MB of archives.\n",
+      "After this operation, 173 MB of additional disk space will be used.\n",
+      "Get:1 http://archive.ubuntu.com/ubuntu jammy-updates/main amd64 libdrm-common all 2.4.113-2~ubuntu0.22.04.1 [5450 B]\n",
+      "Get:2 http://archive.ubuntu.com/ubuntu jammy-updates/main amd64 libdrm2 amd64 2.4.113-2~ubuntu0.22.04.1 [38.1 kB]\n",
+      "Get:3 http://archive.ubuntu.com/ubuntu jammy/main amd64 libxau6 amd64 1:1.0.9-1build5 [7634 B]\n",
+      "Get:4 http://archive.ubuntu.com/ubuntu jammy/main amd64 libxdmcp6 amd64 1:1.1.3-0ubuntu5 [10.9 kB]\n",
+      "Get:5 http://archive.ubuntu.com/ubuntu jammy/main amd64 libxcb1 amd64 1.14-3ubuntu3 [49.0 kB]\n",
+      "Get:6 http://archive.ubuntu.com/ubuntu jammy-updates/main amd64 libx11-data all 2:1.7.5-1ubuntu0.3 [120 kB]\n",
+      "Get:7 http://archive.ubuntu.com/ubuntu jammy-updates/main amd64 libx11-6 amd64 2:1.7.5-1ubuntu0.3 [667 kB]\n",
+      "Get:8 http://archive.ubuntu.com/ubuntu jammy/main amd64 libxext6 amd64 2:1.3.4-1build1 [31.8 kB]\n",
+      "Get:9 http://archive.ubuntu.com/ubuntu jammy-updates/main amd64 libdrm-amdgpu1 amd64 2.4.113-2~ubuntu0.22.04.1 [19.9 kB]\n",
+      "Get:10 http://archive.ubuntu.com/ubuntu jammy/main amd64 libpciaccess0 amd64 0.16-3 [19.1 kB]\n",
+      "Get:11 http://archive.ubuntu.com/ubuntu jammy-updates/main amd64 libdrm-intel1 amd64 2.4.113-2~ubuntu0.22.04.1 [66.7 kB]\n",
+      "Get:12 http://archive.ubuntu.com/ubuntu jammy-updates/main amd64 libdrm-nouveau2 amd64 2.4.113-2~ubuntu0.22.04.1 [17.5 kB]\n",
+      "Get:13 http://archive.ubuntu.com/ubuntu jammy-updates/main amd64 libdrm-radeon1 amd64 2.4.113-2~ubuntu0.22.04.1 [21.6 kB]\n",
+      "Get:14 http://archive.ubuntu.com/ubuntu jammy-updates/main amd64 libglapi-mesa amd64 23.2.1-1ubuntu3.1~22.04.2 [37.1 kB]\n",
+      "Get:15 http://archive.ubuntu.com/ubuntu jammy-updates/main amd64 libgl1-amber-dri amd64 21.3.9-0ubuntu1~22.04.1 [4218 kB]\n",
+      "Get:16 http://archive.ubuntu.com/ubuntu jammy-updates/main amd64 libllvm15 amd64 1:15.0.7-0ubuntu0.22.04.3 [25.4 MB]\n",
+      "Get:17 http://archive.ubuntu.com/ubuntu jammy/main amd64 libsensors-config all 1:3.6.0-7ubuntu1 [5274 B]\n",
+      "Get:18 http://archive.ubuntu.com/ubuntu jammy/main amd64 libsensors5 amd64 1:3.6.0-7ubuntu1 [26.3 kB]\n",
+      "Get:19 http://archive.ubuntu.com/ubuntu jammy/main amd64 libxcb-dri3-0 amd64 1.14-3ubuntu3 [6968 B]\n",
+      "Get:20 http://archive.ubuntu.com/ubuntu jammy-updates/main amd64 libgl1-mesa-dri amd64 23.2.1-1ubuntu3.1~22.04.2 [8860 kB]\n",
+      "Get:21 http://archive.ubuntu.com/ubuntu jammy/main amd64 libglvnd0 amd64 1.4.0-1 [73.6 kB]\n",
+      "Get:22 http://archive.ubuntu.com/ubuntu jammy-updates/main amd64 libx11-xcb1 amd64 2:1.7.5-1ubuntu0.3 [7802 B]\n",
+      "Get:23 http://archive.ubuntu.com/ubuntu jammy/main amd64 libxcb-dri2-0 amd64 1.14-3ubuntu3 [7206 B]\n",
+      "Get:24 http://archive.ubuntu.com/ubuntu jammy/main amd64 libxcb-glx0 amd64 1.14-3ubuntu3 [25.9 kB]\n",
+      "Get:25 http://archive.ubuntu.com/ubuntu jammy/main amd64 libxcb-present0 amd64 1.14-3ubuntu3 [5734 B]\n",
+      "Get:26 http://archive.ubuntu.com/ubuntu jammy/main amd64 libxcb-randr0 amd64 1.14-3ubuntu3 [18.3 kB]\n",
+      "Get:27 http://archive.ubuntu.com/ubuntu jammy/main amd64 libxcb-shm0 amd64 1.14-3ubuntu3 [5780 B]\n",
+      "Get:28 http://archive.ubuntu.com/ubuntu jammy/main amd64 libxcb-sync1 amd64 1.14-3ubuntu3 [9416 B]\n",
+      "Get:29 http://archive.ubuntu.com/ubuntu jammy/main amd64 libxcb-xfixes0 amd64 1.14-3ubuntu3 [9996 B]\n",
+      "Get:30 http://archive.ubuntu.com/ubuntu jammy/main amd64 libxfixes3 amd64 1:6.0.0-1 [11.7 kB]\n",
+      "Get:31 http://archive.ubuntu.com/ubuntu jammy/main amd64 libxshmfence1 amd64 1.3-1build4 [5394 B]\n",
+      "Get:32 http://archive.ubuntu.com/ubuntu jammy/main amd64 libxxf86vm1 amd64 1:1.1.4-1build3 [10.4 kB]\n",
+      "Get:33 http://archive.ubuntu.com/ubuntu jammy-updates/main amd64 libglx-mesa0 amd64 23.2.1-1ubuntu3.1~22.04.2 [158 kB]\n",
+      "Get:34 http://archive.ubuntu.com/ubuntu jammy/main amd64 libglx0 amd64 1.4.0-1 [41.0 kB]\n",
+      "Get:35 http://archive.ubuntu.com/ubuntu jammy/main amd64 libgl1 amd64 1.4.0-1 [110 kB]\n",
+      "Get:36 http://archive.ubuntu.com/ubuntu jammy-updates/universe amd64 libgl1-mesa-glx amd64 23.0.4-0ubuntu1~22.04.1 [5584 B]\n",
+      "Fetched 40.1 MB in 3s (14.7 MB/s)\n",
+      "debconf: delaying package configuration, since apt-utils is not installed\n",
+      "Selecting previously unselected package libdrm-common.\n",
+      "(Reading database ... 22469 files and directories currently installed.)\n",
+      "Preparing to unpack .../00-libdrm-common_2.4.113-2~ubuntu0.22.04.1_all.deb ...\n",
+      "Unpacking libdrm-common (2.4.113-2~ubuntu0.22.04.1) ...\n",
+      "Selecting previously unselected package libdrm2:amd64.\n",
+      "Preparing to unpack .../01-libdrm2_2.4.113-2~ubuntu0.22.04.1_amd64.deb ...\n",
+      "Unpacking libdrm2:amd64 (2.4.113-2~ubuntu0.22.04.1) ...\n",
+      "Selecting previously unselected package libxau6:amd64.\n",
+      "Preparing to unpack .../02-libxau6_1%3a1.0.9-1build5_amd64.deb ...\n",
+      "Unpacking libxau6:amd64 (1:1.0.9-1build5) ...\n",
+      "Selecting previously unselected package libxdmcp6:amd64.\n",
+      "Preparing to unpack .../03-libxdmcp6_1%3a1.1.3-0ubuntu5_amd64.deb ...\n",
+      "Unpacking libxdmcp6:amd64 (1:1.1.3-0ubuntu5) ...\n",
+      "Selecting previously unselected package libxcb1:amd64.\n",
+      "Preparing to unpack .../04-libxcb1_1.14-3ubuntu3_amd64.deb ...\n",
+      "Unpacking libxcb1:amd64 (1.14-3ubuntu3) ...\n",
+      "Selecting previously unselected package libx11-data.\n",
+      "Preparing to unpack .../05-libx11-data_2%3a1.7.5-1ubuntu0.3_all.deb ...\n",
+      "Unpacking libx11-data (2:1.7.5-1ubuntu0.3) ...\n",
+      "Selecting previously unselected package libx11-6:amd64.\n",
+      "Preparing to unpack .../06-libx11-6_2%3a1.7.5-1ubuntu0.3_amd64.deb ...\n",
+      "Unpacking libx11-6:amd64 (2:1.7.5-1ubuntu0.3) ...\n",
+      "Selecting previously unselected package libxext6:amd64.\n",
+      "Preparing to unpack .../07-libxext6_2%3a1.3.4-1build1_amd64.deb ...\n",
+      "Unpacking libxext6:amd64 (2:1.3.4-1build1) ...\n",
+      "Selecting previously unselected package libdrm-amdgpu1:amd64.\n",
+      "Preparing to unpack .../08-libdrm-amdgpu1_2.4.113-2~ubuntu0.22.04.1_amd64.deb ...\n",
+      "Unpacking libdrm-amdgpu1:amd64 (2.4.113-2~ubuntu0.22.04.1) ...\n",
+      "Selecting previously unselected package libpciaccess0:amd64.\n",
+      "Preparing to unpack .../09-libpciaccess0_0.16-3_amd64.deb ...\n",
+      "Unpacking libpciaccess0:amd64 (0.16-3) ...\n",
+      "Selecting previously unselected package libdrm-intel1:amd64.\n",
+      "Preparing to unpack .../10-libdrm-intel1_2.4.113-2~ubuntu0.22.04.1_amd64.deb ...\n",
+      "Unpacking libdrm-intel1:amd64 (2.4.113-2~ubuntu0.22.04.1) ...\n",
+      "Selecting previously unselected package libdrm-nouveau2:amd64.\n",
+      "Preparing to unpack .../11-libdrm-nouveau2_2.4.113-2~ubuntu0.22.04.1_amd64.deb ...\n",
+      "Unpacking libdrm-nouveau2:amd64 (2.4.113-2~ubuntu0.22.04.1) ...\n",
+      "Selecting previously unselected package libdrm-radeon1:amd64.\n",
+      "Preparing to unpack .../12-libdrm-radeon1_2.4.113-2~ubuntu0.22.04.1_amd64.deb ...\n",
+      "Unpacking libdrm-radeon1:amd64 (2.4.113-2~ubuntu0.22.04.1) ...\n",
+      "Selecting previously unselected package libglapi-mesa:amd64.\n",
+      "Preparing to unpack .../13-libglapi-mesa_23.2.1-1ubuntu3.1~22.04.2_amd64.deb ...\n",
+      "Unpacking libglapi-mesa:amd64 (23.2.1-1ubuntu3.1~22.04.2) ...\n",
+      "Selecting previously unselected package libgl1-amber-dri:amd64.\n",
+      "Preparing to unpack .../14-libgl1-amber-dri_21.3.9-0ubuntu1~22.04.1_amd64.deb ...\n",
+      "Unpacking libgl1-amber-dri:amd64 (21.3.9-0ubuntu1~22.04.1) ...\n",
+      "Selecting previously unselected package libllvm15:amd64.\n",
+      "Preparing to unpack .../15-libllvm15_1%3a15.0.7-0ubuntu0.22.04.3_amd64.deb ...\n",
+      "Unpacking libllvm15:amd64 (1:15.0.7-0ubuntu0.22.04.3) ...\n",
+      "Selecting previously unselected package libsensors-config.\n",
+      "Preparing to unpack .../16-libsensors-config_1%3a3.6.0-7ubuntu1_all.deb ...\n",
+      "Unpacking libsensors-config (1:3.6.0-7ubuntu1) ...\n",
+      "Selecting previously unselected package libsensors5:amd64.\n",
+      "Preparing to unpack .../17-libsensors5_1%3a3.6.0-7ubuntu1_amd64.deb ...\n",
+      "Unpacking libsensors5:amd64 (1:3.6.0-7ubuntu1) ...\n",
+      "Selecting previously unselected package libxcb-dri3-0:amd64.\n",
+      "Preparing to unpack .../18-libxcb-dri3-0_1.14-3ubuntu3_amd64.deb ...\n",
+      "Unpacking libxcb-dri3-0:amd64 (1.14-3ubuntu3) ...\n",
+      "Selecting previously unselected package libgl1-mesa-dri:amd64.\n",
+      "Preparing to unpack .../19-libgl1-mesa-dri_23.2.1-1ubuntu3.1~22.04.2_amd64.deb ...\n",
+      "Unpacking libgl1-mesa-dri:amd64 (23.2.1-1ubuntu3.1~22.04.2) ...\n",
+      "Selecting previously unselected package libglvnd0:amd64.\n",
+      "Preparing to unpack .../20-libglvnd0_1.4.0-1_amd64.deb ...\n",
+      "Unpacking libglvnd0:amd64 (1.4.0-1) ...\n",
+      "Selecting previously unselected package libx11-xcb1:amd64.\n",
+      "Preparing to unpack .../21-libx11-xcb1_2%3a1.7.5-1ubuntu0.3_amd64.deb ...\n",
+      "Unpacking libx11-xcb1:amd64 (2:1.7.5-1ubuntu0.3) ...\n",
+      "Selecting previously unselected package libxcb-dri2-0:amd64.\n",
+      "Preparing to unpack .../22-libxcb-dri2-0_1.14-3ubuntu3_amd64.deb ...\n",
+      "Unpacking libxcb-dri2-0:amd64 (1.14-3ubuntu3) ...\n",
+      "Selecting previously unselected package libxcb-glx0:amd64.\n",
+      "Preparing to unpack .../23-libxcb-glx0_1.14-3ubuntu3_amd64.deb ...\n",
+      "Unpacking libxcb-glx0:amd64 (1.14-3ubuntu3) ...\n",
+      "Selecting previously unselected package libxcb-present0:amd64.\n",
+      "Preparing to unpack .../24-libxcb-present0_1.14-3ubuntu3_amd64.deb ...\n",
+      "Unpacking libxcb-present0:amd64 (1.14-3ubuntu3) ...\n",
+      "Selecting previously unselected package libxcb-randr0:amd64.\n",
+      "Preparing to unpack .../25-libxcb-randr0_1.14-3ubuntu3_amd64.deb ...\n",
+      "Unpacking libxcb-randr0:amd64 (1.14-3ubuntu3) ...\n",
+      "Selecting previously unselected package libxcb-shm0:amd64.\n",
+      "Preparing to unpack .../26-libxcb-shm0_1.14-3ubuntu3_amd64.deb ...\n",
+      "Unpacking libxcb-shm0:amd64 (1.14-3ubuntu3) ...\n",
+      "Selecting previously unselected package libxcb-sync1:amd64.\n",
+      "Preparing to unpack .../27-libxcb-sync1_1.14-3ubuntu3_amd64.deb ...\n",
+      "Unpacking libxcb-sync1:amd64 (1.14-3ubuntu3) ...\n",
+      "Selecting previously unselected package libxcb-xfixes0:amd64.\n",
+      "Preparing to unpack .../28-libxcb-xfixes0_1.14-3ubuntu3_amd64.deb ...\n",
+      "Unpacking libxcb-xfixes0:amd64 (1.14-3ubuntu3) ...\n",
+      "Selecting previously unselected package libxfixes3:amd64.\n",
+      "Preparing to unpack .../29-libxfixes3_1%3a6.0.0-1_amd64.deb ...\n",
+      "Unpacking libxfixes3:amd64 (1:6.0.0-1) ...\n",
+      "Selecting previously unselected package libxshmfence1:amd64.\n",
+      "Preparing to unpack .../30-libxshmfence1_1.3-1build4_amd64.deb ...\n",
+      "Unpacking libxshmfence1:amd64 (1.3-1build4) ...\n",
+      "Selecting previously unselected package libxxf86vm1:amd64.\n",
+      "Preparing to unpack .../31-libxxf86vm1_1%3a1.1.4-1build3_amd64.deb ...\n",
+      "Unpacking libxxf86vm1:amd64 (1:1.1.4-1build3) ...\n",
+      "Selecting previously unselected package libglx-mesa0:amd64.\n",
+      "Preparing to unpack .../32-libglx-mesa0_23.2.1-1ubuntu3.1~22.04.2_amd64.deb ...\n",
+      "Unpacking libglx-mesa0:amd64 (23.2.1-1ubuntu3.1~22.04.2) ...\n",
+      "Selecting previously unselected package libglx0:amd64.\n",
+      "Preparing to unpack .../33-libglx0_1.4.0-1_amd64.deb ...\n",
+      "Unpacking libglx0:amd64 (1.4.0-1) ...\n",
+      "Selecting previously unselected package libgl1:amd64.\n",
+      "Preparing to unpack .../34-libgl1_1.4.0-1_amd64.deb ...\n",
+      "Unpacking libgl1:amd64 (1.4.0-1) ...\n",
+      "Selecting previously unselected package libgl1-mesa-glx:amd64.\n",
+      "Preparing to unpack .../35-libgl1-mesa-glx_23.0.4-0ubuntu1~22.04.1_amd64.deb ...\n",
+      "Unpacking libgl1-mesa-glx:amd64 (23.0.4-0ubuntu1~22.04.1) ...\n",
+      "Setting up libpciaccess0:amd64 (0.16-3) ...\n",
+      "Setting up libxau6:amd64 (1:1.0.9-1build5) ...\n",
+      "Setting up libxdmcp6:amd64 (1:1.1.3-0ubuntu5) ...\n",
+      "Setting up libxcb1:amd64 (1.14-3ubuntu3) ...\n",
+      "Setting up libxcb-xfixes0:amd64 (1.14-3ubuntu3) ...\n",
+      "Setting up libglvnd0:amd64 (1.4.0-1) ...\n",
+      "Setting up libxcb-glx0:amd64 (1.14-3ubuntu3) ...\n",
+      "Setting up libsensors-config (1:3.6.0-7ubuntu1) ...\n",
+      "Setting up libxcb-shm0:amd64 (1.14-3ubuntu3) ...\n",
+      "Setting up libxcb-present0:amd64 (1.14-3ubuntu3) ...\n",
+      "Setting up libx11-data (2:1.7.5-1ubuntu0.3) ...\n",
+      "Setting up libxcb-sync1:amd64 (1.14-3ubuntu3) ...\n",
+      "Setting up libsensors5:amd64 (1:3.6.0-7ubuntu1) ...\n",
+      "Setting up libglapi-mesa:amd64 (23.2.1-1ubuntu3.1~22.04.2) ...\n",
+      "Setting up libxcb-dri2-0:amd64 (1.14-3ubuntu3) ...\n",
+      "Setting up libxshmfence1:amd64 (1.3-1build4) ...\n",
+      "Setting up libxcb-randr0:amd64 (1.14-3ubuntu3) ...\n",
+      "Setting up libllvm15:amd64 (1:15.0.7-0ubuntu0.22.04.3) ...\n",
+      "Setting up libx11-6:amd64 (2:1.7.5-1ubuntu0.3) ...\n",
+      "Setting up libdrm-common (2.4.113-2~ubuntu0.22.04.1) ...\n",
+      "Setting up libxcb-dri3-0:amd64 (1.14-3ubuntu3) ...\n",
+      "Setting up libx11-xcb1:amd64 (2:1.7.5-1ubuntu0.3) ...\n",
+      "Setting up libxext6:amd64 (2:1.3.4-1build1) ...\n",
+      "Setting up libxxf86vm1:amd64 (1:1.1.4-1build3) ...\n",
+      "Setting up libxfixes3:amd64 (1:6.0.0-1) ...\n",
+      "Setting up libdrm2:amd64 (2.4.113-2~ubuntu0.22.04.1) ...\n",
+      "Setting up libdrm-amdgpu1:amd64 (2.4.113-2~ubuntu0.22.04.1) ...\n",
+      "Setting up libdrm-nouveau2:amd64 (2.4.113-2~ubuntu0.22.04.1) ...\n",
+      "Setting up libdrm-radeon1:amd64 (2.4.113-2~ubuntu0.22.04.1) ...\n",
+      "Setting up libdrm-intel1:amd64 (2.4.113-2~ubuntu0.22.04.1) ...\n",
+      "Setting up libgl1-mesa-dri:amd64 (23.2.1-1ubuntu3.1~22.04.2) ...\n",
+      "Setting up libgl1-amber-dri:amd64 (21.3.9-0ubuntu1~22.04.1) ...\n",
+      "Setting up libglx-mesa0:amd64 (23.2.1-1ubuntu3.1~22.04.2) ...\n",
+      "Setting up libglx0:amd64 (1.4.0-1) ...\n",
+      "Setting up libgl1:amd64 (1.4.0-1) ...\n",
+      "Setting up libgl1-mesa-glx:amd64 (23.0.4-0ubuntu1~22.04.1) ...\n",
+      "Processing triggers for libc-bin (2.35-0ubuntu3.4) ...\n",
+      "yes: standard output: Broken pipe\n"
+     ]
+    }
+   ],
+   "source": [
+    "!sudo apt-get update\n",
+    "!yes | sudo apt-get install libgl1-mesa-glx"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 9,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "/workspace/yolov9\n"
+     ]
+    }
+   ],
+   "source": [
+    "%cd /workspace/yolov9"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 22,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Collecting Pillow==9.5.0\n",
+      "  Downloading Pillow-9.5.0-cp310-cp310-manylinux_2_28_x86_64.whl.metadata (9.5 kB)\n",
+      "Downloading Pillow-9.5.0-cp310-cp310-manylinux_2_28_x86_64.whl (3.4 MB)\n",
+      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m3.4/3.4 MB\u001b[0m \u001b[31m54.5 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0ma \u001b[36m0:00:01\u001b[0m\n",
+      "\u001b[?25hInstalling collected packages: Pillow\n",
+      "  Attempting uninstall: Pillow\n",
+      "    Found existing installation: Pillow 10.0.1\n",
+      "    Uninstalling Pillow-10.0.1:\n",
+      "      Successfully uninstalled Pillow-10.0.1\n",
+      "Successfully installed Pillow-9.5.0\n",
+      "\u001b[33mWARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv\u001b[0m\u001b[33m\n",
+      "\u001b[0mNote: you may need to restart the kernel to use updated packages.\n"
+     ]
+    }
+   ],
+   "source": [
+    "pip install Pillow==9.5.0"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 23,
+   "metadata": {
+    "colab": {
+     "base_uri": "https://localhost:8080/"
+    },
+    "id": "jjB9EA9yRCyR",
+    "outputId": "2e388fee-6ec5-492d-b698-55d2d1cec2f2"
+   },
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "\u001b[34m\u001b[1mtrain_dual: \u001b[0mweights=, cfg=/workspace/YOLO9_KMUTT/yolov9/models/detect/yolov9-kmuttX.yaml, data=/workspace/6400 images/data.yaml, hyp=hyp.scratch-high.yaml, epochs=500, batch_size=4, imgsz=640, rect=False, resume=False, nosave=False, noval=False, noautoanchor=False, noplots=False, evolve=None, bucket=, cache=None, image_weights=False, device=0, multi_scale=False, single_cls=False, optimizer=SGD, sync_bn=False, workers=8, project=runs/train, name=yolov9-c, exist_ok=False, quad=False, cos_lr=False, flat_cos_lr=False, fixed_lr=False, label_smoothing=0.0, patience=100, freeze=[0], save_period=-1, seed=0, local_rank=-1, min_items=0, close_mosaic=15, entity=None, upload_dataset=False, bbox_interval=-1, artifact_alias=latest\n",
+      "YOLO 🚀 2024-6-7 Python-3.10.13 torch-2.2.0 CUDA:0 (NVIDIA GeForce RTX 4090, 24210MiB)\n",
+      "\n",
+      "\u001b[34m\u001b[1mhyperparameters: \u001b[0mlr0=0.01, lrf=0.01, momentum=0.937, weight_decay=0.0005, warmup_epochs=3.0, warmup_momentum=0.8, warmup_bias_lr=0.1, box=7.5, cls=0.5, cls_pw=1.0, obj=0.7, obj_pw=1.0, dfl=1.5, iou_t=0.2, anchor_t=5.0, fl_gamma=0.0, hsv_h=0.015, hsv_s=0.7, hsv_v=0.4, degrees=0.0, translate=0.1, scale=0.9, shear=0.0, perspective=0.0, flipud=0.0, fliplr=0.5, mosaic=1.0, mixup=0.15, copy_paste=0.3\n",
+      "\u001b[34m\u001b[1mClearML: \u001b[0mrun 'pip install clearml' to automatically track, visualize and remotely train YOLO 🚀 in ClearML\n",
+      "\u001b[34m\u001b[1mComet: \u001b[0mrun 'pip install comet_ml' to automatically track and visualize YOLO 🚀 runs in Comet\n",
+      "\u001b[34m\u001b[1mTensorBoard: \u001b[0mStart with 'tensorboard --logdir runs/train', view at http://localhost:6006/\n",
+      "Overriding model.yaml nc=80 with nc=10\n",
+      "\n",
+      "                 from  n    params  module                                  arguments                     \n",
+      "  0                -1  1         0  models.common.Silence                   []                            \n",
+      "  1                -1  1      1856  models.common.Conv                      [3, 64, 3, 2]                 \n",
+      "  2                -1  1     73984  models.common.Conv                      [64, 128, 3, 2]               \n",
+      "  3                -1  1    252160  models.common.RepNCSPELAN4              [128, 256, 128, 64, 2]        \n",
+      "  4                -1  1    164352  models.common.ADown                     [256, 256]                    \n",
+      "  5                -1  1   1004032  models.common.RepNCSPELAN4              [256, 512, 256, 128, 2]       \n",
+      "  6                -1  1    656384  models.common.ADown                     [512, 512]                    \n",
+      "  7                -1  1   4006912  models.common.RepNCSPELAN4              [512, 1024, 512, 256, 2]      \n",
+      "  8                -1  1   2623488  models.common.ADown                     [1024, 1024]                  \n",
+      "  9                -1  1   4269056  models.common.RepNCSPELAN4              [1024, 1024, 512, 256, 2]     \n",
+      " 10                 1  1      4160  models.common.CBLinear                  [64, [64]]                    \n",
+      " 11                 3  1     49344  models.common.CBLinear                  [256, [64, 128]]              \n",
+      " 12                 5  1    229824  models.common.CBLinear                  [512, [64, 128, 256]]         \n",
+      " 13                 7  1    984000  models.common.CBLinear                  [1024, [64, 128, 256, 512]]   \n",
+      " 14                 9  1   2033600  models.common.CBLinear                  [1024, [64, 128, 256, 512, 1024]]\n",
+      " 15                 0  1      1856  models.common.Conv                      [3, 64, 3, 2]                 \n",
+      " 16[10, 11, 12, 13, 14, -1]  1         0  models.common.CBFuse                    [[0, 0, 0, 0, 0]]             \n",
+      " 17                -1  1     73984  models.common.Conv                      [64, 128, 3, 2]               \n",
+      " 18[11, 12, 13, 14, -1]  1         0  models.common.CBFuse                    [[1, 1, 1, 1]]                \n",
+      " 19                -1  1    252160  models.common.RepNCSPELAN4              [128, 256, 128, 64, 2]        \n",
+      " 20                -1  1    164352  models.common.ADown                     [256, 256]                    \n",
+      " 21  [12, 13, 14, -1]  1         0  models.common.CBFuse                    [[2, 2, 2]]                   \n",
+      " 22                -1  1   1004032  models.common.RepNCSPELAN4              [256, 512, 256, 128, 2]       \n",
+      " 23                -1  1    656384  models.common.ADown                     [512, 512]                    \n",
+      " 24      [13, 14, -1]  1         0  models.common.CBFuse                    [[3, 3]]                      \n",
+      " 25                -1  1   4006912  models.common.RepNCSPELAN4              [512, 1024, 512, 256, 2]      \n",
+      " 26                -1  1   2623488  models.common.ADown                     [1024, 1024]                  \n",
+      " 27          [14, -1]  1         0  models.common.CBFuse                    [[4]]                         \n",
+      " 28                -1  1   4269056  models.common.RepNCSPELAN4              [1024, 1024, 512, 256, 2]     \n",
+      " 29                 9  1    787968  models.common.SPPELAN                   [1024, 512, 256]              \n",
+      " 30                -1  1         0  torch.nn.modules.upsampling.Upsample    [None, 2, 'nearest']          \n",
+      " 31           [-1, 7]  1         0  models.common.Concat                    [1]                           \n",
+      " 32                -1  1   4005888  models.common.RepNCSPELAN4              [1536, 512, 512, 256, 2]      \n",
+      " 33                -1  1         0  torch.nn.modules.upsampling.Upsample    [None, 2, 'nearest']          \n",
+      " 34           [-1, 5]  1         0  models.common.Concat                    [1]                           \n",
+      " 35                -1  1   1069056  models.common.RepNCSPELAN4              [1024, 256, 256, 128, 2]      \n",
+      " 36                28  1    787968  models.common.SPPELAN                   [1024, 512, 256]              \n",
+      " 37                -1  1         0  torch.nn.modules.upsampling.Upsample    [None, 2, 'nearest']          \n",
+      " 38          [-1, 25]  1         0  models.common.Concat                    [1]                           \n",
+      " 39                -1  1   4005888  models.common.RepNCSPELAN4              [1536, 512, 512, 256, 2]      \n",
+      " 40                -1  1         0  torch.nn.modules.upsampling.Upsample    [None, 2, 'nearest']          \n",
+      " 41          [-1, 22]  1         0  models.common.Concat                    [1]                           \n",
+      " 42                -1  1   1069056  models.common.RepNCSPELAN4              [1024, 256, 256, 128, 2]      \n",
+      " 43                -1  1    164352  models.common.ADown                     [256, 256]                    \n",
+      " 44          [-1, 39]  1         0  models.common.Concat                    [1]                           \n",
+      " 45                -1  1   3612672  models.common.RepNCSPELAN4              [768, 512, 512, 256, 2]       \n",
+      " 46                -1  1    656384  models.common.ADown                     [512, 512]                    \n",
+      " 47          [-1, 36]  1         0  models.common.Concat                    [1]                           \n",
+      " 48                -1  1  13910016  models.common.RepNCSPELAN4              [1024, 1024, 1024, 512, 2]    \n",
+      " 49                -1  1   3948032  models.common.PSA                       [1024, 1024]                  \n",
+      " 50                35  1    249728  models.common.PSA                       [256, 256]                    \n",
+      " 51                32  1    990976  models.common.PSA                       [512, 512]                    \n",
+      " 52                29  1    990976  models.common.PSA                       [512, 512]                    \n",
+      " 53                42  1    249728  models.common.PSA                       [256, 256]                    \n",
+      " 54                45  1    990976  models.common.PSA                       [512, 512]                    \n",
+      " 55[50, 51, 52, 53, 54, 49]  1  12471260  models.yolo.DualDDetect                 [10, [256, 512, 512, 256, 512, 1024]]\n",
+      "yolov9-kmuttX summary: 1643 layers, 79366300 parameters, 79366268 gradients, 263.8 GFLOPs\n",
+      "\n",
+      "\u001b[34m\u001b[1mAMP: \u001b[0mchecks passed ✅\n",
+      "\u001b[34m\u001b[1moptimizer:\u001b[0m SGD(lr=0.01) with parameter groups 398 weight(decay=0.0), 417 weight(decay=0.0005), 415 bias\n",
+      "\u001b[34m\u001b[1malbumentations: \u001b[0mBlur(p=0.01, blur_limit=(3, 7)), MedianBlur(p=0.01, blur_limit=(3, 7)), ToGray(p=0.01), CLAHE(p=0.01, clip_limit=(1, 4.0), tile_grid_size=(8, 8))\n",
+      "\u001b[34m\u001b[1mtrain: \u001b[0mScanning /workspace/6400 images/train/labels.cache... 5120 images, 7 back\u001b[0m\n",
+      "\u001b[34m\u001b[1mval: \u001b[0mScanning /workspace/6400 images/valid/labels.cache... 640 images, 0 backgro\u001b[0m\n",
+      "Plotting labels to runs/train/yolov9-c10/labels.jpg... \n",
+      "Image sizes 640 train, 640 val\n",
+      "Using 4 dataloader workers\n",
+      "Logging results to \u001b[1mruns/train/yolov9-c10\u001b[0m\n",
+      "Starting training for 500 epochs...\n",
+      "\n",
+      "      Epoch    GPU_mem   box_loss   cls_loss   dfl_loss  Instances       Size\n",
+      "      0/499      16.5G      4.939      7.156      5.232         11        640:  WARNING ⚠️ TensorBoard graph visualization failure Only tensors, lists, tuples of tensors, or dictionary of tensors can be output from traced functions\n",
+      "      0/499      16.7G      5.236      7.764      5.234         23        640:  ^C\n",
+      "      0/499      16.7G      5.236      7.764      5.234         23        640:  \n",
+      "Traceback (most recent call last):\n",
+      "  File \"/workspace/YOLO9_KMUTT/yolov9/train_dual.py\", line 644, in <module>\n",
+      "    main(opt)\n",
+      "  File \"/workspace/YOLO9_KMUTT/yolov9/train_dual.py\", line 538, in main\n",
+      "    train(opt.hyp, opt, device, callbacks)\n",
+      "  File \"/workspace/YOLO9_KMUTT/yolov9/train_dual.py\", line 314, in train\n",
+      "    pred = model(imgs)  # forward\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1511, in _wrapped_call_impl\n",
+      "    return self._call_impl(*args, **kwargs)\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1520, in _call_impl\n",
+      "    return forward_call(*args, **kwargs)\n",
+      "  File \"/workspace/YOLO9_KMUTT/yolov9/models/yolo.py\", line 633, in forward\n",
+      "    return self._forward_once(x, profile, visualize)  # single-scale inference, train\n",
+      "  File \"/workspace/YOLO9_KMUTT/yolov9/models/yolo.py\", line 533, in _forward_once\n",
+      "    x = m(x)  # run\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1511, in _wrapped_call_impl\n",
+      "    return self._call_impl(*args, **kwargs)\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1520, in _call_impl\n",
+      "    return forward_call(*args, **kwargs)\n",
+      "  File \"/workspace/YOLO9_KMUTT/yolov9/models/common.py\", line 613, in forward\n",
+      "    y.extend((m(y[-1])) for m in [self.cv2, self.cv3])\n",
+      "  File \"/workspace/YOLO9_KMUTT/yolov9/models/common.py\", line 613, in <genexpr>\n",
+      "    y.extend((m(y[-1])) for m in [self.cv2, self.cv3])\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1511, in _wrapped_call_impl\n",
+      "    return self._call_impl(*args, **kwargs)\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1520, in _call_impl\n",
+      "    return forward_call(*args, **kwargs)\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/container.py\", line 217, in forward\n",
+      "    input = module(input)\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1511, in _wrapped_call_impl\n",
+      "    return self._call_impl(*args, **kwargs)\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1520, in _call_impl\n",
+      "    return forward_call(*args, **kwargs)\n",
+      "  File \"/workspace/YOLO9_KMUTT/yolov9/models/common.py\", line 384, in forward\n",
+      "    return self.cv3(torch.cat((self.m(self.cv1(x)), self.cv2(x)), 1))\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1511, in _wrapped_call_impl\n",
+      "    return self._call_impl(*args, **kwargs)\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1520, in _call_impl\n",
+      "    return forward_call(*args, **kwargs)\n",
+      "  File \"/workspace/YOLO9_KMUTT/yolov9/models/common.py\", line 54, in forward\n",
+      "    return self.act(self.bn(self.conv(x)))\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1511, in _wrapped_call_impl\n",
+      "    return self._call_impl(*args, **kwargs)\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1520, in _call_impl\n",
+      "    return forward_call(*args, **kwargs)\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/batchnorm.py\", line 175, in forward\n",
+      "    return F.batch_norm(\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/functional.py\", line 2482, in batch_norm\n",
+      "    return torch.batch_norm(\n",
+      "KeyboardInterrupt\n"
+     ]
+    }
+   ],
+   "source": [
+    "!python train_dual.py --workers 8 --device 0 --batch 4 --data '/workspace/6400 images/data.yaml' --img 640 --cfg /workspace/YOLO9_KMUTT/yolov9/models/detect/yolov9-kmuttX.yaml --weights '' --name yolov9-c --hyp hyp.scratch-high.yaml --min-items 0 --epochs 500 --close-mosaic 15"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 9,
+   "metadata": {
+    "id": "VRMBngDDsIok"
+   },
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "[2024-06-01 23:55:21,192] torch.distributed.run: [WARNING] \n",
+      "[2024-06-01 23:55:21,192] torch.distributed.run: [WARNING] *****************************************\n",
+      "[2024-06-01 23:55:21,192] torch.distributed.run: [WARNING] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. \n",
+      "[2024-06-01 23:55:21,192] torch.distributed.run: [WARNING] *****************************************\n",
+      "\u001b[34m\u001b[1mtrain_dual: \u001b[0mweights=, cfg=models/detect/yolov9-c.yaml, data=/workspace/norway/data.yaml, hyp=hyp.scratch-high.yaml, epochs=500, batch_size=30, imgsz=640, rect=False, resume=False, nosave=False, noval=False, noautoanchor=False, noplots=False, evolve=None, bucket=, cache=None, image_weights=False, device=0,1, multi_scale=False, single_cls=False, optimizer=SGD, sync_bn=True, workers=8, project=runs/train, name=yolov9-c, exist_ok=False, quad=False, cos_lr=False, flat_cos_lr=False, fixed_lr=False, label_smoothing=0.0, patience=100, freeze=[0], save_period=-1, seed=0, local_rank=-1, min_items=0, close_mosaic=15, entity=None, upload_dataset=False, bbox_interval=-1, artifact_alias=latest\n",
+      "Traceback (most recent call last):\n",
+      "  File \"/workspace/yolov9/train_dual.py\", line 644, in <module>\n",
+      "    main(opt)\n",
+      "  File \"/workspace/yolov9/train_dual.py\", line 530, in main\n",
+      "    assert opt.batch_size % WORLD_SIZE == 0, f'--batch-size {opt.batch_size} must be multiple of WORLD_SIZE'\n",
+      "AssertionError: --batch-size 30 must be multiple of WORLD_SIZE\n",
+      "Traceback (most recent call last):\n",
+      "  File \"/workspace/yolov9/train_dual.py\", line 644, in <module>\n",
+      "    main(opt)\n",
+      "  File \"/workspace/yolov9/train_dual.py\", line 530, in main\n",
+      "    assert opt.batch_size % WORLD_SIZE == 0, f'--batch-size {opt.batch_size} must be multiple of WORLD_SIZE'\n",
+      "AssertionError: --batch-size 30 must be multiple of WORLD_SIZE\n",
+      "Traceback (most recent call last):\n",
+      "  File \"/workspace/yolov9/train_dual.py\", line 644, in <module>\n",
+      "    main(opt)\n",
+      "  File \"/workspace/yolov9/train_dual.py\", line 530, in main\n",
+      "    assert opt.batch_size % WORLD_SIZE == 0, f'--batch-size {opt.batch_size} must be multiple of WORLD_SIZE'\n",
+      "AssertionError: --batch-size 30 must be multiple of WORLD_SIZE\n",
+      "YOLO 🚀 v0.1-94-gcd88016 Python-3.10.13 torch-2.2.0 CUDA:0 (NVIDIA GeForce RTX 4090, 24217MiB)\n",
+      "                                                    CUDA:1 (NVIDIA GeForce RTX 4090, 24217MiB)\n",
+      "\n",
+      "Traceback (most recent call last):\n",
+      "  File \"/workspace/yolov9/train_dual.py\", line 644, in <module>\n",
+      "    main(opt)\n",
+      "  File \"/workspace/yolov9/train_dual.py\", line 530, in main\n",
+      "    assert opt.batch_size % WORLD_SIZE == 0, f'--batch-size {opt.batch_size} must be multiple of WORLD_SIZE'\n",
+      "AssertionError: --batch-size 30 must be multiple of WORLD_SIZE\n",
+      "Traceback (most recent call last):\n",
+      "  File \"/workspace/yolov9/train_dual.py\", line 644, in <module>\n",
+      "    main(opt)\n",
+      "  File \"/workspace/yolov9/train_dual.py\", line 530, in main\n",
+      "    assert opt.batch_size % WORLD_SIZE == 0, f'--batch-size {opt.batch_size} must be multiple of WORLD_SIZE'\n",
+      "AssertionError: --batch-size 30 must be multiple of WORLD_SIZE\n",
+      "Traceback (most recent call last):\n",
+      "  File \"/workspace/yolov9/train_dual.py\", line 644, in <module>\n",
+      "    main(opt)\n",
+      "  File \"/workspace/yolov9/train_dual.py\", line 530, in main\n",
+      "    assert opt.batch_size % WORLD_SIZE == 0, f'--batch-size {opt.batch_size} must be multiple of WORLD_SIZE'\n",
+      "AssertionError: --batch-size 30 must be multiple of WORLD_SIZE\n",
+      "Traceback (most recent call last):\n",
+      "  File \"/workspace/yolov9/train_dual.py\", line 644, in <module>\n",
+      "    main(opt)\n",
+      "  File \"/workspace/yolov9/train_dual.py\", line 530, in main\n",
+      "    assert opt.batch_size % WORLD_SIZE == 0, f'--batch-size {opt.batch_size} must be multiple of WORLD_SIZE'\n",
+      "AssertionError: --batch-size 30 must be multiple of WORLD_SIZE\n",
+      "Traceback (most recent call last):\n",
+      "  File \"/workspace/yolov9/train_dual.py\", line 644, in <module>\n",
+      "    main(opt)\n",
+      "  File \"/workspace/yolov9/train_dual.py\", line 530, in main\n",
+      "    assert opt.batch_size % WORLD_SIZE == 0, f'--batch-size {opt.batch_size} must be multiple of WORLD_SIZE'\n",
+      "AssertionError: --batch-size 30 must be multiple of WORLD_SIZE\n",
+      "[2024-06-01 23:55:26,209] torch.distributed.elastic.multiprocessing.api: [ERROR] failed (exitcode: 1) local_rank: 0 (pid: 4360) of binary: /opt/conda/bin/python\n",
+      "Traceback (most recent call last):\n",
+      "  File \"/opt/conda/bin/torchrun\", line 33, in <module>\n",
+      "    sys.exit(load_entry_point('torch==2.2.0', 'console_scripts', 'torchrun')())\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py\", line 347, in wrapper\n",
+      "    return f(*args, **kwargs)\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/distributed/run.py\", line 812, in main\n",
+      "    run(args)\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/distributed/run.py\", line 803, in run\n",
+      "    elastic_launch(\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/distributed/launcher/api.py\", line 135, in __call__\n",
+      "    return launch_agent(self._config, self._entrypoint, list(args))\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/distributed/launcher/api.py\", line 268, in launch_agent\n",
+      "    raise ChildFailedError(\n",
+      "torch.distributed.elastic.multiprocessing.errors.ChildFailedError: \n",
+      "============================================================\n",
+      "train_dual.py FAILED\n",
+      "------------------------------------------------------------\n",
+      "Failures:\n",
+      "[1]:\n",
+      "  time      : 2024-06-01_23:55:26\n",
+      "  host      : de3bed18af84\n",
+      "  rank      : 1 (local_rank: 1)\n",
+      "  exitcode  : 1 (pid: 4361)\n",
+      "  error_file: <N/A>\n",
+      "  traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html\n",
+      "[2]:\n",
+      "  time      : 2024-06-01_23:55:26\n",
+      "  host      : de3bed18af84\n",
+      "  rank      : 2 (local_rank: 2)\n",
+      "  exitcode  : 1 (pid: 4362)\n",
+      "  error_file: <N/A>\n",
+      "  traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html\n",
+      "[3]:\n",
+      "  time      : 2024-06-01_23:55:26\n",
+      "  host      : de3bed18af84\n",
+      "  rank      : 3 (local_rank: 3)\n",
+      "  exitcode  : 1 (pid: 4363)\n",
+      "  error_file: <N/A>\n",
+      "  traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html\n",
+      "[4]:\n",
+      "  time      : 2024-06-01_23:55:26\n",
+      "  host      : de3bed18af84\n",
+      "  rank      : 4 (local_rank: 4)\n",
+      "  exitcode  : 1 (pid: 4364)\n",
+      "  error_file: <N/A>\n",
+      "  traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html\n",
+      "[5]:\n",
+      "  time      : 2024-06-01_23:55:26\n",
+      "  host      : de3bed18af84\n",
+      "  rank      : 5 (local_rank: 5)\n",
+      "  exitcode  : 1 (pid: 4365)\n",
+      "  error_file: <N/A>\n",
+      "  traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html\n",
+      "[6]:\n",
+      "  time      : 2024-06-01_23:55:26\n",
+      "  host      : de3bed18af84\n",
+      "  rank      : 6 (local_rank: 6)\n",
+      "  exitcode  : 1 (pid: 4366)\n",
+      "  error_file: <N/A>\n",
+      "  traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html\n",
+      "[7]:\n",
+      "  time      : 2024-06-01_23:55:26\n",
+      "  host      : de3bed18af84\n",
+      "  rank      : 7 (local_rank: 7)\n",
+      "  exitcode  : 1 (pid: 4367)\n",
+      "  error_file: <N/A>\n",
+      "  traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html\n",
+      "------------------------------------------------------------\n",
+      "Root Cause (first observed failure):\n",
+      "[0]:\n",
+      "  time      : 2024-06-01_23:55:26\n",
+      "  host      : de3bed18af84\n",
+      "  rank      : 0 (local_rank: 0)\n",
+      "  exitcode  : 1 (pid: 4360)\n",
+      "  error_file: <N/A>\n",
+      "  traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html\n",
+      "============================================================\n"
+     ]
+    }
+   ],
+   "source": [
+    "!torchrun  --nproc_per_node 2 --master_port 9527 train_dual.py  \\\n",
+    "--workers 8 --device 0,1 --sync-bn --batch 30 --data '/workspace/norway/data.yaml'  \\\n",
+    "--img 640 --cfg models/detect/yolov9-c.yaml --weights '' --name yolov9-c --hyp hyp.scratch-high.yaml \\\n",
+    "--min-items 0 --epochs 500 --close-mosaic 15"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 10,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "[2024-06-02 02:12:23,563] torch.distributed.run: [WARNING] \n",
+      "[2024-06-02 02:12:23,563] torch.distributed.run: [WARNING] *****************************************\n",
+      "[2024-06-02 02:12:23,563] torch.distributed.run: [WARNING] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. \n",
+      "[2024-06-02 02:12:23,563] torch.distributed.run: [WARNING] *****************************************\n",
+      "\u001b[34m\u001b[1mtrain_dual: \u001b[0mweights=, cfg=models/detect/yolov9-c.yaml, data=/workspace/road damage/data.yaml, hyp=hyp.scratch-high.yaml, epochs=500, batch_size=30, imgsz=640, rect=False, resume=False, nosave=False, noval=False, noautoanchor=False, noplots=False, evolve=None, bucket=, cache=None, image_weights=False, device=0,1, multi_scale=False, single_cls=False, optimizer=SGD, sync_bn=True, workers=8, project=runs/train, name=yolov9-c, exist_ok=False, quad=False, cos_lr=False, flat_cos_lr=False, fixed_lr=False, label_smoothing=0.0, patience=100, freeze=[0], save_period=-1, seed=0, local_rank=-1, min_items=0, close_mosaic=15, entity=None, upload_dataset=False, bbox_interval=-1, artifact_alias=latest\n",
+      "YOLO 🚀 v0.1-94-gcd88016 Python-3.10.13 torch-2.2.0 CUDA:0 (NVIDIA GeForce RTX 4090, 24217MiB)\n",
+      "                                                    CUDA:1 (NVIDIA GeForce RTX 4090, 24217MiB)\n",
+      "\n",
+      "\u001b[34m\u001b[1mhyperparameters: \u001b[0mlr0=0.01, lrf=0.01, momentum=0.937, weight_decay=0.0005, warmup_epochs=3.0, warmup_momentum=0.8, warmup_bias_lr=0.1, box=7.5, cls=0.5, cls_pw=1.0, obj=0.7, obj_pw=1.0, dfl=1.5, iou_t=0.2, anchor_t=5.0, fl_gamma=0.0, hsv_h=0.015, hsv_s=0.7, hsv_v=0.4, degrees=0.0, translate=0.1, scale=0.9, shear=0.0, perspective=0.0, flipud=0.0, fliplr=0.5, mosaic=1.0, mixup=0.15, copy_paste=0.3\n",
+      "\u001b[34m\u001b[1mClearML: \u001b[0mrun 'pip install clearml' to automatically track, visualize and remotely train YOLO 🚀 in ClearML\n",
+      "\u001b[34m\u001b[1mComet: \u001b[0mrun 'pip install comet_ml' to automatically track and visualize YOLO 🚀 runs in Comet\n",
+      "\u001b[34m\u001b[1mTensorBoard: \u001b[0mStart with 'tensorboard --logdir runs/train', view at http://localhost:6006/\n",
+      "Overriding model.yaml nc=80 with nc=7\n",
+      "\n",
+      "                 from  n    params  module                                  arguments                     \n",
+      "  0                -1  1         0  models.common.Silence                   []                            \n",
+      "  1                -1  1      1856  models.common.Conv                      [3, 64, 3, 2]                 \n",
+      "  2                -1  1     73984  models.common.Conv                      [64, 128, 3, 2]               \n",
+      "  3                -1  1    291456  models.common.RepNCSPELAN4              [128, 256, 128, 64, 3]        \n",
+      "  4                -1  1    164352  models.common.ADown                     [256, 256]                    \n",
+      "  5                -1  1   1160448  models.common.RepNCSPELAN4              [256, 512, 256, 128, 3]       \n",
+      "  6                -1  1    656384  models.common.ADown                     [512, 512]                    \n",
+      "  7                -1  1   4105728  models.common.RepNCSPELAN4              [512, 512, 512, 256, 3]       \n",
+      "  8                -1  1    656384  models.common.ADown                     [512, 512]                    \n",
+      "  9                -1  1   4105728  models.common.RepNCSPELAN4              [512, 512, 512, 256, 3]       \n",
+      " 10                -1  1    656896  models.common.SPPELAN                   [512, 512, 256]               \n",
+      " 11                -1  1         0  torch.nn.modules.upsampling.Upsample    [None, 2, 'nearest']          \n",
+      " 12           [-1, 7]  1         0  models.common.Concat                    [1]                           \n",
+      " 13                -1  1   3119616  models.common.RepNCSPELAN4              [1024, 512, 512, 256, 1]      \n",
+      " 14                -1  1         0  torch.nn.modules.upsampling.Upsample    [None, 2, 'nearest']          \n",
+      " 15           [-1, 5]  1         0  models.common.Concat                    [1]                           \n",
+      " 16                -1  1    912640  models.common.RepNCSPELAN4              [1024, 256, 256, 128, 1]      \n",
+      " 17                -1  1    164352  models.common.ADown                     [256, 256]                    \n",
+      " 18          [-1, 13]  1         0  models.common.Concat                    [1]                           \n",
+      " 19                -1  1   2988544  models.common.RepNCSPELAN4              [768, 512, 512, 256, 1]       \n",
+      " 20                -1  1    656384  models.common.ADown                     [512, 512]                    \n",
+      " 21          [-1, 10]  1         0  models.common.Concat                    [1]                           \n",
+      " 22                -1  1   3119616  models.common.RepNCSPELAN4              [1024, 512, 512, 256, 1]      \n",
+      " 23                 5  1    131328  models.common.CBLinear                  [512, [256]]                  \n",
+      " 24                 7  1    393984  models.common.CBLinear                  [512, [256, 512]]             \n",
+      " 25                 9  1    656640  models.common.CBLinear                  [512, [256, 512, 512]]        \n",
+      " 26                 0  1      1856  models.common.Conv                      [3, 64, 3, 2]                 \n",
+      " 27                -1  1     73984  models.common.Conv                      [64, 128, 3, 2]               \n",
+      " 28                -1  1    212864  models.common.RepNCSPELAN4              [128, 256, 128, 64, 1]        \n",
+      " 29                -1  1    164352  models.common.ADown                     [256, 256]                    \n",
+      " 30  [23, 24, 25, -1]  1         0  models.common.CBFuse                    [[0, 0, 0]]                   \n",
+      " 31                -1  1    847616  models.common.RepNCSPELAN4              [256, 512, 256, 128, 1]       \n",
+      " 32                -1  1    656384  models.common.ADown                     [512, 512]                    \n",
+      " 33      [24, 25, -1]  1         0  models.common.CBFuse                    [[1, 1]]                      \n",
+      " 34                -1  1   2857472  models.common.RepNCSPELAN4              [512, 512, 512, 256, 1]       \n",
+      " 35                -1  1    656384  models.common.ADown                     [512, 512]                    \n",
+      " 36          [25, -1]  1         0  models.common.CBFuse                    [[2]]                         \n",
+      " 37                -1  1   2857472  models.common.RepNCSPELAN4              [512, 512, 512, 256, 1]       \n",
+      " 38[31, 34, 37, 16, 19, 22]  1  21556682  models.yolo.DualDDetect                 [7, [512, 512, 512, 256, 512, 512]]\n",
+      "yolov9-c summary: 1170 layers, 53901386 parameters, 53901354 gradients, 252.0 GFLOPs\n",
+      "\n",
+      "\u001b[34m\u001b[1mAMP: \u001b[0mchecks passed ✅\n",
+      "\u001b[34m\u001b[1moptimizer:\u001b[0m SGD(lr=0.01) with parameter groups 286 weight(decay=0.0), 303 weight(decay=0.00046875), 301 bias\n",
+      "Using SyncBatchNorm()\n",
+      "\u001b[34m\u001b[1malbumentations: \u001b[0mBlur(p=0.01, blur_limit=(3, 7)), MedianBlur(p=0.01, blur_limit=(3, 7)), ToGray(p=0.01), CLAHE(p=0.01, clip_limit=(1, 4.0), tile_grid_size=(8, 8))\n",
+      "\u001b[34m\u001b[1mtrain: \u001b[0mScanning /workspace/road damage/train/labels... 4672 images, 345 backgrou\u001b[0m\n",
+      "\u001b[34m\u001b[1mtrain: \u001b[0mWARNING ⚠️ /workspace/road damage/train/images/Japan_006916_jpg.rf.5dc73c813f7053289379ab9cf4e81173.jpg: 1 duplicate labels removed\n",
+      "\u001b[34m\u001b[1mtrain: \u001b[0mNew cache created: /workspace/road damage/train/labels.cache\n",
+      "\u001b[34m\u001b[1mval: \u001b[0mScanning /workspace/road damage/valid/labels... 450 images, 34 backgrounds,\u001b[0m\n",
+      "\u001b[34m\u001b[1mval: \u001b[0mNew cache created: /workspace/road damage/valid/labels.cache\n",
+      "Plotting labels to runs/train/yolov9-c4/labels.jpg... \n",
+      "Image sizes 640 train, 640 val\n",
+      "Using 16 dataloader workers\n",
+      "Logging results to \u001b[1mruns/train/yolov9-c4\u001b[0m\n",
+      "Starting training for 500 epochs...\n",
+      "\n",
+      "      Epoch    GPU_mem   box_loss   cls_loss   dfl_loss  Instances       Size\n",
+      "      0/499      20.4G        4.8      6.503      5.217         74        640:  Exception in thread Thread-11 (plot_images):\n",
+      "Traceback (most recent call last):\n",
+      "  File \"/opt/conda/lib/python3.10/threading.py\", line 1016, in _bootstrap_inner\n",
+      "    self.run()\n",
+      "  File \"/opt/conda/lib/python3.10/threading.py\", line 953, in run\n",
+      "    self._target(*self._args, **self._kwargs)\n",
+      "  File \"/workspace/yolov9/utils/plots.py\", line 300, in plot_images\n",
+      "    annotator.box_label(box, label, color=color)\n",
+      "  File \"/workspace/yolov9/utils/plots.py\", line 86, in box_label\n",
+      "    w, h = self.font.getsize(label)  # text width, height\n",
+      "AttributeError: 'FreeTypeFont' object has no attribute 'getsize'\n",
+      "      0/499      20.4G      4.903      6.657      5.212         61        640:  Exception in thread Thread-12 (plot_images):\n",
+      "Traceback (most recent call last):\n",
+      "  File \"/opt/conda/lib/python3.10/threading.py\", line 1016, in _bootstrap_inner\n",
+      "    self.run()\n",
+      "  File \"/opt/conda/lib/python3.10/threading.py\", line 953, in run\n",
+      "    self._target(*self._args, **self._kwargs)\n",
+      "  File \"/workspace/yolov9/utils/plots.py\", line 300, in plot_images\n",
+      "    annotator.box_label(box, label, color=color)\n",
+      "  File \"/workspace/yolov9/utils/plots.py\", line 86, in box_label\n",
+      "    w, h = self.font.getsize(label)  # text width, height\n",
+      "AttributeError: 'FreeTypeFont' object has no attribute 'getsize'\n",
+      "      0/499      20.4G      4.703      6.511      5.215         57        640:  Exception in thread Thread-13 (plot_images):\n",
+      "Traceback (most recent call last):\n",
+      "  File \"/opt/conda/lib/python3.10/threading.py\", line 1016, in _bootstrap_inner\n",
+      "    self.run()\n",
+      "  File \"/opt/conda/lib/python3.10/threading.py\", line 953, in run\n",
+      "    self._target(*self._args, **self._kwargs)\n",
+      "  File \"/workspace/yolov9/utils/plots.py\", line 300, in plot_images\n",
+      "    annotator.box_label(box, label, color=color)\n",
+      "  File \"/workspace/yolov9/utils/plots.py\", line 86, in box_label\n",
+      "    w, h = self.font.getsize(label)  # text width, height\n",
+      "AttributeError: 'FreeTypeFont' object has no attribute 'getsize'\n",
+      "      0/499      20.5G      4.674      6.804      5.251         49        640:  ^C\n",
+      "[2024-06-02 02:12:52,085] torch.distributed.elastic.agent.server.api: [WARNING] Received Signals.SIGINT death signal, shutting down workers\n",
+      "[2024-06-02 02:12:52,086] torch.distributed.elastic.multiprocessing.api: [WARNING] Sending process 11363 closing signal SIGINT\n",
+      "[2024-06-02 02:12:52,086] torch.distributed.elastic.multiprocessing.api: [WARNING] Sending process 11364 closing signal SIGINT\n",
+      "      0/499      20.5G      4.674      6.804      5.251         49        640:  \n",
+      "Traceback (most recent call last):\n",
+      "  File \"/workspace/yolov9/train_dual.py\", line 644, in <module>\n",
+      "    main(opt)\n",
+      "  File \"/workspace/yolov9/train_dual.py\", line 538, in main\n",
+      "    train(opt.hyp, opt, device, callbacks)\n",
+      "  File \"/workspace/yolov9/train_dual.py\", line 314, in train\n",
+      "    pred = model(imgs)  # forward\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1511, in _wrapped_call_impl\n",
+      "    return self._call_impl(*args, **kwargs)\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1520, in _call_impl\n",
+      "    return forward_call(*args, **kwargs)\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/parallel/distributed.py\", line 1523, in forward\n",
+      "    else self._run_ddp_forward(*inputs, **kwargs)\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/parallel/distributed.py\", line 1359, in _run_ddp_forward\n",
+      "    return self.module(*inputs, **kwargs)  # type: ignore[index]\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1511, in _wrapped_call_impl\n",
+      "    return self._call_impl(*args, **kwargs)\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1520, in _call_impl\n",
+      "    return forward_call(*args, **kwargs)\n",
+      "  File \"/workspace/yolov9/models/yolo.py\", line 633, in forward\n",
+      "    return self._forward_once(x, profile, visualize)  # single-scale inference, train\n",
+      "  File \"/workspace/yolov9/models/yolo.py\", line 533, in _forward_once\n",
+      "    x = m(x)  # run\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1511, in _wrapped_call_impl\n",
+      "    return self._call_impl(*args, **kwargs)\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1520, in _call_impl\n",
+      "    return forward_call(*args, **kwargs)\n",
+      "  File \"/workspace/yolov9/models/yolo.py\", line 225, in forward\n",
+      "    d1.append(torch.cat((self.cv2[i](x[i]), self.cv3[i](x[i])), 1))\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1511, in _wrapped_call_impl\n",
+      "    return self._call_impl(*args, **kwargs)\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1520, in _call_impl\n",
+      "    return forward_call(*args, **kwargs)\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/container.py\", line 217, in forward\n",
+      "    input = module(input)\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1511, in _wrapped_call_impl\n",
+      "    return self._call_impl(*args, **kwargs)\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1520, in _call_impl\n",
+      "    return forward_call(*args, **kwargs)\n",
+      "  File \"/workspace/yolov9/models/common.py\", line 54, in forward\n",
+      "    return self.act(self.bn(self.conv(x)))\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1511, in _wrapped_call_impl\n",
+      "    return self._call_impl(*args, **kwargs)\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1520, in _call_impl\n",
+      "    return forward_call(*args, **kwargs)\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/batchnorm.py\", line 767, in forward\n",
+      "    return sync_batch_norm.apply(\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/autograd/function.py\", line 553, in apply\n",
+      "    return super().apply(*args, **kwargs)  # type: ignore[misc]\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/_functions.py\", line 89, in forward\n",
+      "    counts = count_all.view(-1)\n",
+      "KeyboardInterrupt\n",
+      "Traceback (most recent call last):\n",
+      "  File \"/workspace/yolov9/train_dual.py\", line 644, in <module>\n",
+      "    main(opt)\n",
+      "  File \"/workspace/yolov9/train_dual.py\", line 538, in main\n",
+      "    train(opt.hyp, opt, device, callbacks)\n",
+      "  File \"/workspace/yolov9/train_dual.py\", line 314, in train\n",
+      "    pred = model(imgs)  # forward\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1511, in _wrapped_call_impl\n",
+      "    return self._call_impl(*args, **kwargs)\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1520, in _call_impl\n",
+      "    return forward_call(*args, **kwargs)\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/parallel/distributed.py\", line 1523, in forward\n",
+      "    else self._run_ddp_forward(*inputs, **kwargs)\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/parallel/distributed.py\", line 1359, in _run_ddp_forward\n",
+      "    return self.module(*inputs, **kwargs)  # type: ignore[index]\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1511, in _wrapped_call_impl\n",
+      "    return self._call_impl(*args, **kwargs)\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1520, in _call_impl\n",
+      "    return forward_call(*args, **kwargs)\n",
+      "  File \"/workspace/yolov9/models/yolo.py\", line 633, in forward\n",
+      "    return self._forward_once(x, profile, visualize)  # single-scale inference, train\n",
+      "  File \"/workspace/yolov9/models/yolo.py\", line 533, in _forward_once\n",
+      "    x = m(x)  # run\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1511, in _wrapped_call_impl\n",
+      "    return self._call_impl(*args, **kwargs)\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1520, in _call_impl\n",
+      "    return forward_call(*args, **kwargs)\n",
+      "  File \"/workspace/yolov9/models/yolo.py\", line 225, in forward\n",
+      "    d1.append(torch.cat((self.cv2[i](x[i]), self.cv3[i](x[i])), 1))\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1511, in _wrapped_call_impl\n",
+      "    return self._call_impl(*args, **kwargs)\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1520, in _call_impl\n",
+      "    return forward_call(*args, **kwargs)\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/container.py\", line 217, in forward\n",
+      "    input = module(input)\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1511, in _wrapped_call_impl\n",
+      "    return self._call_impl(*args, **kwargs)\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1520, in _call_impl\n",
+      "    return forward_call(*args, **kwargs)\n",
+      "  File \"/workspace/yolov9/models/common.py\", line 54, in forward\n",
+      "    return self.act(self.bn(self.conv(x)))\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1511, in _wrapped_call_impl\n",
+      "    return self._call_impl(*args, **kwargs)\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py\", line 1520, in _call_impl\n",
+      "    return forward_call(*args, **kwargs)\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/batchnorm.py\", line 767, in forward\n",
+      "    return sync_batch_norm.apply(\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/autograd/function.py\", line 553, in apply\n",
+      "    return super().apply(*args, **kwargs)  # type: ignore[misc]\n",
+      "  File \"/opt/conda/lib/python3.10/site-packages/torch/nn/modules/_functions.py\", line 89, in forward\n",
+      "    counts = count_all.view(-1)\n",
+      "KeyboardInterrupt\n"
+     ]
+    }
+   ],
+   "source": [
+    "!torchrun  --nproc_per_node 2 --master_port 9527 train_dual.py  \\\n",
+    "--workers 8 --device 0,1 --sync-bn --batch 30 --data '/workspace/road damage/data.yaml'  \\\n",
+    "--img 640 --cfg models/detect/yolov9-c.yaml --weights '' --name yolov9-c --hyp hyp.scratch-high.yaml \\\n",
+    "--min-items 0 --epochs 500 --close-mosaic 15"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "!torchrun  --nproc_per_node 2 --master_port 9527 train_dual.py  \\\n",
+    "--workers 8 --device 0,1 --sync-bn --batch 30 --data '/workspace/road damage/data.yaml'  \\\n",
+    "--img 640 --cfg models/detect/yolov9-c.yaml --weights '' --name yolov9-c --hyp hyp.scratch-high.yaml \\\n",
+    "--min-items 0 --epochs 500 --close-mosaic 15"
+   ]
+  }
+ ],
+ "metadata": {
+  "accelerator": "GPU",
+  "colab": {
+   "gpuType": "L4",
+   "machine_shape": "hm",
+   "provenance": []
+  },
+  "kernelspec": {
+   "display_name": "Python 3 (ipykernel)",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.13"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 4
+}

panoptic/predict.py

@@ -0,0 +1,246 @@
+import argparse
+import os
+import platform
+import sys
+from pathlib import Path
+
+import torch
+
+FILE = Path(__file__).resolve()
+ROOT = FILE.parents[1]  # YOLO root directory
+if str(ROOT) not in sys.path:
+    sys.path.append(str(ROOT))  # add ROOT to PATH
+ROOT = Path(os.path.relpath(ROOT, Path.cwd()))  # relative
+
+from models.common import DetectMultiBackend
+from utils.dataloaders import IMG_FORMATS, VID_FORMATS, LoadImages, LoadScreenshots, LoadStreams
+from utils.general import (LOGGER, Profile, check_file, check_img_size, check_imshow, check_requirements, colorstr, cv2,
+                           increment_path, non_max_suppression, print_args, scale_boxes, scale_segments,
+                           strip_optimizer, xyxy2xywh)
+from utils.plots import Annotator, colors, save_one_box
+from utils.segment.general import masks2segments, process_mask
+from utils.torch_utils import select_device, smart_inference_mode
+
+
+@smart_inference_mode()
+def run(
+    weights=ROOT / 'yolo-pan.pt',  # model.pt path(s)
+    source=ROOT / 'data/images',  # file/dir/URL/glob/screen/0(webcam)
+    data=ROOT / 'data/coco128.yaml',  # dataset.yaml path
+    imgsz=(640, 640),  # inference size (height, width)
+    conf_thres=0.25,  # confidence threshold
+    iou_thres=0.45,  # NMS IOU threshold
+    max_det=1000,  # maximum detections per image
+    device='',  # cuda device, i.e. 0 or 0,1,2,3 or cpu
+    view_img=False,  # show results
+    save_txt=False,  # save results to *.txt
+    save_conf=False,  # save confidences in --save-txt labels
+    save_crop=False,  # save cropped prediction boxes
+    nosave=False,  # do not save images/videos
+    classes=None,  # filter by class: --class 0, or --class 0 2 3
+    agnostic_nms=False,  # class-agnostic NMS
+    augment=False,  # augmented inference
+    visualize=False,  # visualize features
+    update=False,  # update all models
+    project=ROOT / 'runs/predict-seg',  # save results to project/name
+    name='exp',  # save results to project/name
+    exist_ok=False,  # existing project/name ok, do not increment
+    line_thickness=3,  # bounding box thickness (pixels)
+    hide_labels=False,  # hide labels
+    hide_conf=False,  # hide confidences
+    half=False,  # use FP16 half-precision inference
+    dnn=False,  # use OpenCV DNN for ONNX inference
+    vid_stride=1,  # video frame-rate stride
+    retina_masks=False,
+):
+    source = str(source)
+    save_img = not nosave and not source.endswith('.txt')  # save inference images
+    is_file = Path(source).suffix[1:] in (IMG_FORMATS + VID_FORMATS)
+    is_url = source.lower().startswith(('rtsp://', 'rtmp://', 'http://', 'https://'))
+    webcam = source.isnumeric() or source.endswith('.txt') or (is_url and not is_file)
+    screenshot = source.lower().startswith('screen')
+    if is_url and is_file:
+        source = check_file(source)  # download
+
+    # Directories
+    save_dir = increment_path(Path(project) / name, exist_ok=exist_ok)  # increment run
+    (save_dir / 'labels' if save_txt else save_dir).mkdir(parents=True, exist_ok=True)  # make dir
+
+    # Load model
+    device = select_device(device)
+    model = DetectMultiBackend(weights, device=device, dnn=dnn, data=data, fp16=half)
+    stride, names, pt = model.stride, model.names, model.pt
+    imgsz = check_img_size(imgsz, s=stride)  # check image size
+
+    # Dataloader
+    bs = 1  # batch_size
+    if webcam:
+        view_img = check_imshow(warn=True)
+        dataset = LoadStreams(source, img_size=imgsz, stride=stride, auto=pt, vid_stride=vid_stride)
+        bs = len(dataset)
+    elif screenshot:
+        dataset = LoadScreenshots(source, img_size=imgsz, stride=stride, auto=pt)
+    else:
+        dataset = LoadImages(source, img_size=imgsz, stride=stride, auto=pt, vid_stride=vid_stride)
+    vid_path, vid_writer = [None] * bs, [None] * bs
+
+    # Run inference
+    model.warmup(imgsz=(1 if pt else bs, 3, *imgsz))  # warmup
+    seen, windows, dt = 0, [], (Profile(), Profile(), Profile())
+    for path, im, im0s, vid_cap, s in dataset:
+        with dt[0]:
+            im = torch.from_numpy(im).to(model.device)
+            im = im.half() if model.fp16 else im.float()  # uint8 to fp16/32
+            im /= 255  # 0 - 255 to 0.0 - 1.0
+            if len(im.shape) == 3:
+                im = im[None]  # expand for batch dim
+
+        # Inference
+        with dt[1]:
+            visualize = increment_path(save_dir / Path(path).stem, mkdir=True) if visualize else False
+            pred, proto = model(im, augment=augment, visualize=visualize)[:2]
+
+        # NMS
+        with dt[2]:
+            pred = non_max_suppression(pred, conf_thres, iou_thres, classes, agnostic_nms, max_det=max_det, nm=32)
+
+        # Second-stage classifier (optional)
+        # pred = utils.general.apply_classifier(pred, classifier_model, im, im0s)
+
+        # Process predictions
+        for i, det in enumerate(pred):  # per image
+            seen += 1
+            if webcam:  # batch_size >= 1
+                p, im0, frame = path[i], im0s[i].copy(), dataset.count
+                s += f'{i}: '
+            else:
+                p, im0, frame = path, im0s.copy(), getattr(dataset, 'frame', 0)
+
+            p = Path(p)  # to Path
+            save_path = str(save_dir / p.name)  # im.jpg
+            txt_path = str(save_dir / 'labels' / p.stem) + ('' if dataset.mode == 'image' else f'_{frame}')  # im.txt
+            s += '%gx%g ' % im.shape[2:]  # print string
+            imc = im0.copy() if save_crop else im0  # for save_crop
+            annotator = Annotator(im0, line_width=line_thickness, example=str(names))
+            if len(det):
+                masks = process_mask(proto[i], det[:, 6:], det[:, :4], im.shape[2:], upsample=True)  # HWC
+                det[:, :4] = scale_boxes(im.shape[2:], det[:, :4], im0.shape).round()  # rescale boxes to im0 size
+
+                # Segments
+                if save_txt:
+                    segments = reversed(masks2segments(masks))
+                    segments = [scale_segments(im.shape[2:], x, im0.shape, normalize=True) for x in segments]
+
+                # Print results
+                for c in det[:, 5].unique():
+                    n = (det[:, 5] == c).sum()  # detections per class
+                    s += f"{n} {names[int(c)]}{'s' * (n > 1)}, "  # add to string
+
+                # Mask plotting
+                annotator.masks(masks,
+                                colors=[colors(x, True) for x in det[:, 5]],
+                                im_gpu=None if retina_masks else im[i])
+
+                # Write results
+                for j, (*xyxy, conf, cls) in enumerate(reversed(det[:, :6])):
+                    if save_txt:  # Write to file
+                        segj = segments[j].reshape(-1)  # (n,2) to (n*2)
+                        line = (cls, *segj, conf) if save_conf else (cls, *segj)  # label format
+                        with open(f'{txt_path}.txt', 'a') as f:
+                            f.write(('%g ' * len(line)).rstrip() % line + '\n')
+
+                    if save_img or save_crop or view_img:  # Add bbox to image
+                        c = int(cls)  # integer class
+                        label = None if hide_labels else (names[c] if hide_conf else f'{names[c]} {conf:.2f}')
+                        annotator.box_label(xyxy, label, color=colors(c, True))
+                        # annotator.draw.polygon(segments[j], outline=colors(c, True), width=3)
+                    if save_crop:
+                        save_one_box(xyxy, imc, file=save_dir / 'crops' / names[c] / f'{p.stem}.jpg', BGR=True)
+
+            # Stream results
+            im0 = annotator.result()
+            if view_img:
+                if platform.system() == 'Linux' and p not in windows:
+                    windows.append(p)
+                    cv2.namedWindow(str(p), cv2.WINDOW_NORMAL | cv2.WINDOW_KEEPRATIO)  # allow window resize (Linux)
+                    cv2.resizeWindow(str(p), im0.shape[1], im0.shape[0])
+                cv2.imshow(str(p), im0)
+                if cv2.waitKey(1) == ord('q'):  # 1 millisecond
+                    exit()
+
+            # Save results (image with detections)
+            if save_img:
+                if dataset.mode == 'image':
+                    cv2.imwrite(save_path, im0)
+                else:  # 'video' or 'stream'
+                    if vid_path[i] != save_path:  # new video
+                        vid_path[i] = save_path
+                        if isinstance(vid_writer[i], cv2.VideoWriter):
+                            vid_writer[i].release()  # release previous video writer
+                        if vid_cap:  # video
+                            fps = vid_cap.get(cv2.CAP_PROP_FPS)
+                            w = int(vid_cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+                            h = int(vid_cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+                        else:  # stream
+                            fps, w, h = 30, im0.shape[1], im0.shape[0]
+                        save_path = str(Path(save_path).with_suffix('.mp4'))  # force *.mp4 suffix on results videos
+                        vid_writer[i] = cv2.VideoWriter(save_path, cv2.VideoWriter_fourcc(*'mp4v'), fps, (w, h))
+                    vid_writer[i].write(im0)
+
+        # Print time (inference-only)
+        LOGGER.info(f"{s}{'' if len(det) else '(no detections), '}{dt[1].dt * 1E3:.1f}ms")
+
+    # Print results
+    t = tuple(x.t / seen * 1E3 for x in dt)  # speeds per image
+    LOGGER.info(f'Speed: %.1fms pre-process, %.1fms inference, %.1fms NMS per image at shape {(1, 3, *imgsz)}' % t)
+    if save_txt or save_img:
+        s = f"\n{len(list(save_dir.glob('labels/*.txt')))} labels saved to {save_dir / 'labels'}" if save_txt else ''
+        LOGGER.info(f"Results saved to {colorstr('bold', save_dir)}{s}")
+    if update:
+        strip_optimizer(weights[0])  # update model (to fix SourceChangeWarning)
+
+
+def parse_opt():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--weights', nargs='+', type=str, default=ROOT / 'yolo-pan.pt', help='model path(s)')
+    parser.add_argument('--source', type=str, default=ROOT / 'data/images', help='file/dir/URL/glob/screen/0(webcam)')
+    parser.add_argument('--data', type=str, default=ROOT / 'data/coco128.yaml', help='(optional) dataset.yaml path')
+    parser.add_argument('--imgsz', '--img', '--img-size', nargs='+', type=int, default=[640], help='inference size h,w')
+    parser.add_argument('--conf-thres', type=float, default=0.25, help='confidence threshold')
+    parser.add_argument('--iou-thres', type=float, default=0.45, help='NMS IoU threshold')
+    parser.add_argument('--max-det', type=int, default=1000, help='maximum detections per image')
+    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
+    parser.add_argument('--view-img', action='store_true', help='show results')
+    parser.add_argument('--save-txt', action='store_true', help='save results to *.txt')
+    parser.add_argument('--save-conf', action='store_true', help='save confidences in --save-txt labels')
+    parser.add_argument('--save-crop', action='store_true', help='save cropped prediction boxes')
+    parser.add_argument('--nosave', action='store_true', help='do not save images/videos')
+    parser.add_argument('--classes', nargs='+', type=int, help='filter by class: --classes 0, or --classes 0 2 3')
+    parser.add_argument('--agnostic-nms', action='store_true', help='class-agnostic NMS')
+    parser.add_argument('--augment', action='store_true', help='augmented inference')
+    parser.add_argument('--visualize', action='store_true', help='visualize features')
+    parser.add_argument('--update', action='store_true', help='update all models')
+    parser.add_argument('--project', default=ROOT / 'runs/predict-seg', help='save results to project/name')
+    parser.add_argument('--name', default='exp', help='save results to project/name')
+    parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')
+    parser.add_argument('--line-thickness', default=3, type=int, help='bounding box thickness (pixels)')
+    parser.add_argument('--hide-labels', default=False, action='store_true', help='hide labels')
+    parser.add_argument('--hide-conf', default=False, action='store_true', help='hide confidences')
+    parser.add_argument('--half', action='store_true', help='use FP16 half-precision inference')
+    parser.add_argument('--dnn', action='store_true', help='use OpenCV DNN for ONNX inference')
+    parser.add_argument('--vid-stride', type=int, default=1, help='video frame-rate stride')
+    parser.add_argument('--retina-masks', action='store_true', help='whether to plot masks in native resolution')
+    opt = parser.parse_args()
+    opt.imgsz *= 2 if len(opt.imgsz) == 1 else 1  # expand
+    print_args(vars(opt))
+    return opt
+
+
+def main(opt):
+    check_requirements(exclude=('tensorboard', 'thop'))
+    run(**vars(opt))
+
+
+if __name__ == "__main__":
+    opt = parse_opt()
+    main(opt)

panoptic/train.py

@@ -0,0 +1,662 @@
+import argparse
+import math
+import os
+import random
+import sys
+import time
+from copy import deepcopy
+from datetime import datetime
+from pathlib import Path
+
+import numpy as np
+import torch
+import torch.distributed as dist
+import torch.nn as nn
+import yaml
+from torch.optim import lr_scheduler
+from tqdm import tqdm
+
+FILE = Path(__file__).resolve()
+ROOT = FILE.parents[1]  # YOLO root directory
+if str(ROOT) not in sys.path:
+    sys.path.append(str(ROOT))  # add ROOT to PATH
+ROOT = Path(os.path.relpath(ROOT, Path.cwd()))  # relative
+
+import panoptic.val as validate  # for end-of-epoch mAP
+from models.experimental import attempt_load
+from models.yolo import SegmentationModel
+from utils.autoanchor import check_anchors
+from utils.autobatch import check_train_batch_size
+from utils.callbacks import Callbacks
+from utils.downloads import attempt_download, is_url
+from utils.general import (LOGGER, TQDM_BAR_FORMAT, check_amp, check_dataset, check_file, check_git_info,
+                           check_git_status, check_img_size, check_requirements, check_suffix, check_yaml, colorstr,
+                           get_latest_run, increment_path, init_seeds, intersect_dicts, labels_to_class_weights,
+                           labels_to_image_weights, one_cycle, one_flat_cycle, print_args, print_mutation, strip_optimizer, yaml_save)
+from utils.loggers import GenericLogger
+from utils.plots import plot_evolve, plot_labels
+from utils.panoptic.dataloaders import create_dataloader
+from utils.panoptic.loss_tal import ComputeLoss
+from utils.panoptic.metrics import KEYS, fitness
+from utils.panoptic.plots import plot_images_and_masks, plot_results_with_masks
+from utils.torch_utils import (EarlyStopping, ModelEMA, de_parallel, select_device, smart_DDP, smart_optimizer,
+                               smart_resume, torch_distributed_zero_first)
+
+LOCAL_RANK = int(os.getenv('LOCAL_RANK', -1))  # https://pytorch.org/docs/stable/elastic/run.html
+RANK = int(os.getenv('RANK', -1))
+WORLD_SIZE = int(os.getenv('WORLD_SIZE', 1))
+GIT_INFO = None#check_git_info()
+
+
+def train(hyp, opt, device, callbacks):  # hyp is path/to/hyp.yaml or hyp dictionary
+    save_dir, epochs, batch_size, weights, single_cls, evolve, data, cfg, resume, noval, nosave, workers, freeze, mask_ratio = \
+        Path(opt.save_dir), opt.epochs, opt.batch_size, opt.weights, opt.single_cls, opt.evolve, opt.data, opt.cfg, \
+        opt.resume, opt.noval, opt.nosave, opt.workers, opt.freeze, opt.mask_ratio
+    # callbacks.run('on_pretrain_routine_start')
+
+    # Directories
+    w = save_dir / 'weights'  # weights dir
+    (w.parent if evolve else w).mkdir(parents=True, exist_ok=True)  # make dir
+    last, best = w / 'last.pt', w / 'best.pt'
+
+    # Hyperparameters
+    if isinstance(hyp, str):
+        with open(hyp, errors='ignore') as f:
+            hyp = yaml.safe_load(f)  # load hyps dict
+    LOGGER.info(colorstr('hyperparameters: ') + ', '.join(f'{k}={v}' for k, v in hyp.items()))
+    opt.hyp = hyp.copy()  # for saving hyps to checkpoints
+
+    # Save run settings
+    if not evolve:
+        yaml_save(save_dir / 'hyp.yaml', hyp)
+        yaml_save(save_dir / 'opt.yaml', vars(opt))
+
+    # Loggers
+    data_dict = None
+    if RANK in {-1, 0}:
+        logger = GenericLogger(opt=opt, console_logger=LOGGER)
+
+    # Config
+    plots = not evolve and not opt.noplots  # create plots
+    overlap = not opt.no_overlap
+    cuda = device.type != 'cpu'
+    init_seeds(opt.seed + 1 + RANK, deterministic=True)
+    with torch_distributed_zero_first(LOCAL_RANK):
+        data_dict = data_dict or check_dataset(data)  # check if None
+    train_path, val_path = data_dict['train'], data_dict['val']
+    nc = 1 if single_cls else int(data_dict['nc'])  # number of classes
+    names = {0: 'item'} if single_cls and len(data_dict['names']) != 1 else data_dict['names']  # class names
+    #is_coco = isinstance(val_path, str) and val_path.endswith('coco/val2017.txt')  # COCO dataset
+    is_coco = isinstance(val_path, str) and val_path.endswith('val2017.txt')  # COCO dataset
+
+    # Model
+    check_suffix(weights, '.pt')  # check weights
+    pretrained = weights.endswith('.pt')
+    if pretrained:
+        with torch_distributed_zero_first(LOCAL_RANK):
+            weights = attempt_download(weights)  # download if not found locally
+        ckpt = torch.load(weights, map_location='cpu')  # load checkpoint to CPU to avoid CUDA memory leak
+        model = SegmentationModel(cfg or ckpt['model'].yaml, ch=3, nc=nc).to(device)
+        exclude = ['anchor'] if (cfg or hyp.get('anchors')) and not resume else []  # exclude keys
+        csd = ckpt['model'].float().state_dict()  # checkpoint state_dict as FP32
+        csd = intersect_dicts(csd, model.state_dict(), exclude=exclude)  # intersect
+        model.load_state_dict(csd, strict=False)  # load
+        LOGGER.info(f'Transferred {len(csd)}/{len(model.state_dict())} items from {weights}')  # report
+    else:
+        model = SegmentationModel(cfg, ch=3, nc=nc).to(device)  # create
+    amp = check_amp(model)  # check AMP
+
+    # Freeze
+    freeze = [f'model.{x}.' for x in (freeze if len(freeze) > 1 else range(freeze[0]))]  # layers to freeze
+    for k, v in model.named_parameters():
+        #v.requires_grad = True  # train all layers
+        # v.register_hook(lambda x: torch.nan_to_num(x))  # NaN to 0 (commented for erratic training results)
+        if any(x in k for x in freeze):
+            LOGGER.info(f'freezing {k}')
+            v.requires_grad = False
+
+    # Image size
+    gs = max(int(model.stride.max()), 32)  # grid size (max stride)
+    imgsz = check_img_size(opt.imgsz, gs, floor=gs * 2)  # verify imgsz is gs-multiple
+
+    # Batch size
+    if RANK == -1 and batch_size == -1:  # single-GPU only, estimate best batch size
+        batch_size = check_train_batch_size(model, imgsz, amp)
+        logger.update_params({"batch_size": batch_size})
+        # loggers.on_params_update({"batch_size": batch_size})
+
+    # Optimizer
+    nbs = 64  # nominal batch size
+    accumulate = max(round(nbs / batch_size), 1)  # accumulate loss before optimizing
+    hyp['weight_decay'] *= batch_size * accumulate / nbs  # scale weight_decay
+    optimizer = smart_optimizer(model, opt.optimizer, hyp['lr0'], hyp['momentum'], hyp['weight_decay'])
+
+    # Scheduler
+    if opt.cos_lr:
+        lf = one_cycle(1, hyp['lrf'], epochs)  # cosine 1->hyp['lrf']
+    elif opt.flat_cos_lr:
+        lf = one_flat_cycle(1, hyp['lrf'], epochs)  # cosine 1->hyp['lrf']
+    elif opt.fixed_lr:
+        lf = lambda x: 1.0
+    elif opt.poly_lr:
+        power = 0.9
+        lf = lambda x: ((1 - (x / epochs)) ** power) * (1.0 - hyp['lrf']) + hyp['lrf']
+    else:
+        lf = lambda x: (1 - x / epochs) * (1.0 - hyp['lrf']) + hyp['lrf']  # linear
+    scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)  # plot_lr_scheduler(optimizer, scheduler, epochs)
+
+    # EMA
+    ema = ModelEMA(model) if RANK in {-1, 0} else None
+
+    # Resume
+    best_fitness, start_epoch = 0.0, 0
+    if pretrained:
+        if resume:
+            best_fitness, start_epoch, epochs = smart_resume(ckpt, optimizer, ema, weights, epochs, resume)
+        del ckpt, csd
+
+    # DP mode
+    if cuda and RANK == -1 and torch.cuda.device_count() > 1:
+        LOGGER.warning('WARNING ⚠️ DP not recommended, use torch.distributed.run for best DDP Multi-GPU results.')
+        model = torch.nn.DataParallel(model)
+
+    # SyncBatchNorm
+    if opt.sync_bn and cuda and RANK != -1:
+        model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device)
+        LOGGER.info('Using SyncBatchNorm()')
+
+    # Trainloader
+    train_loader, dataset = create_dataloader(
+        train_path,
+        imgsz,
+        batch_size // WORLD_SIZE,
+        gs,
+        single_cls,
+        hyp=hyp,
+        augment=True,
+        cache=None if opt.cache == 'val' else opt.cache,
+        rect=opt.rect,
+        rank=LOCAL_RANK,
+        workers=workers,
+        image_weights=opt.image_weights,
+        close_mosaic=opt.close_mosaic != 0,
+        quad=opt.quad,
+        prefix=colorstr('train: '),
+        shuffle=True,
+        mask_downsample_ratio=mask_ratio,
+        overlap_mask=overlap,
+    )
+    labels = np.concatenate(dataset.labels, 0)
+    mlc = int(labels[:, 0].max())  # max label class
+    assert mlc < nc, f'Label class {mlc} exceeds nc={nc} in {data}. Possible class labels are 0-{nc - 1}'
+
+    # Process 0
+    if RANK in {-1, 0}:
+        val_loader = create_dataloader(val_path,
+                                       imgsz,
+                                       batch_size // WORLD_SIZE * 2,
+                                       gs,
+                                       single_cls,
+                                       hyp=hyp,
+                                       cache=None if noval else opt.cache,
+                                       rect=True,
+                                       rank=-1,
+                                       workers=workers * 2,
+                                       pad=0.5,
+                                       mask_downsample_ratio=mask_ratio,
+                                       overlap_mask=overlap,
+                                       prefix=colorstr('val: '))[0]
+
+        if not resume:
+            #if not opt.noautoanchor:
+            #    check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz)  # run AutoAnchor
+            model.half().float()  # pre-reduce anchor precision
+
+            if plots:
+                plot_labels(labels, names, save_dir)
+        # callbacks.run('on_pretrain_routine_end', labels, names)
+
+    # DDP mode
+    if cuda and RANK != -1:
+        model = smart_DDP(model)
+
+    # Model attributes
+    nl = de_parallel(model).model[-1].nl  # number of detection layers (to scale hyps)
+    #hyp['box'] *= 3 / nl  # scale to layers
+    #hyp['cls'] *= nc / 80 * 3 / nl  # scale to classes and layers
+    #hyp['obj'] *= (imgsz / 640) ** 2 * 3 / nl  # scale to image size and layers
+    hyp['label_smoothing'] = opt.label_smoothing
+    model.nc = nc  # attach number of classes to model
+    model.hyp = hyp  # attach hyperparameters to model
+    model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) * nc  # attach class weights
+    model.names = names
+
+    # Start training
+    t0 = time.time()
+    nb = len(train_loader)  # number of batches
+    nw = max(round(hyp['warmup_epochs'] * nb), 100)  # number of warmup iterations, max(3 epochs, 100 iterations)
+    # nw = min(nw, (epochs - start_epoch) / 2 * nb)  # limit warmup to < 1/2 of training
+    last_opt_step = -1
+    maps = np.zeros(nc)  # mAP per class
+    results = (0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0)  # P, R, [email protected], [email protected], val_loss(box, obj, cls)
+    scheduler.last_epoch = start_epoch - 1  # do not move
+    scaler = torch.cuda.amp.GradScaler(enabled=amp)
+    stopper, stop = EarlyStopping(patience=opt.patience), False
+    compute_loss = ComputeLoss(model, overlap=overlap)  # init loss class
+    # callbacks.run('on_train_start')
+    LOGGER.info(f'Image sizes {imgsz} train, {imgsz} val\n'
+                f'Using {train_loader.num_workers * WORLD_SIZE} dataloader workers\n'
+                f"Logging results to {colorstr('bold', save_dir)}\n"
+                f'Starting training for {epochs} epochs...')
+    for epoch in range(start_epoch, epochs):  # epoch ------------------------------------------------------------------
+        # callbacks.run('on_train_epoch_start')
+        model.train()
+
+        # Update image weights (optional, single-GPU only)
+        if opt.image_weights:
+            cw = model.class_weights.cpu().numpy() * (1 - maps) ** 2 / nc  # class weights
+            iw = labels_to_image_weights(dataset.labels, nc=nc, class_weights=cw)  # image weights
+            dataset.indices = random.choices(range(dataset.n), weights=iw, k=dataset.n)  # rand weighted idx
+        if epoch == (epochs - opt.close_mosaic):
+            LOGGER.info("Closing dataloader mosaic")
+            dataset.mosaic = False
+
+        # Update mosaic border (optional)
+        # b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs)
+        # dataset.mosaic_border = [b - imgsz, -b]  # height, width borders
+
+        mloss = torch.zeros(6, device=device)  # mean losses
+        if RANK != -1:
+            train_loader.sampler.set_epoch(epoch)
+        pbar = enumerate(train_loader)
+        LOGGER.info(('\n' + '%11s' * 10) %
+                    ('Epoch', 'GPU_mem', 'box_loss', 'seg_loss', 'cls_loss', 'dfl_loss', 'fcl_loss', 'dic_loss', 'Instances', 'Size'))
+        if RANK in {-1, 0}:
+            pbar = tqdm(pbar, total=nb, bar_format=TQDM_BAR_FORMAT)  # progress bar
+        optimizer.zero_grad()
+        for i, (imgs, targets, paths, _, masks, semasks) in pbar:  # batch ------------------------------------------------------
+            # callbacks.run('on_train_batch_start')
+            #print(imgs.shape)
+            #print(semasks.shape)
+            #print(masks.shape)
+            ni = i + nb * epoch  # number integrated batches (since train start)
+            imgs = imgs.to(device, non_blocking=True).float() / 255  # uint8 to float32, 0-255 to 0.0-1.0
+
+            # Warmup
+            if ni <= nw:
+                xi = [0, nw]  # x interp
+                # compute_loss.gr = np.interp(ni, xi, [0.0, 1.0])  # iou loss ratio (obj_loss = 1.0 or iou)
+                accumulate = max(1, np.interp(ni, xi, [1, nbs / batch_size]).round())
+                for j, x in enumerate(optimizer.param_groups):
+                    # bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
+                    x['lr'] = np.interp(ni, xi, [hyp['warmup_bias_lr'] if j == 0 else 0.0, x['initial_lr'] * lf(epoch)])
+                    if 'momentum' in x:
+                        x['momentum'] = np.interp(ni, xi, [hyp['warmup_momentum'], hyp['momentum']])
+
+            # Multi-scale
+            if opt.multi_scale:
+                sz = random.randrange(imgsz * 0.5, imgsz * 1.5 + gs) // gs * gs  # size
+                sf = sz / max(imgs.shape[2:])  # scale factor
+                if sf != 1:
+                    ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]]  # new shape (stretched to gs-multiple)
+                    imgs = nn.functional.interpolate(imgs, size=ns, mode='bilinear', align_corners=False)
+
+            # Forward
+            with torch.cuda.amp.autocast(amp):
+                pred = model(imgs)  # forward
+                loss, loss_items = compute_loss(pred, targets.to(device), masks=masks.to(device).float(),
+                                                semasks=semasks.to(device).float())
+                if RANK != -1:
+                    loss *= WORLD_SIZE  # gradient averaged between devices in DDP mode
+                if opt.quad:
+                    loss *= 4.
+
+            # Backward
+            torch.use_deterministic_algorithms(False)
+            scaler.scale(loss).backward()
+
+            # Optimize - https://pytorch.org/docs/master/notes/amp_examples.html
+            if ni - last_opt_step >= accumulate:
+                scaler.unscale_(optimizer)  # unscale gradients
+                torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=10.0)  # clip gradients
+                scaler.step(optimizer)  # optimizer.step
+                scaler.update()
+                optimizer.zero_grad()
+                if ema:
+                    ema.update(model)
+                last_opt_step = ni
+
+            # Log
+            if RANK in {-1, 0}:
+                mloss = (mloss * i + loss_items) / (i + 1)  # update mean losses
+                mem = f'{torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0:.3g}G'  # (GB)
+                pbar.set_description(('%11s' * 2 + '%11.4g' * 8) %
+                                     (f'{epoch}/{epochs - 1}', mem, *mloss, targets.shape[0], imgs.shape[-1]))
+                # callbacks.run('on_train_batch_end', model, ni, imgs, targets, paths)
+                # if callbacks.stop_training:
+                #    return
+
+                # Mosaic plots
+                if plots:
+                    if ni < 10:
+                        plot_images_and_masks(imgs, targets, masks, semasks, paths, save_dir / f"train_batch{ni}.jpg")
+                    if ni == 10:
+                        files = sorted(save_dir.glob('train*.jpg'))
+                        logger.log_images(files, "Mosaics", epoch)
+            # end batch ------------------------------------------------------------------------------------------------
+
+        # Scheduler
+        lr = [x['lr'] for x in optimizer.param_groups]  # for loggers
+        scheduler.step()
+
+        if RANK in {-1, 0}:
+            # mAP
+            # callbacks.run('on_train_epoch_end', epoch=epoch)
+            ema.update_attr(model, include=['yaml', 'nc', 'hyp', 'names', 'stride', 'class_weights'])
+            final_epoch = (epoch + 1 == epochs) or stopper.possible_stop
+            if not noval or final_epoch:  # Calculate mAP
+                if (opt.save_period > 0 and epoch % opt.save_period == 0) or (epoch > (epochs - 2 * opt.close_mosaic)):
+                    results, maps, _ = validate.run(data_dict,
+                                                    batch_size=batch_size // WORLD_SIZE * 2,
+                                                    imgsz=imgsz,
+                                                    half=amp,
+                                                    model=ema.ema,
+                                                    single_cls=single_cls,
+                                                    dataloader=val_loader,
+                                                    save_dir=save_dir,
+                                                    plots=False,
+                                                    callbacks=callbacks,
+                                                    compute_loss=compute_loss,
+                                                    mask_downsample_ratio=mask_ratio,
+                                                    overlap=overlap)
+
+            # Update best mAP
+            fi = fitness(np.array(results).reshape(1, -1))  # weighted combination of [P, R, [email protected], [email protected]]
+            stop = stopper(epoch=epoch, fitness=fi)  # early stop check
+            if fi > best_fitness:
+                best_fitness = fi
+            log_vals = list(mloss) + list(results) + lr
+            # callbacks.run('on_fit_epoch_end', log_vals, epoch, best_fitness, fi)
+            # Log val metrics and media
+            metrics_dict = dict(zip(KEYS, log_vals))
+            logger.log_metrics(metrics_dict, epoch)
+
+            # Save model
+            if (not nosave) or (final_epoch and not evolve):  # if save
+                ckpt = {
+                    'epoch': epoch,
+                    'best_fitness': best_fitness,
+                    'model': deepcopy(de_parallel(model)).half(),
+                    'ema': deepcopy(ema.ema).half(),
+                    'updates': ema.updates,
+                    'optimizer': optimizer.state_dict(),
+                    'opt': vars(opt),
+                    'git': GIT_INFO,  # {remote, branch, commit} if a git repo
+                    'date': datetime.now().isoformat()}
+
+                # Save last, best and delete
+                torch.save(ckpt, last)
+                if best_fitness == fi:
+                    torch.save(ckpt, best)
+                if opt.save_period > 0 and epoch % opt.save_period == 0:
+                    torch.save(ckpt, w / f'epoch{epoch}.pt')
+                    logger.log_model(w / f'epoch{epoch}.pt')
+                del ckpt
+                # callbacks.run('on_model_save', last, epoch, final_epoch, best_fitness, fi)
+
+        # EarlyStopping
+        if RANK != -1:  # if DDP training
+            broadcast_list = [stop if RANK == 0 else None]
+            dist.broadcast_object_list(broadcast_list, 0)  # broadcast 'stop' to all ranks
+            if RANK != 0:
+                stop = broadcast_list[0]
+        if stop:
+            break  # must break all DDP ranks
+
+        # end epoch ----------------------------------------------------------------------------------------------------
+    # end training -----------------------------------------------------------------------------------------------------
+    if RANK in {-1, 0}:
+        LOGGER.info(f'\n{epoch - start_epoch + 1} epochs completed in {(time.time() - t0) / 3600:.3f} hours.')
+        for f in last, best:
+            if f.exists():
+                strip_optimizer(f)  # strip optimizers
+                if f is best:
+                    LOGGER.info(f'\nValidating {f}...')
+                    results, _, _ = validate.run(
+                        data_dict,
+                        batch_size=batch_size // WORLD_SIZE * 2,
+                        imgsz=imgsz,
+                        model=attempt_load(f, device).half(),
+                        iou_thres=0.65 if is_coco else 0.60,  # best pycocotools at iou 0.65
+                        single_cls=single_cls,
+                        dataloader=val_loader,
+                        save_dir=save_dir,
+                        save_json=is_coco,
+                        verbose=True,
+                        plots=plots,
+                        callbacks=callbacks,
+                        compute_loss=compute_loss,
+                        mask_downsample_ratio=mask_ratio,
+                        overlap=overlap)  # val best model with plots
+                    if is_coco:
+                        # callbacks.run('on_fit_epoch_end', list(mloss) + list(results) + lr, epoch, best_fitness, fi)
+                        metrics_dict = dict(zip(KEYS, list(mloss) + list(results) + lr))
+                        logger.log_metrics(metrics_dict, epoch)
+
+        # callbacks.run('on_train_end', last, best, epoch, results)
+        # on train end callback using genericLogger
+        logger.log_metrics(dict(zip(KEYS[6:22], results)), epochs)
+        if not opt.evolve:
+            logger.log_model(best, epoch)
+        if plots:
+            plot_results_with_masks(file=save_dir / 'results.csv')  # save results.png
+            files = ['results.png', 'confusion_matrix.png', *(f'{x}_curve.png' for x in ('F1', 'PR', 'P', 'R'))]
+            files = [(save_dir / f) for f in files if (save_dir / f).exists()]  # filter
+            LOGGER.info(f"Results saved to {colorstr('bold', save_dir)}")
+            logger.log_images(files, "Results", epoch + 1)
+            logger.log_images(sorted(save_dir.glob('val*.jpg')), "Validation", epoch + 1)
+    torch.cuda.empty_cache()
+    return results
+
+
+def parse_opt(known=False):
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--weights', type=str, default=ROOT / 'yolo-pan.pt', help='initial weights path')
+    parser.add_argument('--cfg', type=str, default='', help='model.yaml path')
+    parser.add_argument('--data', type=str, default=ROOT / 'data/coco128-seg.yaml', help='dataset.yaml path')
+    parser.add_argument('--hyp', type=str, default=ROOT / 'data/hyps/hyp.scratch-low.yaml', help='hyperparameters path')
+    parser.add_argument('--epochs', type=int, default=100, help='total training epochs')
+    parser.add_argument('--batch-size', type=int, default=16, help='total batch size for all GPUs, -1 for autobatch')
+    parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=640, help='train, val image size (pixels)')
+    parser.add_argument('--rect', action='store_true', help='rectangular training')
+    parser.add_argument('--resume', nargs='?', const=True, default=False, help='resume most recent training')
+    parser.add_argument('--nosave', action='store_true', help='only save final checkpoint')
+    parser.add_argument('--noval', action='store_true', help='only validate final epoch')
+    parser.add_argument('--noautoanchor', action='store_true', help='disable AutoAnchor')
+    parser.add_argument('--noplots', action='store_true', help='save no plot files')
+    parser.add_argument('--evolve', type=int, nargs='?', const=300, help='evolve hyperparameters for x generations')
+    parser.add_argument('--bucket', type=str, default='', help='gsutil bucket')
+    parser.add_argument('--cache', type=str, nargs='?', const='ram', help='image --cache ram/disk')
+    parser.add_argument('--image-weights', action='store_true', help='use weighted image selection for training')
+    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
+    parser.add_argument('--multi-scale', action='store_true', help='vary img-size +/- 50%%')
+    parser.add_argument('--single-cls', action='store_true', help='train multi-class data as single-class')
+    parser.add_argument('--optimizer', type=str, choices=['SGD', 'Adam', 'AdamW', 'LION'], default='SGD', help='optimizer')
+    parser.add_argument('--sync-bn', action='store_true', help='use SyncBatchNorm, only available in DDP mode')
+    parser.add_argument('--workers', type=int, default=8, help='max dataloader workers (per RANK in DDP mode)')
+    parser.add_argument('--project', default=ROOT / 'runs/train-pan', help='save to project/name')
+    parser.add_argument('--name', default='exp', help='save to project/name')
+    parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')
+    parser.add_argument('--quad', action='store_true', help='quad dataloader')
+    parser.add_argument('--cos-lr', action='store_true', help='cosine LR scheduler')
+    parser.add_argument('--flat-cos-lr', action='store_true', help='cosine LR scheduler')
+    parser.add_argument('--fixed-lr', action='store_true', help='fixed LR scheduler')
+    parser.add_argument('--poly-lr', action='store_true', help='fixed LR scheduler')
+    parser.add_argument('--label-smoothing', type=float, default=0.0, help='Label smoothing epsilon')
+    parser.add_argument('--patience', type=int, default=100, help='EarlyStopping patience (epochs without improvement)')
+    parser.add_argument('--freeze', nargs='+', type=int, default=[0], help='Freeze layers: backbone=10, first3=0 1 2')
+    parser.add_argument('--save-period', type=int, default=-1, help='Save checkpoint every x epochs (disabled if < 1)')
+    parser.add_argument('--seed', type=int, default=0, help='Global training seed')
+    parser.add_argument('--local_rank', type=int, default=-1, help='Automatic DDP Multi-GPU argument, do not modify')
+    parser.add_argument('--close-mosaic', type=int, default=0, help='Experimental')
+
+    # Instance Segmentation Args
+    parser.add_argument('--mask-ratio', type=int, default=4, help='Downsample the truth masks to saving memory')
+    parser.add_argument('--no-overlap', action='store_true', help='Overlap masks train faster at slightly less mAP')
+
+    return parser.parse_known_args()[0] if known else parser.parse_args()
+
+
+def main(opt, callbacks=Callbacks()):
+    # Checks
+    if RANK in {-1, 0}:
+        print_args(vars(opt))
+        #check_git_status()
+        #check_requirements()
+
+    # Resume
+    if opt.resume and not opt.evolve:  # resume from specified or most recent last.pt
+        last = Path(check_file(opt.resume) if isinstance(opt.resume, str) else get_latest_run())
+        opt_yaml = last.parent.parent / 'opt.yaml'  # train options yaml
+        opt_data = opt.data  # original dataset
+        if opt_yaml.is_file():
+            with open(opt_yaml, errors='ignore') as f:
+                d = yaml.safe_load(f)
+        else:
+            d = torch.load(last, map_location='cpu')['opt']
+        opt = argparse.Namespace(**d)  # replace
+        opt.cfg, opt.weights, opt.resume = '', str(last), True  # reinstate
+        if is_url(opt_data):
+            opt.data = check_file(opt_data)  # avoid HUB resume auth timeout
+    else:
+        opt.data, opt.cfg, opt.hyp, opt.weights, opt.project = \
+            check_file(opt.data), check_yaml(opt.cfg), check_yaml(opt.hyp), str(opt.weights), str(opt.project)  # checks
+        assert len(opt.cfg) or len(opt.weights), 'either --cfg or --weights must be specified'
+        if opt.evolve:
+            if opt.project == str(ROOT / 'runs/train'):  # if default project name, rename to runs/evolve
+                opt.project = str(ROOT / 'runs/evolve')
+            opt.exist_ok, opt.resume = opt.resume, False  # pass resume to exist_ok and disable resume
+        if opt.name == 'cfg':
+            opt.name = Path(opt.cfg).stem  # use model.yaml as name
+        opt.save_dir = str(increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok))
+
+    # DDP mode
+    device = select_device(opt.device, batch_size=opt.batch_size)
+    if LOCAL_RANK != -1:
+        msg = 'is not compatible with YOLO Multi-GPU DDP training'
+        assert not opt.image_weights, f'--image-weights {msg}'
+        assert not opt.evolve, f'--evolve {msg}'
+        assert opt.batch_size != -1, f'AutoBatch with --batch-size -1 {msg}, please pass a valid --batch-size'
+        assert opt.batch_size % WORLD_SIZE == 0, f'--batch-size {opt.batch_size} must be multiple of WORLD_SIZE'
+        assert torch.cuda.device_count() > LOCAL_RANK, 'insufficient CUDA devices for DDP command'
+        torch.cuda.set_device(LOCAL_RANK)
+        device = torch.device('cuda', LOCAL_RANK)
+        dist.init_process_group(backend="nccl" if dist.is_nccl_available() else "gloo")
+
+    # Train
+    if not opt.evolve:
+        train(opt.hyp, opt, device, callbacks)
+
+    # Evolve hyperparameters (optional)
+    else:
+        # Hyperparameter evolution metadata (mutation scale 0-1, lower_limit, upper_limit)
+        meta = {
+            'lr0': (1, 1e-5, 1e-1),  # initial learning rate (SGD=1E-2, Adam=1E-3)
+            'lrf': (1, 0.01, 1.0),  # final OneCycleLR learning rate (lr0 * lrf)
+            'momentum': (0.3, 0.6, 0.98),  # SGD momentum/Adam beta1
+            'weight_decay': (1, 0.0, 0.001),  # optimizer weight decay
+            'warmup_epochs': (1, 0.0, 5.0),  # warmup epochs (fractions ok)
+            'warmup_momentum': (1, 0.0, 0.95),  # warmup initial momentum
+            'warmup_bias_lr': (1, 0.0, 0.2),  # warmup initial bias lr
+            'box': (1, 0.02, 0.2),  # box loss gain
+            'cls': (1, 0.2, 4.0),  # cls loss gain
+            'cls_pw': (1, 0.5, 2.0),  # cls BCELoss positive_weight
+            'obj': (1, 0.2, 4.0),  # obj loss gain (scale with pixels)
+            'obj_pw': (1, 0.5, 2.0),  # obj BCELoss positive_weight
+            'iou_t': (0, 0.1, 0.7),  # IoU training threshold
+            'anchor_t': (1, 2.0, 8.0),  # anchor-multiple threshold
+            'anchors': (2, 2.0, 10.0),  # anchors per output grid (0 to ignore)
+            'fl_gamma': (0, 0.0, 2.0),  # focal loss gamma (efficientDet default gamma=1.5)
+            'hsv_h': (1, 0.0, 0.1),  # image HSV-Hue augmentation (fraction)
+            'hsv_s': (1, 0.0, 0.9),  # image HSV-Saturation augmentation (fraction)
+            'hsv_v': (1, 0.0, 0.9),  # image HSV-Value augmentation (fraction)
+            'degrees': (1, 0.0, 45.0),  # image rotation (+/- deg)
+            'translate': (1, 0.0, 0.9),  # image translation (+/- fraction)
+            'scale': (1, 0.0, 0.9),  # image scale (+/- gain)
+            'shear': (1, 0.0, 10.0),  # image shear (+/- deg)
+            'perspective': (0, 0.0, 0.001),  # image perspective (+/- fraction), range 0-0.001
+            'flipud': (1, 0.0, 1.0),  # image flip up-down (probability)
+            'fliplr': (0, 0.0, 1.0),  # image flip left-right (probability)
+            'mosaic': (1, 0.0, 1.0),  # image mixup (probability)
+            'mixup': (1, 0.0, 1.0),  # image mixup (probability)
+            'copy_paste': (1, 0.0, 1.0)}  # segment copy-paste (probability)
+
+        with open(opt.hyp, errors='ignore') as f:
+            hyp = yaml.safe_load(f)  # load hyps dict
+            if 'anchors' not in hyp:  # anchors commented in hyp.yaml
+                hyp['anchors'] = 3
+        if opt.noautoanchor:
+            del hyp['anchors'], meta['anchors']
+        opt.noval, opt.nosave, save_dir = True, True, Path(opt.save_dir)  # only val/save final epoch
+        # ei = [isinstance(x, (int, float)) for x in hyp.values()]  # evolvable indices
+        evolve_yaml, evolve_csv = save_dir / 'hyp_evolve.yaml', save_dir / 'evolve.csv'
+        if opt.bucket:
+            os.system(f'gsutil cp gs://{opt.bucket}/evolve.csv {evolve_csv}')  # download evolve.csv if exists
+
+        for _ in range(opt.evolve):  # generations to evolve
+            if evolve_csv.exists():  # if evolve.csv exists: select best hyps and mutate
+                # Select parent(s)
+                parent = 'single'  # parent selection method: 'single' or 'weighted'
+                x = np.loadtxt(evolve_csv, ndmin=2, delimiter=',', skiprows=1)
+                n = min(5, len(x))  # number of previous results to consider
+                x = x[np.argsort(-fitness(x))][:n]  # top n mutations
+                w = fitness(x) - fitness(x).min() + 1E-6  # weights (sum > 0)
+                if parent == 'single' or len(x) == 1:
+                    # x = x[random.randint(0, n - 1)]  # random selection
+                    x = x[random.choices(range(n), weights=w)[0]]  # weighted selection
+                elif parent == 'weighted':
+                    x = (x * w.reshape(n, 1)).sum(0) / w.sum()  # weighted combination
+
+                # Mutate
+                mp, s = 0.8, 0.2  # mutation probability, sigma
+                npr = np.random
+                npr.seed(int(time.time()))
+                g = np.array([meta[k][0] for k in hyp.keys()])  # gains 0-1
+                ng = len(meta)
+                v = np.ones(ng)
+                while all(v == 1):  # mutate until a change occurs (prevent duplicates)
+                    v = (g * (npr.random(ng) < mp) * npr.randn(ng) * npr.random() * s + 1).clip(0.3, 3.0)
+                for i, k in enumerate(hyp.keys()):  # plt.hist(v.ravel(), 300)
+                    hyp[k] = float(x[i + 7] * v[i])  # mutate
+
+            # Constrain to limits
+            for k, v in meta.items():
+                hyp[k] = max(hyp[k], v[1])  # lower limit
+                hyp[k] = min(hyp[k], v[2])  # upper limit
+                hyp[k] = round(hyp[k], 5)  # significant digits
+
+            # Train mutation
+            results = train(hyp.copy(), opt, device, callbacks)
+            callbacks = Callbacks()
+            # Write mutation results
+            print_mutation(KEYS, results, hyp.copy(), save_dir, opt.bucket)
+
+        # Plot results
+        plot_evolve(evolve_csv)
+        LOGGER.info(f'Hyperparameter evolution finished {opt.evolve} generations\n'
+                    f"Results saved to {colorstr('bold', save_dir)}\n"
+                    f'Usage example: $ python train.py --hyp {evolve_yaml}')
+
+
+def run(**kwargs):
+    # Usage: import train; train.run(data='coco128.yaml', imgsz=320, weights='yolo.pt')
+    opt = parse_opt(True)
+    for k, v in kwargs.items():
+        setattr(opt, k, v)
+    main(opt)
+    return opt
+
+
+if __name__ == "__main__":
+    opt = parse_opt()
+    main(opt)

panoptic/val.py

@@ -0,0 +1,597 @@
+import argparse
+import json
+import os
+import sys
+from multiprocessing.pool import ThreadPool
+from pathlib import Path
+
+import numpy as np
+import torch
+from tqdm import tqdm
+
+FILE = Path(__file__).resolve()
+ROOT = FILE.parents[1]  # YOLO root directory
+if str(ROOT) not in sys.path:
+    sys.path.append(str(ROOT))  # add ROOT to PATH
+ROOT = Path(os.path.relpath(ROOT, Path.cwd()))  # relative
+
+import torch.nn.functional as F
+import torchvision.transforms as transforms
+from pycocotools import mask as maskUtils
+from models.common import DetectMultiBackend
+from models.yolo import SegmentationModel
+from utils.callbacks import Callbacks
+from utils.coco_utils import getCocoIds, getMappingId, getMappingIndex
+from utils.general import (LOGGER, NUM_THREADS, TQDM_BAR_FORMAT, Profile, check_dataset, check_img_size,
+                           check_requirements, check_yaml, coco80_to_coco91_class, colorstr, increment_path,
+                           non_max_suppression, print_args, scale_boxes, xywh2xyxy, xyxy2xywh)
+from utils.metrics import ConfusionMatrix, box_iou
+from utils.plots import output_to_target, plot_val_study
+from utils.panoptic.dataloaders import create_dataloader
+from utils.panoptic.general import mask_iou, process_mask, process_mask_upsample, scale_image
+from utils.panoptic.metrics import Metrics, ap_per_class_box_and_mask, Semantic_Metrics
+from utils.panoptic.plots import plot_images_and_masks
+from utils.torch_utils import de_parallel, select_device, smart_inference_mode
+
+
+def save_one_txt(predn, save_conf, shape, file):
+    # Save one txt result
+    gn = torch.tensor(shape)[[1, 0, 1, 0]]  # normalization gain whwh
+    for *xyxy, conf, cls in predn.tolist():
+        xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist()  # normalized xywh
+        line = (cls, *xywh, conf) if save_conf else (cls, *xywh)  # label format
+        with open(file, 'a') as f:
+            f.write(('%g ' * len(line)).rstrip() % line + '\n')
+
+
+def save_one_json(predn, jdict, path, class_map, pred_masks):
+    # Save one JSON result {"image_id": 42, "category_id": 18, "bbox": [258.15, 41.29, 348.26, 243.78], "score": 0.236}
+    from pycocotools.mask import encode
+
+    def single_encode(x):
+        rle = encode(np.asarray(x[:, :, None], order="F", dtype="uint8"))[0]
+        rle["counts"] = rle["counts"].decode("utf-8")
+        return rle
+
+    image_id = int(path.stem) if path.stem.isnumeric() else path.stem
+    box = xyxy2xywh(predn[:, :4])  # xywh
+    box[:, :2] -= box[:, 2:] / 2  # xy center to top-left corner
+    pred_masks = np.transpose(pred_masks, (2, 0, 1))
+    with ThreadPool(NUM_THREADS) as pool:
+        rles = pool.map(single_encode, pred_masks)
+    for i, (p, b) in enumerate(zip(predn.tolist(), box.tolist())):
+        jdict.append({
+            'image_id': image_id,
+            'category_id': class_map[int(p[5])],
+            'bbox': [round(x, 3) for x in b],
+            'score': round(p[4], 5),
+            'segmentation': rles[i]})
+
+
+def process_batch(detections, labels, iouv, pred_masks=None, gt_masks=None, overlap=False, masks=False):
+    """
+    Return correct prediction matrix
+    Arguments:
+        detections (array[N, 6]), x1, y1, x2, y2, conf, class
+        labels (array[M, 5]), class, x1, y1, x2, y2
+    Returns:
+        correct (array[N, 10]), for 10 IoU levels
+    """
+    if masks:
+        if overlap:
+            nl = len(labels)
+            index = torch.arange(nl, device=gt_masks.device).view(nl, 1, 1) + 1
+            gt_masks = gt_masks.repeat(nl, 1, 1)  # shape(1,640,640) -> (n,640,640)
+            gt_masks = torch.where(gt_masks == index, 1.0, 0.0)
+        if gt_masks.shape[1:] != pred_masks.shape[1:]:
+            gt_masks = F.interpolate(gt_masks[None], pred_masks.shape[1:], mode="bilinear", align_corners=False)[0]
+            gt_masks = gt_masks.gt_(0.5)
+        iou = mask_iou(gt_masks.view(gt_masks.shape[0], -1), pred_masks.view(pred_masks.shape[0], -1))
+    else:  # boxes
+        iou = box_iou(labels[:, 1:], detections[:, :4])
+
+    correct = np.zeros((detections.shape[0], iouv.shape[0])).astype(bool)
+    correct_class = labels[:, 0:1] == detections[:, 5]
+    for i in range(len(iouv)):
+        x = torch.where((iou >= iouv[i]) & correct_class)  # IoU > threshold and classes match
+        if x[0].shape[0]:
+            matches = torch.cat((torch.stack(x, 1), iou[x[0], x[1]][:, None]), 1).cpu().numpy()  # [label, detect, iou]
+            if x[0].shape[0] > 1:
+                matches = matches[matches[:, 2].argsort()[::-1]]
+                matches = matches[np.unique(matches[:, 1], return_index=True)[1]]
+                # matches = matches[matches[:, 2].argsort()[::-1]]
+                matches = matches[np.unique(matches[:, 0], return_index=True)[1]]
+            correct[matches[:, 1].astype(int), i] = True
+    return torch.tensor(correct, dtype=torch.bool, device=iouv.device)
+
+
+@smart_inference_mode()
+def run(
+        data,
+        weights=None,  # model.pt path(s)
+        batch_size=32,  # batch size
+        imgsz=640,  # inference size (pixels)
+        conf_thres=0.001,  # confidence threshold
+        iou_thres=0.6,  # NMS IoU threshold
+        max_det=300,  # maximum detections per image
+        task='val',  # train, val, test, speed or study
+        device='',  # cuda device, i.e. 0 or 0,1,2,3 or cpu
+        workers=8,  # max dataloader workers (per RANK in DDP mode)
+        single_cls=False,  # treat as single-class dataset
+        augment=False,  # augmented inference
+        verbose=False,  # verbose output
+        save_txt=False,  # save results to *.txt
+        save_hybrid=False,  # save label+prediction hybrid results to *.txt
+        save_conf=False,  # save confidences in --save-txt labels
+        save_json=False,  # save a COCO-JSON results file
+        project=ROOT / 'runs/val-pan',  # save to project/name
+        name='exp',  # save to project/name
+        exist_ok=False,  # existing project/name ok, do not increment
+        half=True,  # use FP16 half-precision inference
+        dnn=False,  # use OpenCV DNN for ONNX inference
+        model=None,
+        dataloader=None,
+        save_dir=Path(''),
+        plots=True,
+        overlap=False,
+        mask_downsample_ratio=1,
+        compute_loss=None,
+        callbacks=Callbacks(),
+):
+    if save_json:
+        check_requirements(['pycocotools'])
+        process = process_mask_upsample  # more accurate
+    else:
+        process = process_mask  # faster
+
+    # Initialize/load model and set device
+    training = model is not None
+    if training:  # called by train.py
+        device, pt, jit, engine = next(model.parameters()).device, True, False, False  # get model device, PyTorch model
+        half &= device.type != 'cpu'  # half precision only supported on CUDA
+        model.half() if half else model.float()
+        nm = de_parallel(model).model[-1].nm  # number of masks
+    else:  # called directly
+        device = select_device(device, batch_size=batch_size)
+
+        # Directories
+        save_dir = increment_path(Path(project) / name, exist_ok=exist_ok)  # increment run
+        (save_dir / 'labels' if save_txt else save_dir).mkdir(parents=True, exist_ok=True)  # make dir
+
+        # Load model
+        model = DetectMultiBackend(weights, device=device, dnn=dnn, data=data, fp16=half)
+        stride, pt, jit, engine = model.stride, model.pt, model.jit, model.engine
+        imgsz = check_img_size(imgsz, s=stride)  # check image size
+        half = model.fp16  # FP16 supported on limited backends with CUDA
+        nm = de_parallel(model).model.model[-1].nm if isinstance(model, SegmentationModel) else 32  # number of masks
+        if engine:
+            batch_size = model.batch_size
+        else:
+            device = model.device
+            if not (pt or jit):
+                batch_size = 1  # export.py models default to batch-size 1
+                LOGGER.info(f'Forcing --batch-size 1 square inference (1,3,{imgsz},{imgsz}) for non-PyTorch models')
+
+        # Data
+        data = check_dataset(data)  # check
+
+    # Configure
+    model.eval()
+    cuda = device.type != 'cpu'
+    #is_coco = isinstance(data.get('val'), str) and data['val'].endswith(f'coco{os.sep}val2017.txt')  # COCO dataset
+    is_coco = isinstance(data.get('val'), str) and data['val'].endswith(f'val2017.txt')  # COCO dataset
+    nc = 1 if single_cls else int(data['nc'])  # number of classes
+    stuff_names = data.get('stuff_names', [])  # names of stuff classes
+    stuff_nc = len(stuff_names)  # number of stuff classes
+    iouv = torch.linspace(0.5, 0.95, 10, device=device)  # iou vector for [email protected]:0.95
+    niou = iouv.numel()
+
+    # Semantic Segmentation
+    img_id_list = []
+
+    # Dataloader
+    if not training:
+        if pt and not single_cls:  # check --weights are trained on --data
+            ncm = model.model.nc
+            assert ncm == nc, f'{weights} ({ncm} classes) trained on different --data than what you passed ({nc} ' \
+                              f'classes). Pass correct combination of --weights and --data that are trained together.'
+        model.warmup(imgsz=(1 if pt else batch_size, 3, imgsz, imgsz))  # warmup
+        pad, rect = (0.0, False) if task == 'speed' else (0.5, pt)  # square inference for benchmarks
+        task = task if task in ('train', 'val', 'test') else 'val'  # path to train/val/test images
+        dataloader = create_dataloader(data[task],
+                                       imgsz,
+                                       batch_size,
+                                       stride,
+                                       single_cls,
+                                       pad=pad,
+                                       rect=rect,
+                                       workers=workers,
+                                       prefix=colorstr(f'{task}: '),
+                                       overlap_mask=overlap,
+                                       mask_downsample_ratio=mask_downsample_ratio)[0]
+
+    seen = 0
+    confusion_matrix = ConfusionMatrix(nc=nc)
+    names = model.names if hasattr(model, 'names') else model.module.names  # get class names
+    if isinstance(names, (list, tuple)):  # old format
+        names = dict(enumerate(names))
+    class_map = coco80_to_coco91_class() if is_coco else list(range(1000))
+    s = ('%22s' + '%11s' * 12) % ('Class', 'Images', 'Instances', 'Box(P', "R", "mAP50", "mAP50-95)", "Mask(P", "R",
+                                  "mAP50", "mAP50-95)", 'S(MIoU', 'FWIoU)')
+    dt = Profile(), Profile(), Profile()
+    metrics = Metrics()
+    semantic_metrics = Semantic_Metrics(nc = (nc + stuff_nc), device = device)
+    loss = torch.zeros(6, device=device)
+    jdict, stats = [], []
+    semantic_jdict = []
+    # callbacks.run('on_val_start')
+    pbar = tqdm(dataloader, desc=s, bar_format=TQDM_BAR_FORMAT)  # progress bar
+    for batch_i, (im, targets, paths, shapes, masks, semasks) in enumerate(pbar):
+        # callbacks.run('on_val_batch_start')
+        with dt[0]:
+            if cuda:
+                im = im.to(device, non_blocking=True)
+                targets = targets.to(device)
+                masks = masks.to(device)
+                semasks = semasks.to(device)
+            masks = masks.float()
+            semasks = semasks.float()
+            im = im.half() if half else im.float()  # uint8 to fp16/32
+            im /= 255  # 0 - 255 to 0.0 - 1.0
+            nb, _, height, width = im.shape  # batch size, channels, height, width
+
+        # Inference
+        with dt[1]:
+            preds, train_out = model(im)# if compute_loss else (*model(im, augment=augment)[:2], None)
+            #train_out, preds, protos = p if len(p) == 3 else p[1]
+            #preds = p
+            #train_out = p[1][0] if len(p[1]) == 3 else p[0]
+            # protos = train_out[-1]
+            #print(preds.shape)
+            #print(train_out[0].shape)
+            #print(train_out[1].shape)
+            #print(train_out[2].shape)
+            _, pred_masks, protos, psemasks = train_out
+
+        # Loss
+        if compute_loss:
+            loss += compute_loss(train_out, targets, masks, semasks = semasks)[1]  # box, obj, cls
+
+        # NMS
+        targets[:, 2:] *= torch.tensor((width, height, width, height), device=device)  # to pixels
+        lb = [targets[targets[:, 0] == i, 1:] for i in range(nb)] if save_hybrid else []  # for autolabelling
+        with dt[2]:
+            preds = non_max_suppression(preds,
+                                        conf_thres,
+                                        iou_thres,
+                                        labels=lb,
+                                        multi_label=True,
+                                        agnostic=single_cls,
+                                        max_det=max_det,
+                                        nm=nm)
+
+        # Metrics
+        plot_masks = []  # masks for plotting
+        plot_semasks = []  # masks for plotting
+
+        if training:
+            semantic_metrics.update(psemasks, semasks)
+        else:
+            _, _, smh, smw = semasks.shape
+            semantic_metrics.update(torch.nn.functional.interpolate(psemasks, size = (smh, smw), mode = 'bilinear', align_corners = False), semasks)
+
+        if plots and batch_i < 3:
+            plot_semasks.append(psemasks.clone().detach().cpu())
+
+        for si, (pred, proto, psemask) in enumerate(zip(preds, protos, psemasks)):
+            labels = targets[targets[:, 0] == si, 1:]
+            nl, npr = labels.shape[0], pred.shape[0]  # number of labels, predictions
+            path, shape = Path(paths[si]), shapes[si][0]
+            image_id = path.stem
+            img_id_list.append(image_id)
+            correct_masks = torch.zeros(npr, niou, dtype=torch.bool, device=device)  # init
+            correct_bboxes = torch.zeros(npr, niou, dtype=torch.bool, device=device)  # init
+            seen += 1
+
+            if npr == 0:
+                if nl:
+                    stats.append((correct_masks, correct_bboxes, *torch.zeros((2, 0), device=device), labels[:, 0]))
+                    if plots:
+                        confusion_matrix.process_batch(detections=None, labels=labels[:, 0])
+            else:
+                # Masks
+                midx = [si] if overlap else targets[:, 0] == si
+                gt_masks = masks[midx]
+                pred_masks = process(proto, pred[:, 6:], pred[:, :4], shape=im[si].shape[1:])
+
+                # Predictions
+                if single_cls:
+                    pred[:, 5] = 0
+                predn = pred.clone()
+                scale_boxes(im[si].shape[1:], predn[:, :4], shape, shapes[si][1])  # native-space pred
+
+                # Evaluate
+                if nl:
+                    tbox = xywh2xyxy(labels[:, 1:5])  # target boxes
+                    scale_boxes(im[si].shape[1:], tbox, shape, shapes[si][1])  # native-space labels
+                    labelsn = torch.cat((labels[:, 0:1], tbox), 1)  # native-space labels
+                    correct_bboxes = process_batch(predn, labelsn, iouv)
+                    correct_masks = process_batch(predn, labelsn, iouv, pred_masks, gt_masks, overlap=overlap, masks=True)
+                    if plots:
+                        confusion_matrix.process_batch(predn, labelsn)
+                stats.append((correct_masks, correct_bboxes, pred[:, 4], pred[:, 5], labels[:, 0]))  # (conf, pcls, tcls)
+
+                pred_masks = torch.as_tensor(pred_masks, dtype=torch.uint8)
+                if plots and batch_i < 3:
+                    plot_masks.append(pred_masks[:15].cpu())  # filter top 15 to plot
+
+                # Save/log
+                if save_txt:
+                    save_one_txt(predn, save_conf, shape, file=save_dir / 'labels' / f'{path.stem}.txt')
+                if save_json:
+                    pred_masks = scale_image(im[si].shape[1:],
+                                            pred_masks.permute(1, 2, 0).contiguous().cpu().numpy(), shape, shapes[si][1])
+                    save_one_json(predn, jdict, path, class_map, pred_masks)  # append to COCO-JSON dictionary
+                # callbacks.run('on_val_image_end', pred, predn, path, names, im[si])
+
+            # Semantic Segmentation
+            h0, w0 = shape
+
+            # resize
+            _, mask_h, mask_w = psemask.shape
+            h_ratio = mask_h / h0
+            w_ratio = mask_w / w0
+
+            if h_ratio == w_ratio:
+                psemask = torch.nn.functional.interpolate(psemask[None, :], size = (h0, w0), mode = 'bilinear', align_corners = False)
+            else:
+                transform = transforms.CenterCrop((h0, w0))
+
+                if (1 != h_ratio) and (1 != w_ratio):
+                    h_new = h0 if (h_ratio < w_ratio) else int(mask_h / w_ratio)
+                    w_new = w0 if (h_ratio > w_ratio) else int(mask_w / h_ratio)
+                    psemask = torch.nn.functional.interpolate(psemask[None, :], size = (h_new, w_new), mode = 'bilinear', align_corners = False)
+
+                psemask = transform(psemask)
+
+            psemask = torch.squeeze(psemask)
+
+            nc, h, w = psemask.shape
+
+            semantic_mask = torch.flatten(psemask, start_dim = 1).permute(1, 0) # class x h x w -> (h x w) x class
+
+            max_idx = semantic_mask.argmax(1)
+            output_masks = torch.zeros(semantic_mask.shape).scatter(1, max_idx.cpu().unsqueeze(1), 1.0) # one hot: (h x w) x class
+            output_masks = torch.reshape(output_masks.permute(1, 0), (nc, h, w)) # (h x w) x class -> class x h x w
+            psemask = output_masks.to(device = device)
+
+            # TODO: check is_coco
+            instances_ids = getCocoIds(name = 'instances')
+            stuff_mask = torch.zeros((h, w), device = device)
+            check_semantic_mask = False
+            for idx, pred_semantic_mask in enumerate(psemask):
+                category_id = int(getMappingId(idx))
+                if 183 == category_id:
+                    # set all non-stuff pixels to other
+                    pred_semantic_mask = (torch.logical_xor(stuff_mask, torch.ones((h, w), device = device))).int()
+
+                # ignore the classes which all zeros / unlabeled class
+                if (0 >= torch.max(pred_semantic_mask)) or (0 >= category_id):
+                    continue
+
+                if category_id not in instances_ids:
+                    # record all stuff mask
+                    stuff_mask = torch.logical_or(stuff_mask, pred_semantic_mask)
+
+                if (category_id not in instances_ids):
+                    rle = maskUtils.encode(np.asfortranarray(pred_semantic_mask.cpu(), dtype = np.uint8))
+                    rle['counts'] = rle['counts'].decode('utf-8')
+
+                    temp_d = {
+                        'image_id': int(image_id) if image_id.isnumeric() else image_id,
+                        'category_id': category_id,
+                        'segmentation': rle,
+                        'score': 1
+                    }
+
+                    semantic_jdict.append(temp_d)
+                    check_semantic_mask = True
+
+            if not check_semantic_mask:
+                # append a other mask for evaluation if the image without any mask
+                other_mask = (torch.ones((h, w), device = device)).int()
+
+                rle = maskUtils.encode(np.asfortranarray(other_mask.cpu(), dtype = np.uint8))
+                rle['counts'] = rle['counts'].decode('utf-8')
+
+                temp_d = {
+                    'image_id': int(image_id) if image_id.isnumeric() else image_id,
+                    'category_id': 183,
+                    'segmentation': rle,
+                    'score': 1
+                }
+
+                semantic_jdict.append(temp_d)
+
+        # Plot images
+        if plots and batch_i < 3:
+            if len(plot_masks):
+                plot_masks = torch.cat(plot_masks, dim=0)
+            if len(plot_semasks):
+                plot_semasks = torch.cat(plot_semasks, dim = 0)
+            plot_images_and_masks(im, targets, masks, semasks, paths, save_dir / f'val_batch{batch_i}_labels.jpg', names)
+            plot_images_and_masks(im, output_to_target(preds, max_det=15), plot_masks, plot_semasks, paths,
+                                  save_dir / f'val_batch{batch_i}_pred.jpg', names)  # pred
+
+        # callbacks.run('on_val_batch_end')
+
+    # Compute metrics
+    stats = [torch.cat(x, 0).cpu().numpy() for x in zip(*stats)]  # to numpy
+    if len(stats) and stats[0].any():
+        results = ap_per_class_box_and_mask(*stats, plot=plots, save_dir=save_dir, names=names)
+        metrics.update(results)
+    nt = np.bincount(stats[4].astype(int), minlength=nc)  # number of targets per class
+
+    # Print results
+    pf = '%22s' + '%11i' * 2 + '%11.3g' * 10  # print format
+    LOGGER.info(pf % ("all", seen, nt.sum(), *metrics.mean_results(), *semantic_metrics.results()))
+    if nt.sum() == 0:
+        LOGGER.warning(f'WARNING ⚠️ no labels found in {task} set, can not compute metrics without labels')
+
+    # Print results per class
+    if (verbose or (nc < 50 and not training)) and nc > 1 and len(stats):
+        for i, c in enumerate(metrics.ap_class_index):
+            LOGGER.info(pf % (names[c], seen, nt[c], *metrics.class_result(i), *semantic_metrics.results()))
+
+    # Print speeds
+    t = tuple(x.t / seen * 1E3 for x in dt)  # speeds per image
+    if not training:
+        shape = (batch_size, 3, imgsz, imgsz)
+        LOGGER.info(f'Speed: %.1fms pre-process, %.1fms inference, %.1fms NMS per image at shape {shape}' % t)
+
+    # Plots
+    if plots:
+        confusion_matrix.plot(save_dir=save_dir, names=list(names.values()))
+    # callbacks.run('on_val_end')
+
+    mp_bbox, mr_bbox, map50_bbox, map_bbox, mp_mask, mr_mask, map50_mask, map_mask = metrics.mean_results()
+    miou_sem, fwiou_sem = semantic_metrics.results()
+    semantic_metrics.reset()
+
+    # Save JSON
+    if save_json and len(jdict):
+        w = Path(weights[0] if isinstance(weights, list) else weights).stem if weights is not None else ''  # weights
+        anno_path = Path(data.get('path', '../coco'))
+        anno_json = str(anno_path / 'annotations/instances_val2017.json')  # annotations json
+        pred_json = str(save_dir / f"{w}_predictions.json")  # predictions json
+        LOGGER.info(f'\nEvaluating pycocotools mAP... saving {pred_json}...')
+        with open(pred_json, 'w') as f:
+            json.dump(jdict, f)
+
+        semantic_anno_json = str(anno_path / 'annotations/stuff_val2017.json')  # annotations json
+        semantic_pred_json = str(save_dir / f"{w}_predictions_stuff.json")  # predictions json
+        LOGGER.info(f'\nsaving {semantic_pred_json}...')
+        with open(semantic_pred_json, 'w') as f:
+            json.dump(semantic_jdict, f)
+
+        try:  # https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocoEvalDemo.ipynb
+            from pycocotools.coco import COCO
+            from pycocotools.cocoeval import COCOeval
+
+            anno = COCO(anno_json)  # init annotations api
+            pred = anno.loadRes(pred_json)  # init predictions api
+            results = []
+            for eval in COCOeval(anno, pred, 'bbox'), COCOeval(anno, pred, 'segm'):
+                if is_coco:
+                    eval.params.imgIds = [int(Path(x).stem) for x in dataloader.dataset.im_files]  # img ID to evaluate
+                eval.evaluate()
+                eval.accumulate()
+                eval.summarize()
+                results.extend(eval.stats[:2])  # update results ([email protected]:0.95, [email protected])
+            map_bbox, map50_bbox, map_mask, map50_mask = results
+
+            # Semantic Segmentation
+            from utils.stuff_seg.cocostuffeval import COCOStuffeval
+
+            LOGGER.info(f'\nEvaluating pycocotools stuff... ')
+            imgIds = [int(x) for x in img_id_list]
+
+            stuffGt = COCO(semantic_anno_json)  # initialize COCO ground truth api
+            stuffDt = stuffGt.loadRes(semantic_pred_json)  # initialize COCO pred api
+
+            cocoStuffEval = COCOStuffeval(stuffGt, stuffDt)
+            cocoStuffEval.params.imgIds = imgIds  # image IDs to evaluate
+            cocoStuffEval.evaluate()
+            stats, statsClass = cocoStuffEval.summarize()
+            stuffIds = getCocoIds(name = 'stuff')
+            title = ' {:<5} | {:^6} | {:^6} '.format('class', 'iou', 'macc') if (0 >= len(stuff_names)) else \
+                    ' {:<5} | {:<20} | {:^6} | {:^6} '.format('class', 'class name', 'iou', 'macc')
+            print(title)
+            for idx, (iou, macc) in enumerate(zip(statsClass['ious'], statsClass['maccs'])):
+                id = (idx + 1)
+                if id not in stuffIds:
+                    continue
+                content = ' {:<5} | {:0.4f} | {:0.4f} '.format(str(id), iou, macc) if (0 >= len(stuff_names)) else \
+                            ' {:<5} | {:<20} | {:0.4f} | {:0.4f} '.format(str(id), str(stuff_names[getMappingIndex(id, name = 'stuff')]), iou, macc)
+                print(content)
+
+        except Exception as e:
+            LOGGER.info(f'pycocotools unable to run: {e}')
+
+    # Return results
+    model.float()  # for training
+    if not training:
+        s = f"\n{len(list(save_dir.glob('labels/*.txt')))} labels saved to {save_dir / 'labels'}" if save_txt else ''
+        LOGGER.info(f"Results saved to {colorstr('bold', save_dir)}{s}")
+    final_metric = mp_bbox, mr_bbox, map50_bbox, map_bbox, mp_mask, mr_mask, map50_mask, map_mask, miou_sem, fwiou_sem
+    return (*final_metric, *(loss.cpu() / len(dataloader)).tolist()), metrics.get_maps(nc), t
+
+
+def parse_opt():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--data', type=str, default=ROOT / 'data/coco128-pan.yaml', help='dataset.yaml path')
+    parser.add_argument('--weights', nargs='+', type=str, default=ROOT / 'yolo-pan.pt', help='model path(s)')
+    parser.add_argument('--batch-size', type=int, default=32, help='batch size')
+    parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=640, help='inference size (pixels)')
+    parser.add_argument('--conf-thres', type=float, default=0.001, help='confidence threshold')
+    parser.add_argument('--iou-thres', type=float, default=0.6, help='NMS IoU threshold')
+    parser.add_argument('--max-det', type=int, default=300, help='maximum detections per image')
+    parser.add_argument('--task', default='val', help='train, val, test, speed or study')
+    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
+    parser.add_argument('--workers', type=int, default=8, help='max dataloader workers (per RANK in DDP mode)')
+    parser.add_argument('--single-cls', action='store_true', help='treat as single-class dataset')
+    parser.add_argument('--augment', action='store_true', help='augmented inference')
+    parser.add_argument('--verbose', action='store_true', help='report mAP by class')
+    parser.add_argument('--save-txt', action='store_true', help='save results to *.txt')
+    parser.add_argument('--save-hybrid', action='store_true', help='save label+prediction hybrid results to *.txt')
+    parser.add_argument('--save-conf', action='store_true', help='save confidences in --save-txt labels')
+    parser.add_argument('--save-json', action='store_true', help='save a COCO-JSON results file')
+    parser.add_argument('--project', default=ROOT / 'runs/val-pan', help='save results to project/name')
+    parser.add_argument('--name', default='exp', help='save to project/name')
+    parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')
+    parser.add_argument('--half', action='store_true', help='use FP16 half-precision inference')
+    parser.add_argument('--dnn', action='store_true', help='use OpenCV DNN for ONNX inference')
+    opt = parser.parse_args()
+    opt.data = check_yaml(opt.data)  # check YAML
+    # opt.save_json |= opt.data.endswith('coco.yaml')
+    opt.save_txt |= opt.save_hybrid
+    print_args(vars(opt))
+    return opt
+
+
+def main(opt):
+    #check_requirements(requirements=ROOT / 'requirements.txt', exclude=('tensorboard', 'thop'))
+
+    if opt.task in ('train', 'val', 'test'):  # run normally
+        if opt.conf_thres > 0.001:  # https://github.com/
+            LOGGER.warning(f'WARNING ⚠️ confidence threshold {opt.conf_thres} > 0.001 produces invalid results')
+        if opt.save_hybrid:
+            LOGGER.warning('WARNING ⚠️ --save-hybrid returns high mAP from hybrid labels, not from predictions alone')
+        run(**vars(opt))
+
+    else:
+        weights = opt.weights if isinstance(opt.weights, list) else [opt.weights]
+        opt.half = torch.cuda.is_available() and opt.device != 'cpu'  # FP16 for fastest results
+        if opt.task == 'speed':  # speed benchmarks
+            # python val.py --task speed --data coco.yaml --batch 1 --weights yolo.pt...
+            opt.conf_thres, opt.iou_thres, opt.save_json = 0.25, 0.45, False
+            for opt.weights in weights:
+                run(**vars(opt), plots=False)
+
+        elif opt.task == 'study':  # speed vs mAP benchmarks
+            # python val.py --task study --data coco.yaml --iou 0.7 --weights yolo.pt...
+            for opt.weights in weights:
+                f = f'study_{Path(opt.data).stem}_{Path(opt.weights).stem}.txt'  # filename to save to
+                x, y = list(range(256, 1536 + 128, 128)), []  # x axis (image sizes), y axis
+                for opt.imgsz in x:  # img-size
+                    LOGGER.info(f'\nRunning {f} --imgsz {opt.imgsz}...')
+                    r, _, t = run(**vars(opt), plots=False)
+                    y.append(r + t)  # results and times
+                np.savetxt(f, y, fmt='%10.4g')  # save
+            os.system('zip -r study.zip study_*.txt')
+            plot_val_study(x=x)  # plot
+
+
+if __name__ == "__main__":
+    opt = parse_opt()
+    main(opt)

requirements.txt

@@ -0,0 +1,49 @@
+# requirements
+# Usage: pip install -r requirements.txt
+
+# Base ------------------------------------------------------------------------
+gitpython
+ipython
+matplotlib>=3.2.2
+numpy>=1.18.5
+opencv-python>=4.1.1
+Pillow>=7.1.2
+psutil
+PyYAML>=5.3.1
+requests>=2.23.0
+scipy>=1.4.1
+thop>=0.1.1
+opencv-python
+opencv-contrib-python
+torch
+torchvision
+tqdm>=4.64.0
+# protobuf<=3.20.1
+
+# Logging ---------------------------------------------------------------------
+tensorboard>=2.4.1
+# clearml>=1.2.0
+# comet
+
+# Plotting --------------------------------------------------------------------
+pandas>=1.1.4
+seaborn>=0.11.0
+
+# Export ----------------------------------------------------------------------
+# coremltools>=6.0
+# onnx>=1.9.0
+# onnx-simplifier>=0.4.1
+# nvidia-pyindex
+# nvidia-tensorrt
+# scikit-learn<=1.1.2
+# tensorflow>=2.4.1
+# tensorflowjs>=3.9.0
+# openvino-dev
+
+# Deploy ----------------------------------------------------------------------
+# tritonclient[all]~=2.24.0
+
+# Extras ----------------------------------------------------------------------
+# mss
+albumentations>=1.0.3
+pycocotools>=2.0

scripts/get_coco.sh

@@ -0,0 +1,22 @@
+#!/bin/bash
+# COCO 2017 dataset http://cocodataset.org
+# Download command: bash ./scripts/get_coco.sh
+
+# Download/unzip labels
+d='./' # unzip directory
+url=https://github.com/ultralytics/yolov5/releases/download/v1.0/
+f='coco2017labels-segments.zip' # or 'coco2017labels.zip', 68 MB
+echo 'Downloading' $url$f ' ...'
+curl -L $url$f -o $f && unzip -q $f -d $d && rm $f & # download, unzip, remove in background
+
+# Download/unzip images
+d='./coco/images' # unzip directory
+url=http://images.cocodataset.org/zips/
+f1='train2017.zip' # 19G, 118k images
+f2='val2017.zip'   # 1G, 5k images
+f3='test2017.zip'  # 7G, 41k images (optional)
+for f in $f1 $f2 $f3; do
+  echo 'Downloading' $url$f '...'
+  curl -L $url$f -o $f && unzip -q $f -d $d && rm $f & # download, unzip, remove in background
+done
+wait # finish background tasks

segment/predict.py

@@ -0,0 +1,246 @@
+import argparse
+import os
+import platform
+import sys
+from pathlib import Path
+
+import torch
+
+FILE = Path(__file__).resolve()
+ROOT = FILE.parents[1]  # YOLO root directory
+if str(ROOT) not in sys.path:
+    sys.path.append(str(ROOT))  # add ROOT to PATH
+ROOT = Path(os.path.relpath(ROOT, Path.cwd()))  # relative
+
+from models.common import DetectMultiBackend
+from utils.dataloaders import IMG_FORMATS, VID_FORMATS, LoadImages, LoadScreenshots, LoadStreams
+from utils.general import (LOGGER, Profile, check_file, check_img_size, check_imshow, check_requirements, colorstr, cv2,
+                           increment_path, non_max_suppression, print_args, scale_boxes, scale_segments,
+                           strip_optimizer, xyxy2xywh)
+from utils.plots import Annotator, colors, save_one_box
+from utils.segment.general import masks2segments, process_mask
+from utils.torch_utils import select_device, smart_inference_mode
+
+
+@smart_inference_mode()
+def run(
+    weights=ROOT / 'yolo-seg.pt',  # model.pt path(s)
+    source=ROOT / 'data/images',  # file/dir/URL/glob/screen/0(webcam)
+    data=ROOT / 'data/coco.yaml',  # dataset.yaml path
+    imgsz=(640, 640),  # inference size (height, width)
+    conf_thres=0.25,  # confidence threshold
+    iou_thres=0.45,  # NMS IOU threshold
+    max_det=1000,  # maximum detections per image
+    device='',  # cuda device, i.e. 0 or 0,1,2,3 or cpu
+    view_img=False,  # show results
+    save_txt=False,  # save results to *.txt
+    save_conf=False,  # save confidences in --save-txt labels
+    save_crop=False,  # save cropped prediction boxes
+    nosave=False,  # do not save images/videos
+    classes=None,  # filter by class: --class 0, or --class 0 2 3
+    agnostic_nms=False,  # class-agnostic NMS
+    augment=False,  # augmented inference
+    visualize=False,  # visualize features
+    update=False,  # update all models
+    project=ROOT / 'runs/predict-seg',  # save results to project/name
+    name='exp',  # save results to project/name
+    exist_ok=False,  # existing project/name ok, do not increment
+    line_thickness=3,  # bounding box thickness (pixels)
+    hide_labels=False,  # hide labels
+    hide_conf=False,  # hide confidences
+    half=False,  # use FP16 half-precision inference
+    dnn=False,  # use OpenCV DNN for ONNX inference
+    vid_stride=1,  # video frame-rate stride
+    retina_masks=False,
+):
+    source = str(source)
+    save_img = not nosave and not source.endswith('.txt')  # save inference images
+    is_file = Path(source).suffix[1:] in (IMG_FORMATS + VID_FORMATS)
+    is_url = source.lower().startswith(('rtsp://', 'rtmp://', 'http://', 'https://'))
+    webcam = source.isnumeric() or source.endswith('.txt') or (is_url and not is_file)
+    screenshot = source.lower().startswith('screen')
+    if is_url and is_file:
+        source = check_file(source)  # download
+
+    # Directories
+    save_dir = increment_path(Path(project) / name, exist_ok=exist_ok)  # increment run
+    (save_dir / 'labels' if save_txt else save_dir).mkdir(parents=True, exist_ok=True)  # make dir
+
+    # Load model
+    device = select_device(device)
+    model = DetectMultiBackend(weights, device=device, dnn=dnn, data=data, fp16=half)
+    stride, names, pt = model.stride, model.names, model.pt
+    imgsz = check_img_size(imgsz, s=stride)  # check image size
+
+    # Dataloader
+    bs = 1  # batch_size
+    if webcam:
+        view_img = check_imshow(warn=True)
+        dataset = LoadStreams(source, img_size=imgsz, stride=stride, auto=pt, vid_stride=vid_stride)
+        bs = len(dataset)
+    elif screenshot:
+        dataset = LoadScreenshots(source, img_size=imgsz, stride=stride, auto=pt)
+    else:
+        dataset = LoadImages(source, img_size=imgsz, stride=stride, auto=pt, vid_stride=vid_stride)
+    vid_path, vid_writer = [None] * bs, [None] * bs
+
+    # Run inference
+    model.warmup(imgsz=(1 if pt else bs, 3, *imgsz))  # warmup
+    seen, windows, dt = 0, [], (Profile(), Profile(), Profile())
+    for path, im, im0s, vid_cap, s in dataset:
+        with dt[0]:
+            im = torch.from_numpy(im).to(model.device)
+            im = im.half() if model.fp16 else im.float()  # uint8 to fp16/32
+            im /= 255  # 0 - 255 to 0.0 - 1.0
+            if len(im.shape) == 3:
+                im = im[None]  # expand for batch dim
+
+        # Inference
+        with dt[1]:
+            visualize = increment_path(save_dir / Path(path).stem, mkdir=True) if visualize else False
+            pred, proto = model(im, augment=augment, visualize=visualize)[:2]
+
+        # NMS
+        with dt[2]:
+            pred = non_max_suppression(pred, conf_thres, iou_thres, classes, agnostic_nms, max_det=max_det, nm=32)
+
+        # Second-stage classifier (optional)
+        # pred = utils.general.apply_classifier(pred, classifier_model, im, im0s)
+
+        # Process predictions
+        for i, det in enumerate(pred):  # per image
+            seen += 1
+            if webcam:  # batch_size >= 1
+                p, im0, frame = path[i], im0s[i].copy(), dataset.count
+                s += f'{i}: '
+            else:
+                p, im0, frame = path, im0s.copy(), getattr(dataset, 'frame', 0)
+
+            p = Path(p)  # to Path
+            save_path = str(save_dir / p.name)  # im.jpg
+            txt_path = str(save_dir / 'labels' / p.stem) + ('' if dataset.mode == 'image' else f'_{frame}')  # im.txt
+            s += '%gx%g ' % im.shape[2:]  # print string
+            imc = im0.copy() if save_crop else im0  # for save_crop
+            annotator = Annotator(im0, line_width=line_thickness, example=str(names))
+            if len(det):
+                masks = process_mask(proto[i], det[:, 6:], det[:, :4], im.shape[2:], upsample=True)  # HWC
+                det[:, :4] = scale_boxes(im.shape[2:], det[:, :4], im0.shape).round()  # rescale boxes to im0 size
+
+                # Segments
+                if save_txt:
+                    segments = reversed(masks2segments(masks))
+                    segments = [scale_segments(im.shape[2:], x, im0.shape, normalize=True) for x in segments]
+
+                # Print results
+                for c in det[:, 5].unique():
+                    n = (det[:, 5] == c).sum()  # detections per class
+                    s += f"{n} {names[int(c)]}{'s' * (n > 1)}, "  # add to string
+
+                # Mask plotting
+                annotator.masks(masks,
+                                colors=[colors(x, True) for x in det[:, 5]],
+                                im_gpu=None if retina_masks else im[i])
+
+                # Write results
+                for j, (*xyxy, conf, cls) in enumerate(reversed(det[:, :6])):
+                    if save_txt:  # Write to file
+                        segj = segments[j].reshape(-1)  # (n,2) to (n*2)
+                        line = (cls, *segj, conf) if save_conf else (cls, *segj)  # label format
+                        with open(f'{txt_path}.txt', 'a') as f:
+                            f.write(('%g ' * len(line)).rstrip() % line + '\n')
+
+                    if save_img or save_crop or view_img:  # Add bbox to image
+                        c = int(cls)  # integer class
+                        label = None if hide_labels else (names[c] if hide_conf else f'{names[c]} {conf:.2f}')
+                        annotator.box_label(xyxy, label, color=colors(c, True))
+                        # annotator.draw.polygon(segments[j], outline=colors(c, True), width=3)
+                    if save_crop:
+                        save_one_box(xyxy, imc, file=save_dir / 'crops' / names[c] / f'{p.stem}.jpg', BGR=True)
+
+            # Stream results
+            im0 = annotator.result()
+            if view_img:
+                if platform.system() == 'Linux' and p not in windows:
+                    windows.append(p)
+                    cv2.namedWindow(str(p), cv2.WINDOW_NORMAL | cv2.WINDOW_KEEPRATIO)  # allow window resize (Linux)
+                    cv2.resizeWindow(str(p), im0.shape[1], im0.shape[0])
+                cv2.imshow(str(p), im0)
+                if cv2.waitKey(1) == ord('q'):  # 1 millisecond
+                    exit()
+
+            # Save results (image with detections)
+            if save_img:
+                if dataset.mode == 'image':
+                    cv2.imwrite(save_path, im0)
+                else:  # 'video' or 'stream'
+                    if vid_path[i] != save_path:  # new video
+                        vid_path[i] = save_path
+                        if isinstance(vid_writer[i], cv2.VideoWriter):
+                            vid_writer[i].release()  # release previous video writer
+                        if vid_cap:  # video
+                            fps = vid_cap.get(cv2.CAP_PROP_FPS)
+                            w = int(vid_cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+                            h = int(vid_cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+                        else:  # stream
+                            fps, w, h = 30, im0.shape[1], im0.shape[0]
+                        save_path = str(Path(save_path).with_suffix('.mp4'))  # force *.mp4 suffix on results videos
+                        vid_writer[i] = cv2.VideoWriter(save_path, cv2.VideoWriter_fourcc(*'mp4v'), fps, (w, h))
+                    vid_writer[i].write(im0)
+
+        # Print time (inference-only)
+        LOGGER.info(f"{s}{'' if len(det) else '(no detections), '}{dt[1].dt * 1E3:.1f}ms")
+
+    # Print results
+    t = tuple(x.t / seen * 1E3 for x in dt)  # speeds per image
+    LOGGER.info(f'Speed: %.1fms pre-process, %.1fms inference, %.1fms NMS per image at shape {(1, 3, *imgsz)}' % t)
+    if save_txt or save_img:
+        s = f"\n{len(list(save_dir.glob('labels/*.txt')))} labels saved to {save_dir / 'labels'}" if save_txt else ''
+        LOGGER.info(f"Results saved to {colorstr('bold', save_dir)}{s}")
+    if update:
+        strip_optimizer(weights[0])  # update model (to fix SourceChangeWarning)
+
+
+def parse_opt():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--weights', nargs='+', type=str, default=ROOT / 'yolo-seg.pt', help='model path(s)')
+    parser.add_argument('--source', type=str, default=ROOT / 'data/images', help='file/dir/URL/glob/screen/0(webcam)')
+    parser.add_argument('--data', type=str, default=ROOT / 'data/coco128.yaml', help='(optional) dataset.yaml path')
+    parser.add_argument('--imgsz', '--img', '--img-size', nargs='+', type=int, default=[640], help='inference size h,w')
+    parser.add_argument('--conf-thres', type=float, default=0.25, help='confidence threshold')
+    parser.add_argument('--iou-thres', type=float, default=0.45, help='NMS IoU threshold')
+    parser.add_argument('--max-det', type=int, default=1000, help='maximum detections per image')
+    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
+    parser.add_argument('--view-img', action='store_true', help='show results')
+    parser.add_argument('--save-txt', action='store_true', help='save results to *.txt')
+    parser.add_argument('--save-conf', action='store_true', help='save confidences in --save-txt labels')
+    parser.add_argument('--save-crop', action='store_true', help='save cropped prediction boxes')
+    parser.add_argument('--nosave', action='store_true', help='do not save images/videos')
+    parser.add_argument('--classes', nargs='+', type=int, help='filter by class: --classes 0, or --classes 0 2 3')
+    parser.add_argument('--agnostic-nms', action='store_true', help='class-agnostic NMS')
+    parser.add_argument('--augment', action='store_true', help='augmented inference')
+    parser.add_argument('--visualize', action='store_true', help='visualize features')
+    parser.add_argument('--update', action='store_true', help='update all models')
+    parser.add_argument('--project', default=ROOT / 'runs/predict-seg', help='save results to project/name')
+    parser.add_argument('--name', default='exp', help='save results to project/name')
+    parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')
+    parser.add_argument('--line-thickness', default=3, type=int, help='bounding box thickness (pixels)')
+    parser.add_argument('--hide-labels', default=False, action='store_true', help='hide labels')
+    parser.add_argument('--hide-conf', default=False, action='store_true', help='hide confidences')
+    parser.add_argument('--half', action='store_true', help='use FP16 half-precision inference')
+    parser.add_argument('--dnn', action='store_true', help='use OpenCV DNN for ONNX inference')
+    parser.add_argument('--vid-stride', type=int, default=1, help='video frame-rate stride')
+    parser.add_argument('--retina-masks', action='store_true', help='whether to plot masks in native resolution')
+    opt = parser.parse_args()
+    opt.imgsz *= 2 if len(opt.imgsz) == 1 else 1  # expand
+    print_args(vars(opt))
+    return opt
+
+
+def main(opt):
+    check_requirements(exclude=('tensorboard', 'thop'))
+    run(**vars(opt))
+
+
+if __name__ == "__main__":
+    opt = parse_opt()
+    main(opt)

segment/train.py

@@ -0,0 +1,646 @@
+import argparse
+import math
+import os
+import random
+import sys
+import time
+from copy import deepcopy
+from datetime import datetime
+from pathlib import Path
+
+import numpy as np
+import torch
+import torch.distributed as dist
+import torch.nn as nn
+import yaml
+from torch.optim import lr_scheduler
+from tqdm import tqdm
+
+FILE = Path(__file__).resolve()
+ROOT = FILE.parents[1]  # YOLO root directory
+if str(ROOT) not in sys.path:
+    sys.path.append(str(ROOT))  # add ROOT to PATH
+ROOT = Path(os.path.relpath(ROOT, Path.cwd()))  # relative
+
+import segment.val as validate  # for end-of-epoch mAP
+from models.experimental import attempt_load
+from models.yolo import SegmentationModel
+from utils.autoanchor import check_anchors
+from utils.autobatch import check_train_batch_size
+from utils.callbacks import Callbacks
+from utils.downloads import attempt_download, is_url
+from utils.general import (LOGGER, TQDM_BAR_FORMAT, check_amp, check_dataset, check_file, check_git_info,
+                           check_git_status, check_img_size, check_requirements, check_suffix, check_yaml, colorstr,
+                           get_latest_run, increment_path, init_seeds, intersect_dicts, labels_to_class_weights,
+                           labels_to_image_weights, one_cycle, print_args, print_mutation, strip_optimizer, yaml_save)
+from utils.loggers import GenericLogger
+from utils.plots import plot_evolve, plot_labels
+from utils.segment.dataloaders import create_dataloader
+from utils.segment.loss_tal import ComputeLoss
+from utils.segment.metrics import KEYS, fitness
+from utils.segment.plots import plot_images_and_masks, plot_results_with_masks
+from utils.torch_utils import (EarlyStopping, ModelEMA, de_parallel, select_device, smart_DDP, smart_optimizer,
+                               smart_resume, torch_distributed_zero_first)
+
+LOCAL_RANK = int(os.getenv('LOCAL_RANK', -1))  # https://pytorch.org/docs/stable/elastic/run.html
+RANK = int(os.getenv('RANK', -1))
+WORLD_SIZE = int(os.getenv('WORLD_SIZE', 1))
+GIT_INFO = None#check_git_info()
+
+
+def train(hyp, opt, device, callbacks):  # hyp is path/to/hyp.yaml or hyp dictionary
+    save_dir, epochs, batch_size, weights, single_cls, evolve, data, cfg, resume, noval, nosave, workers, freeze, mask_ratio = \
+        Path(opt.save_dir), opt.epochs, opt.batch_size, opt.weights, opt.single_cls, opt.evolve, opt.data, opt.cfg, \
+        opt.resume, opt.noval, opt.nosave, opt.workers, opt.freeze, opt.mask_ratio
+    # callbacks.run('on_pretrain_routine_start')
+
+    # Directories
+    w = save_dir / 'weights'  # weights dir
+    (w.parent if evolve else w).mkdir(parents=True, exist_ok=True)  # make dir
+    last, best = w / 'last.pt', w / 'best.pt'
+
+    # Hyperparameters
+    if isinstance(hyp, str):
+        with open(hyp, errors='ignore') as f:
+            hyp = yaml.safe_load(f)  # load hyps dict
+    LOGGER.info(colorstr('hyperparameters: ') + ', '.join(f'{k}={v}' for k, v in hyp.items()))
+    opt.hyp = hyp.copy()  # for saving hyps to checkpoints
+
+    # Save run settings
+    if not evolve:
+        yaml_save(save_dir / 'hyp.yaml', hyp)
+        yaml_save(save_dir / 'opt.yaml', vars(opt))
+
+    # Loggers
+    data_dict = None
+    if RANK in {-1, 0}:
+        logger = GenericLogger(opt=opt, console_logger=LOGGER)
+
+    # Config
+    plots = not evolve and not opt.noplots  # create plots
+    overlap = not opt.no_overlap
+    cuda = device.type != 'cpu'
+    init_seeds(opt.seed + 1 + RANK, deterministic=True)
+    with torch_distributed_zero_first(LOCAL_RANK):
+        data_dict = data_dict or check_dataset(data)  # check if None
+    train_path, val_path = data_dict['train'], data_dict['val']
+    nc = 1 if single_cls else int(data_dict['nc'])  # number of classes
+    names = {0: 'item'} if single_cls and len(data_dict['names']) != 1 else data_dict['names']  # class names
+    #is_coco = isinstance(val_path, str) and val_path.endswith('coco/val2017.txt')  # COCO dataset
+    is_coco = isinstance(val_path, str) and val_path.endswith('val2017.txt')  # COCO dataset
+
+    # Model
+    check_suffix(weights, '.pt')  # check weights
+    pretrained = weights.endswith('.pt')
+    if pretrained:
+        with torch_distributed_zero_first(LOCAL_RANK):
+            weights = attempt_download(weights)  # download if not found locally
+        ckpt = torch.load(weights, map_location='cpu')  # load checkpoint to CPU to avoid CUDA memory leak
+        model = SegmentationModel(cfg or ckpt['model'].yaml, ch=3, nc=nc).to(device)
+        exclude = ['anchor'] if (cfg or hyp.get('anchors')) and not resume else []  # exclude keys
+        csd = ckpt['model'].float().state_dict()  # checkpoint state_dict as FP32
+        csd = intersect_dicts(csd, model.state_dict(), exclude=exclude)  # intersect
+        model.load_state_dict(csd, strict=False)  # load
+        LOGGER.info(f'Transferred {len(csd)}/{len(model.state_dict())} items from {weights}')  # report
+    else:
+        model = SegmentationModel(cfg, ch=3, nc=nc).to(device)  # create
+    amp = check_amp(model)  # check AMP
+
+    # Freeze
+    freeze = [f'model.{x}.' for x in (freeze if len(freeze) > 1 else range(freeze[0]))]  # layers to freeze
+    for k, v in model.named_parameters():
+        #v.requires_grad = True  # train all layers
+        # v.register_hook(lambda x: torch.nan_to_num(x))  # NaN to 0 (commented for erratic training results)
+        if any(x in k for x in freeze):
+            LOGGER.info(f'freezing {k}')
+            v.requires_grad = False
+
+    # Image size
+    gs = max(int(model.stride.max()), 32)  # grid size (max stride)
+    imgsz = check_img_size(opt.imgsz, gs, floor=gs * 2)  # verify imgsz is gs-multiple
+
+    # Batch size
+    if RANK == -1 and batch_size == -1:  # single-GPU only, estimate best batch size
+        batch_size = check_train_batch_size(model, imgsz, amp)
+        logger.update_params({"batch_size": batch_size})
+        # loggers.on_params_update({"batch_size": batch_size})
+
+    # Optimizer
+    nbs = 64  # nominal batch size
+    accumulate = max(round(nbs / batch_size), 1)  # accumulate loss before optimizing
+    hyp['weight_decay'] *= batch_size * accumulate / nbs  # scale weight_decay
+    optimizer = smart_optimizer(model, opt.optimizer, hyp['lr0'], hyp['momentum'], hyp['weight_decay'])
+
+    # Scheduler
+    if opt.cos_lr:
+        lf = one_cycle(1, hyp['lrf'], epochs)  # cosine 1->hyp['lrf']
+    else:
+        lf = lambda x: (1 - x / epochs) * (1.0 - hyp['lrf']) + hyp['lrf']  # linear
+    scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)  # plot_lr_scheduler(optimizer, scheduler, epochs)
+
+    # EMA
+    ema = ModelEMA(model) if RANK in {-1, 0} else None
+
+    # Resume
+    best_fitness, start_epoch = 0.0, 0
+    if pretrained:
+        if resume:
+            best_fitness, start_epoch, epochs = smart_resume(ckpt, optimizer, ema, weights, epochs, resume)
+        del ckpt, csd
+
+    # DP mode
+    if cuda and RANK == -1 and torch.cuda.device_count() > 1:
+        LOGGER.warning('WARNING ⚠️ DP not recommended, use torch.distributed.run for best DDP Multi-GPU results.')
+        model = torch.nn.DataParallel(model)
+
+    # SyncBatchNorm
+    if opt.sync_bn and cuda and RANK != -1:
+        model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device)
+        LOGGER.info('Using SyncBatchNorm()')
+
+    # Trainloader
+    train_loader, dataset = create_dataloader(
+        train_path,
+        imgsz,
+        batch_size // WORLD_SIZE,
+        gs,
+        single_cls,
+        hyp=hyp,
+        augment=True,
+        cache=None if opt.cache == 'val' else opt.cache,
+        rect=opt.rect,
+        rank=LOCAL_RANK,
+        workers=workers,
+        image_weights=opt.image_weights,
+        close_mosaic=opt.close_mosaic != 0,
+        quad=opt.quad,
+        prefix=colorstr('train: '),
+        shuffle=True,
+        mask_downsample_ratio=mask_ratio,
+        overlap_mask=overlap,
+    )
+    labels = np.concatenate(dataset.labels, 0)
+    mlc = int(labels[:, 0].max())  # max label class
+    assert mlc < nc, f'Label class {mlc} exceeds nc={nc} in {data}. Possible class labels are 0-{nc - 1}'
+
+    # Process 0
+    if RANK in {-1, 0}:
+        val_loader = create_dataloader(val_path,
+                                       imgsz,
+                                       batch_size // WORLD_SIZE * 2,
+                                       gs,
+                                       single_cls,
+                                       hyp=hyp,
+                                       cache=None if noval else opt.cache,
+                                       rect=True,
+                                       rank=-1,
+                                       workers=workers * 2,
+                                       pad=0.5,
+                                       mask_downsample_ratio=mask_ratio,
+                                       overlap_mask=overlap,
+                                       prefix=colorstr('val: '))[0]
+
+        if not resume:
+            #if not opt.noautoanchor:
+            #    check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz)  # run AutoAnchor
+            model.half().float()  # pre-reduce anchor precision
+
+            if plots:
+                plot_labels(labels, names, save_dir)
+        # callbacks.run('on_pretrain_routine_end', labels, names)
+
+    # DDP mode
+    if cuda and RANK != -1:
+        model = smart_DDP(model)
+
+    # Model attributes
+    nl = de_parallel(model).model[-1].nl  # number of detection layers (to scale hyps)
+    #hyp['box'] *= 3 / nl  # scale to layers
+    #hyp['cls'] *= nc / 80 * 3 / nl  # scale to classes and layers
+    #hyp['obj'] *= (imgsz / 640) ** 2 * 3 / nl  # scale to image size and layers
+    hyp['label_smoothing'] = opt.label_smoothing
+    model.nc = nc  # attach number of classes to model
+    model.hyp = hyp  # attach hyperparameters to model
+    model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) * nc  # attach class weights
+    model.names = names
+
+    # Start training
+    t0 = time.time()
+    nb = len(train_loader)  # number of batches
+    nw = max(round(hyp['warmup_epochs'] * nb), 100)  # number of warmup iterations, max(3 epochs, 100 iterations)
+    # nw = min(nw, (epochs - start_epoch) / 2 * nb)  # limit warmup to < 1/2 of training
+    last_opt_step = -1
+    maps = np.zeros(nc)  # mAP per class
+    results = (0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0)  # P, R, [email protected], [email protected], val_loss(box, obj, cls)
+    scheduler.last_epoch = start_epoch - 1  # do not move
+    scaler = torch.cuda.amp.GradScaler(enabled=amp)
+    stopper, stop = EarlyStopping(patience=opt.patience), False
+    compute_loss = ComputeLoss(model, overlap=overlap)  # init loss class
+    # callbacks.run('on_train_start')
+    LOGGER.info(f'Image sizes {imgsz} train, {imgsz} val\n'
+                f'Using {train_loader.num_workers * WORLD_SIZE} dataloader workers\n'
+                f"Logging results to {colorstr('bold', save_dir)}\n"
+                f'Starting training for {epochs} epochs...')
+    for epoch in range(start_epoch, epochs):  # epoch ------------------------------------------------------------------
+        # callbacks.run('on_train_epoch_start')
+        model.train()
+
+        # Update image weights (optional, single-GPU only)
+        if opt.image_weights:
+            cw = model.class_weights.cpu().numpy() * (1 - maps) ** 2 / nc  # class weights
+            iw = labels_to_image_weights(dataset.labels, nc=nc, class_weights=cw)  # image weights
+            dataset.indices = random.choices(range(dataset.n), weights=iw, k=dataset.n)  # rand weighted idx
+        if epoch == (epochs - opt.close_mosaic):
+            LOGGER.info("Closing dataloader mosaic")
+            dataset.mosaic = False
+
+        # Update mosaic border (optional)
+        # b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs)
+        # dataset.mosaic_border = [b - imgsz, -b]  # height, width borders
+
+        mloss = torch.zeros(4, device=device)  # mean losses
+        if RANK != -1:
+            train_loader.sampler.set_epoch(epoch)
+        pbar = enumerate(train_loader)
+        LOGGER.info(('\n' + '%11s' * 8) %
+                    ('Epoch', 'GPU_mem', 'box_loss', 'seg_loss', 'cls_loss', 'dfl_loss', 'Instances', 'Size'))
+        if RANK in {-1, 0}:
+            pbar = tqdm(pbar, total=nb, bar_format=TQDM_BAR_FORMAT)  # progress bar
+        optimizer.zero_grad()
+        for i, (imgs, targets, paths, _, masks) in pbar:  # batch ------------------------------------------------------
+            # callbacks.run('on_train_batch_start')
+            ni = i + nb * epoch  # number integrated batches (since train start)
+            imgs = imgs.to(device, non_blocking=True).float() / 255  # uint8 to float32, 0-255 to 0.0-1.0
+
+            # Warmup
+            if ni <= nw:
+                xi = [0, nw]  # x interp
+                # compute_loss.gr = np.interp(ni, xi, [0.0, 1.0])  # iou loss ratio (obj_loss = 1.0 or iou)
+                accumulate = max(1, np.interp(ni, xi, [1, nbs / batch_size]).round())
+                for j, x in enumerate(optimizer.param_groups):
+                    # bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
+                    x['lr'] = np.interp(ni, xi, [hyp['warmup_bias_lr'] if j == 0 else 0.0, x['initial_lr'] * lf(epoch)])
+                    if 'momentum' in x:
+                        x['momentum'] = np.interp(ni, xi, [hyp['warmup_momentum'], hyp['momentum']])
+
+            # Multi-scale
+            if opt.multi_scale:
+                sz = random.randrange(imgsz * 0.5, imgsz * 1.5 + gs) // gs * gs  # size
+                sf = sz / max(imgs.shape[2:])  # scale factor
+                if sf != 1:
+                    ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]]  # new shape (stretched to gs-multiple)
+                    imgs = nn.functional.interpolate(imgs, size=ns, mode='bilinear', align_corners=False)
+
+            # Forward
+            with torch.cuda.amp.autocast(amp):
+                pred = model(imgs)  # forward
+                loss, loss_items = compute_loss(pred, targets.to(device), masks=masks.to(device).float())
+                if RANK != -1:
+                    loss *= WORLD_SIZE  # gradient averaged between devices in DDP mode
+                if opt.quad:
+                    loss *= 4.
+
+            # Backward
+            scaler.scale(loss).backward()
+
+            # Optimize - https://pytorch.org/docs/master/notes/amp_examples.html
+            if ni - last_opt_step >= accumulate:
+                scaler.unscale_(optimizer)  # unscale gradients
+                torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=10.0)  # clip gradients
+                scaler.step(optimizer)  # optimizer.step
+                scaler.update()
+                optimizer.zero_grad()
+                if ema:
+                    ema.update(model)
+                last_opt_step = ni
+
+            # Log
+            if RANK in {-1, 0}:
+                mloss = (mloss * i + loss_items) / (i + 1)  # update mean losses
+                mem = f'{torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0:.3g}G'  # (GB)
+                pbar.set_description(('%11s' * 2 + '%11.4g' * 6) %
+                                     (f'{epoch}/{epochs - 1}', mem, *mloss, targets.shape[0], imgs.shape[-1]))
+                # callbacks.run('on_train_batch_end', model, ni, imgs, targets, paths)
+                # if callbacks.stop_training:
+                #    return
+
+                # Mosaic plots
+                if plots:
+                    if ni < 3:
+                        plot_images_and_masks(imgs, targets, masks, paths, save_dir / f"train_batch{ni}.jpg")
+                    if ni == 10:
+                        files = sorted(save_dir.glob('train*.jpg'))
+                        logger.log_images(files, "Mosaics", epoch)
+            # end batch ------------------------------------------------------------------------------------------------
+
+        # Scheduler
+        lr = [x['lr'] for x in optimizer.param_groups]  # for loggers
+        scheduler.step()
+
+        if RANK in {-1, 0}:
+            # mAP
+            # callbacks.run('on_train_epoch_end', epoch=epoch)
+            ema.update_attr(model, include=['yaml', 'nc', 'hyp', 'names', 'stride', 'class_weights'])
+            final_epoch = (epoch + 1 == epochs) or stopper.possible_stop
+            if not noval or final_epoch:  # Calculate mAP
+                results, maps, _ = validate.run(data_dict,
+                                                batch_size=batch_size // WORLD_SIZE * 2,
+                                                imgsz=imgsz,
+                                                half=amp,
+                                                model=ema.ema,
+                                                single_cls=single_cls,
+                                                dataloader=val_loader,
+                                                save_dir=save_dir,
+                                                plots=False,
+                                                callbacks=callbacks,
+                                                compute_loss=compute_loss,
+                                                mask_downsample_ratio=mask_ratio,
+                                                overlap=overlap)
+
+            # Update best mAP
+            fi = fitness(np.array(results).reshape(1, -1))  # weighted combination of [P, R, [email protected], [email protected]]
+            stop = stopper(epoch=epoch, fitness=fi)  # early stop check
+            if fi > best_fitness:
+                best_fitness = fi
+            log_vals = list(mloss) + list(results) + lr
+            # callbacks.run('on_fit_epoch_end', log_vals, epoch, best_fitness, fi)
+            # Log val metrics and media
+            metrics_dict = dict(zip(KEYS, log_vals))
+            logger.log_metrics(metrics_dict, epoch)
+
+            # Save model
+            if (not nosave) or (final_epoch and not evolve):  # if save
+                ckpt = {
+                    'epoch': epoch,
+                    'best_fitness': best_fitness,
+                    'model': deepcopy(de_parallel(model)).half(),
+                    'ema': deepcopy(ema.ema).half(),
+                    'updates': ema.updates,
+                    'optimizer': optimizer.state_dict(),
+                    'opt': vars(opt),
+                    'git': GIT_INFO,  # {remote, branch, commit} if a git repo
+                    'date': datetime.now().isoformat()}
+
+                # Save last, best and delete
+                torch.save(ckpt, last)
+                if best_fitness == fi:
+                    torch.save(ckpt, best)
+                if opt.save_period > 0 and epoch % opt.save_period == 0:
+                    torch.save(ckpt, w / f'epoch{epoch}.pt')
+                    logger.log_model(w / f'epoch{epoch}.pt')
+                del ckpt
+                # callbacks.run('on_model_save', last, epoch, final_epoch, best_fitness, fi)
+
+        # EarlyStopping
+        if RANK != -1:  # if DDP training
+            broadcast_list = [stop if RANK == 0 else None]
+            dist.broadcast_object_list(broadcast_list, 0)  # broadcast 'stop' to all ranks
+            if RANK != 0:
+                stop = broadcast_list[0]
+        if stop:
+            break  # must break all DDP ranks
+
+        # end epoch ----------------------------------------------------------------------------------------------------
+    # end training -----------------------------------------------------------------------------------------------------
+    if RANK in {-1, 0}:
+        LOGGER.info(f'\n{epoch - start_epoch + 1} epochs completed in {(time.time() - t0) / 3600:.3f} hours.')
+        for f in last, best:
+            if f.exists():
+                strip_optimizer(f)  # strip optimizers
+                if f is best:
+                    LOGGER.info(f'\nValidating {f}...')
+                    results, _, _ = validate.run(
+                        data_dict,
+                        batch_size=batch_size // WORLD_SIZE * 2,
+                        imgsz=imgsz,
+                        model=attempt_load(f, device).half(),
+                        iou_thres=0.65 if is_coco else 0.60,  # best pycocotools at iou 0.65
+                        single_cls=single_cls,
+                        dataloader=val_loader,
+                        save_dir=save_dir,
+                        save_json=is_coco,
+                        verbose=True,
+                        plots=plots,
+                        callbacks=callbacks,
+                        compute_loss=compute_loss,
+                        mask_downsample_ratio=mask_ratio,
+                        overlap=overlap)  # val best model with plots
+                    if is_coco:
+                        # callbacks.run('on_fit_epoch_end', list(mloss) + list(results) + lr, epoch, best_fitness, fi)
+                        metrics_dict = dict(zip(KEYS, list(mloss) + list(results) + lr))
+                        logger.log_metrics(metrics_dict, epoch)
+
+        # callbacks.run('on_train_end', last, best, epoch, results)
+        # on train end callback using genericLogger
+        logger.log_metrics(dict(zip(KEYS[4:16], results)), epochs)
+        if not opt.evolve:
+            logger.log_model(best, epoch)
+        if plots:
+            plot_results_with_masks(file=save_dir / 'results.csv')  # save results.png
+            files = ['results.png', 'confusion_matrix.png', *(f'{x}_curve.png' for x in ('F1', 'PR', 'P', 'R'))]
+            files = [(save_dir / f) for f in files if (save_dir / f).exists()]  # filter
+            LOGGER.info(f"Results saved to {colorstr('bold', save_dir)}")
+            logger.log_images(files, "Results", epoch + 1)
+            logger.log_images(sorted(save_dir.glob('val*.jpg')), "Validation", epoch + 1)
+    torch.cuda.empty_cache()
+    return results
+
+
+def parse_opt(known=False):
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--weights', type=str, default=ROOT / 'yolo-seg.pt', help='initial weights path')
+    parser.add_argument('--cfg', type=str, default='', help='model.yaml path')
+    parser.add_argument('--data', type=str, default=ROOT / 'data/coco128-seg.yaml', help='dataset.yaml path')
+    parser.add_argument('--hyp', type=str, default=ROOT / 'data/hyps/hyp.scratch-low.yaml', help='hyperparameters path')
+    parser.add_argument('--epochs', type=int, default=100, help='total training epochs')
+    parser.add_argument('--batch-size', type=int, default=16, help='total batch size for all GPUs, -1 for autobatch')
+    parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=640, help='train, val image size (pixels)')
+    parser.add_argument('--rect', action='store_true', help='rectangular training')
+    parser.add_argument('--resume', nargs='?', const=True, default=False, help='resume most recent training')
+    parser.add_argument('--nosave', action='store_true', help='only save final checkpoint')
+    parser.add_argument('--noval', action='store_true', help='only validate final epoch')
+    parser.add_argument('--noautoanchor', action='store_true', help='disable AutoAnchor')
+    parser.add_argument('--noplots', action='store_true', help='save no plot files')
+    parser.add_argument('--evolve', type=int, nargs='?', const=300, help='evolve hyperparameters for x generations')
+    parser.add_argument('--bucket', type=str, default='', help='gsutil bucket')
+    parser.add_argument('--cache', type=str, nargs='?', const='ram', help='image --cache ram/disk')
+    parser.add_argument('--image-weights', action='store_true', help='use weighted image selection for training')
+    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
+    parser.add_argument('--multi-scale', action='store_true', help='vary img-size +/- 50%%')
+    parser.add_argument('--single-cls', action='store_true', help='train multi-class data as single-class')
+    parser.add_argument('--optimizer', type=str, choices=['SGD', 'Adam', 'AdamW', 'LION'], default='SGD', help='optimizer')
+    parser.add_argument('--sync-bn', action='store_true', help='use SyncBatchNorm, only available in DDP mode')
+    parser.add_argument('--workers', type=int, default=8, help='max dataloader workers (per RANK in DDP mode)')
+    parser.add_argument('--project', default=ROOT / 'runs/train-seg', help='save to project/name')
+    parser.add_argument('--name', default='exp', help='save to project/name')
+    parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')
+    parser.add_argument('--quad', action='store_true', help='quad dataloader')
+    parser.add_argument('--cos-lr', action='store_true', help='cosine LR scheduler')
+    parser.add_argument('--label-smoothing', type=float, default=0.0, help='Label smoothing epsilon')
+    parser.add_argument('--patience', type=int, default=100, help='EarlyStopping patience (epochs without improvement)')
+    parser.add_argument('--freeze', nargs='+', type=int, default=[0], help='Freeze layers: backbone=10, first3=0 1 2')
+    parser.add_argument('--save-period', type=int, default=-1, help='Save checkpoint every x epochs (disabled if < 1)')
+    parser.add_argument('--seed', type=int, default=0, help='Global training seed')
+    parser.add_argument('--local_rank', type=int, default=-1, help='Automatic DDP Multi-GPU argument, do not modify')
+    parser.add_argument('--close-mosaic', type=int, default=0, help='Experimental')
+
+    # Instance Segmentation Args
+    parser.add_argument('--mask-ratio', type=int, default=4, help='Downsample the truth masks to saving memory')
+    parser.add_argument('--no-overlap', action='store_true', help='Overlap masks train faster at slightly less mAP')
+
+    return parser.parse_known_args()[0] if known else parser.parse_args()
+
+
+def main(opt, callbacks=Callbacks()):
+    # Checks
+    if RANK in {-1, 0}:
+        print_args(vars(opt))
+        #check_git_status()
+        #check_requirements()
+
+    # Resume
+    if opt.resume and not opt.evolve:  # resume from specified or most recent last.pt
+        last = Path(check_file(opt.resume) if isinstance(opt.resume, str) else get_latest_run())
+        opt_yaml = last.parent.parent / 'opt.yaml'  # train options yaml
+        opt_data = opt.data  # original dataset
+        if opt_yaml.is_file():
+            with open(opt_yaml, errors='ignore') as f:
+                d = yaml.safe_load(f)
+        else:
+            d = torch.load(last, map_location='cpu')['opt']
+        opt = argparse.Namespace(**d)  # replace
+        opt.cfg, opt.weights, opt.resume = '', str(last), True  # reinstate
+        if is_url(opt_data):
+            opt.data = check_file(opt_data)  # avoid HUB resume auth timeout
+    else:
+        opt.data, opt.cfg, opt.hyp, opt.weights, opt.project = \
+            check_file(opt.data), check_yaml(opt.cfg), check_yaml(opt.hyp), str(opt.weights), str(opt.project)  # checks
+        assert len(opt.cfg) or len(opt.weights), 'either --cfg or --weights must be specified'
+        if opt.evolve:
+            if opt.project == str(ROOT / 'runs/train'):  # if default project name, rename to runs/evolve
+                opt.project = str(ROOT / 'runs/evolve')
+            opt.exist_ok, opt.resume = opt.resume, False  # pass resume to exist_ok and disable resume
+        if opt.name == 'cfg':
+            opt.name = Path(opt.cfg).stem  # use model.yaml as name
+        opt.save_dir = str(increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok))
+
+    # DDP mode
+    device = select_device(opt.device, batch_size=opt.batch_size)
+    if LOCAL_RANK != -1:
+        msg = 'is not compatible with YOLO Multi-GPU DDP training'
+        assert not opt.image_weights, f'--image-weights {msg}'
+        assert not opt.evolve, f'--evolve {msg}'
+        assert opt.batch_size != -1, f'AutoBatch with --batch-size -1 {msg}, please pass a valid --batch-size'
+        assert opt.batch_size % WORLD_SIZE == 0, f'--batch-size {opt.batch_size} must be multiple of WORLD_SIZE'
+        assert torch.cuda.device_count() > LOCAL_RANK, 'insufficient CUDA devices for DDP command'
+        torch.cuda.set_device(LOCAL_RANK)
+        device = torch.device('cuda', LOCAL_RANK)
+        dist.init_process_group(backend="nccl" if dist.is_nccl_available() else "gloo")
+
+    # Train
+    if not opt.evolve:
+        train(opt.hyp, opt, device, callbacks)
+
+    # Evolve hyperparameters (optional)
+    else:
+        # Hyperparameter evolution metadata (mutation scale 0-1, lower_limit, upper_limit)
+        meta = {
+            'lr0': (1, 1e-5, 1e-1),  # initial learning rate (SGD=1E-2, Adam=1E-3)
+            'lrf': (1, 0.01, 1.0),  # final OneCycleLR learning rate (lr0 * lrf)
+            'momentum': (0.3, 0.6, 0.98),  # SGD momentum/Adam beta1
+            'weight_decay': (1, 0.0, 0.001),  # optimizer weight decay
+            'warmup_epochs': (1, 0.0, 5.0),  # warmup epochs (fractions ok)
+            'warmup_momentum': (1, 0.0, 0.95),  # warmup initial momentum
+            'warmup_bias_lr': (1, 0.0, 0.2),  # warmup initial bias lr
+            'box': (1, 0.02, 0.2),  # box loss gain
+            'cls': (1, 0.2, 4.0),  # cls loss gain
+            'cls_pw': (1, 0.5, 2.0),  # cls BCELoss positive_weight
+            'obj': (1, 0.2, 4.0),  # obj loss gain (scale with pixels)
+            'obj_pw': (1, 0.5, 2.0),  # obj BCELoss positive_weight
+            'iou_t': (0, 0.1, 0.7),  # IoU training threshold
+            'anchor_t': (1, 2.0, 8.0),  # anchor-multiple threshold
+            'anchors': (2, 2.0, 10.0),  # anchors per output grid (0 to ignore)
+            'fl_gamma': (0, 0.0, 2.0),  # focal loss gamma (efficientDet default gamma=1.5)
+            'hsv_h': (1, 0.0, 0.1),  # image HSV-Hue augmentation (fraction)
+            'hsv_s': (1, 0.0, 0.9),  # image HSV-Saturation augmentation (fraction)
+            'hsv_v': (1, 0.0, 0.9),  # image HSV-Value augmentation (fraction)
+            'degrees': (1, 0.0, 45.0),  # image rotation (+/- deg)
+            'translate': (1, 0.0, 0.9),  # image translation (+/- fraction)
+            'scale': (1, 0.0, 0.9),  # image scale (+/- gain)
+            'shear': (1, 0.0, 10.0),  # image shear (+/- deg)
+            'perspective': (0, 0.0, 0.001),  # image perspective (+/- fraction), range 0-0.001
+            'flipud': (1, 0.0, 1.0),  # image flip up-down (probability)
+            'fliplr': (0, 0.0, 1.0),  # image flip left-right (probability)
+            'mosaic': (1, 0.0, 1.0),  # image mixup (probability)
+            'mixup': (1, 0.0, 1.0),  # image mixup (probability)
+            'copy_paste': (1, 0.0, 1.0)}  # segment copy-paste (probability)
+
+        with open(opt.hyp, errors='ignore') as f:
+            hyp = yaml.safe_load(f)  # load hyps dict
+            if 'anchors' not in hyp:  # anchors commented in hyp.yaml
+                hyp['anchors'] = 3
+        if opt.noautoanchor:
+            del hyp['anchors'], meta['anchors']
+        opt.noval, opt.nosave, save_dir = True, True, Path(opt.save_dir)  # only val/save final epoch
+        # ei = [isinstance(x, (int, float)) for x in hyp.values()]  # evolvable indices
+        evolve_yaml, evolve_csv = save_dir / 'hyp_evolve.yaml', save_dir / 'evolve.csv'
+        if opt.bucket:
+            os.system(f'gsutil cp gs://{opt.bucket}/evolve.csv {evolve_csv}')  # download evolve.csv if exists
+
+        for _ in range(opt.evolve):  # generations to evolve
+            if evolve_csv.exists():  # if evolve.csv exists: select best hyps and mutate
+                # Select parent(s)
+                parent = 'single'  # parent selection method: 'single' or 'weighted'
+                x = np.loadtxt(evolve_csv, ndmin=2, delimiter=',', skiprows=1)
+                n = min(5, len(x))  # number of previous results to consider
+                x = x[np.argsort(-fitness(x))][:n]  # top n mutations
+                w = fitness(x) - fitness(x).min() + 1E-6  # weights (sum > 0)
+                if parent == 'single' or len(x) == 1:
+                    # x = x[random.randint(0, n - 1)]  # random selection
+                    x = x[random.choices(range(n), weights=w)[0]]  # weighted selection
+                elif parent == 'weighted':
+                    x = (x * w.reshape(n, 1)).sum(0) / w.sum()  # weighted combination
+
+                # Mutate
+                mp, s = 0.8, 0.2  # mutation probability, sigma
+                npr = np.random
+                npr.seed(int(time.time()))
+                g = np.array([meta[k][0] for k in hyp.keys()])  # gains 0-1
+                ng = len(meta)
+                v = np.ones(ng)
+                while all(v == 1):  # mutate until a change occurs (prevent duplicates)
+                    v = (g * (npr.random(ng) < mp) * npr.randn(ng) * npr.random() * s + 1).clip(0.3, 3.0)
+                for i, k in enumerate(hyp.keys()):  # plt.hist(v.ravel(), 300)
+                    hyp[k] = float(x[i + 7] * v[i])  # mutate
+
+            # Constrain to limits
+            for k, v in meta.items():
+                hyp[k] = max(hyp[k], v[1])  # lower limit
+                hyp[k] = min(hyp[k], v[2])  # upper limit
+                hyp[k] = round(hyp[k], 5)  # significant digits
+
+            # Train mutation
+            results = train(hyp.copy(), opt, device, callbacks)
+            callbacks = Callbacks()
+            # Write mutation results
+            print_mutation(KEYS, results, hyp.copy(), save_dir, opt.bucket)
+
+        # Plot results
+        plot_evolve(evolve_csv)
+        LOGGER.info(f'Hyperparameter evolution finished {opt.evolve} generations\n'
+                    f"Results saved to {colorstr('bold', save_dir)}\n"
+                    f'Usage example: $ python train.py --hyp {evolve_yaml}')
+
+
+def run(**kwargs):
+    # Usage: import train; train.run(data='coco128.yaml', imgsz=320, weights='yolo.pt')
+    opt = parse_opt(True)
+    for k, v in kwargs.items():
+        setattr(opt, k, v)
+    main(opt)
+    return opt
+
+
+if __name__ == "__main__":
+    opt = parse_opt()
+    main(opt)

segment/train_dual.py

@@ -0,0 +1,647 @@
+import argparse
+import math
+import os
+import random
+import sys
+import time
+from copy import deepcopy
+from datetime import datetime
+from pathlib import Path
+
+import numpy as np
+import torch
+import torch.distributed as dist
+import torch.nn as nn
+import yaml
+from torch.optim import lr_scheduler
+from tqdm import tqdm
+
+FILE = Path(__file__).resolve()
+ROOT = FILE.parents[1]  # YOLO root directory
+if str(ROOT) not in sys.path:
+    sys.path.append(str(ROOT))  # add ROOT to PATH
+ROOT = Path(os.path.relpath(ROOT, Path.cwd()))  # relative
+
+import segment.val_dual as validate  # for end-of-epoch mAP
+from models.experimental import attempt_load
+from models.yolo import SegmentationModel
+from utils.autoanchor import check_anchors
+from utils.autobatch import check_train_batch_size
+from utils.callbacks import Callbacks
+from utils.downloads import attempt_download, is_url
+from utils.general import (LOGGER, TQDM_BAR_FORMAT, check_amp, check_dataset, check_file, check_git_info,
+                           check_git_status, check_img_size, check_requirements, check_suffix, check_yaml, colorstr,
+                           get_latest_run, increment_path, init_seeds, intersect_dicts, labels_to_class_weights,
+                           labels_to_image_weights, one_cycle, print_args, print_mutation, strip_optimizer, yaml_save)
+from utils.loggers import GenericLogger
+from utils.plots import plot_evolve, plot_labels
+from utils.segment.dataloaders import create_dataloader
+from utils.segment.loss_tal_dual import ComputeLoss
+#from utils.segment.loss_tal_dual import ComputeLossLH as ComputeLoss
+from utils.segment.metrics import KEYS, fitness
+from utils.segment.plots import plot_images_and_masks, plot_results_with_masks
+from utils.torch_utils import (EarlyStopping, ModelEMA, de_parallel, select_device, smart_DDP, smart_optimizer,
+                               smart_resume, torch_distributed_zero_first)
+
+LOCAL_RANK = int(os.getenv('LOCAL_RANK', -1))  # https://pytorch.org/docs/stable/elastic/run.html
+RANK = int(os.getenv('RANK', -1))
+WORLD_SIZE = int(os.getenv('WORLD_SIZE', 1))
+GIT_INFO = None#check_git_info()
+
+
+def train(hyp, opt, device, callbacks):  # hyp is path/to/hyp.yaml or hyp dictionary
+    save_dir, epochs, batch_size, weights, single_cls, evolve, data, cfg, resume, noval, nosave, workers, freeze, mask_ratio = \
+        Path(opt.save_dir), opt.epochs, opt.batch_size, opt.weights, opt.single_cls, opt.evolve, opt.data, opt.cfg, \
+        opt.resume, opt.noval, opt.nosave, opt.workers, opt.freeze, opt.mask_ratio
+    # callbacks.run('on_pretrain_routine_start')
+
+    # Directories
+    w = save_dir / 'weights'  # weights dir
+    (w.parent if evolve else w).mkdir(parents=True, exist_ok=True)  # make dir
+    last, best = w / 'last.pt', w / 'best.pt'
+
+    # Hyperparameters
+    if isinstance(hyp, str):
+        with open(hyp, errors='ignore') as f:
+            hyp = yaml.safe_load(f)  # load hyps dict
+    LOGGER.info(colorstr('hyperparameters: ') + ', '.join(f'{k}={v}' for k, v in hyp.items()))
+    opt.hyp = hyp.copy()  # for saving hyps to checkpoints
+
+    # Save run settings
+    if not evolve:
+        yaml_save(save_dir / 'hyp.yaml', hyp)
+        yaml_save(save_dir / 'opt.yaml', vars(opt))
+
+    # Loggers
+    data_dict = None
+    if RANK in {-1, 0}:
+        logger = GenericLogger(opt=opt, console_logger=LOGGER)
+
+    # Config
+    plots = not evolve and not opt.noplots  # create plots
+    overlap = not opt.no_overlap
+    cuda = device.type != 'cpu'
+    init_seeds(opt.seed + 1 + RANK, deterministic=True)
+    with torch_distributed_zero_first(LOCAL_RANK):
+        data_dict = data_dict or check_dataset(data)  # check if None
+    train_path, val_path = data_dict['train'], data_dict['val']
+    nc = 1 if single_cls else int(data_dict['nc'])  # number of classes
+    names = {0: 'item'} if single_cls and len(data_dict['names']) != 1 else data_dict['names']  # class names
+    #is_coco = isinstance(val_path, str) and val_path.endswith('coco/val2017.txt')  # COCO dataset
+    is_coco = isinstance(val_path, str) and val_path.endswith('val2017.txt')  # COCO dataset
+
+    # Model
+    check_suffix(weights, '.pt')  # check weights
+    pretrained = weights.endswith('.pt')
+    if pretrained:
+        with torch_distributed_zero_first(LOCAL_RANK):
+            weights = attempt_download(weights)  # download if not found locally
+        ckpt = torch.load(weights, map_location='cpu')  # load checkpoint to CPU to avoid CUDA memory leak
+        model = SegmentationModel(cfg or ckpt['model'].yaml, ch=3, nc=nc).to(device)
+        exclude = ['anchor'] if (cfg or hyp.get('anchors')) and not resume else []  # exclude keys
+        csd = ckpt['model'].float().state_dict()  # checkpoint state_dict as FP32
+        csd = intersect_dicts(csd, model.state_dict(), exclude=exclude)  # intersect
+        model.load_state_dict(csd, strict=False)  # load
+        LOGGER.info(f'Transferred {len(csd)}/{len(model.state_dict())} items from {weights}')  # report
+    else:
+        model = SegmentationModel(cfg, ch=3, nc=nc).to(device)  # create
+    amp = check_amp(model)  # check AMP
+
+    # Freeze
+    freeze = [f'model.{x}.' for x in (freeze if len(freeze) > 1 else range(freeze[0]))]  # layers to freeze
+    for k, v in model.named_parameters():
+        #v.requires_grad = True  # train all layers
+        # v.register_hook(lambda x: torch.nan_to_num(x))  # NaN to 0 (commented for erratic training results)
+        if any(x in k for x in freeze):
+            LOGGER.info(f'freezing {k}')
+            v.requires_grad = False
+
+    # Image size
+    gs = max(int(model.stride.max()), 32)  # grid size (max stride)
+    imgsz = check_img_size(opt.imgsz, gs, floor=gs * 2)  # verify imgsz is gs-multiple
+
+    # Batch size
+    if RANK == -1 and batch_size == -1:  # single-GPU only, estimate best batch size
+        batch_size = check_train_batch_size(model, imgsz, amp)
+        logger.update_params({"batch_size": batch_size})
+        # loggers.on_params_update({"batch_size": batch_size})
+
+    # Optimizer
+    nbs = 64  # nominal batch size
+    accumulate = max(round(nbs / batch_size), 1)  # accumulate loss before optimizing
+    hyp['weight_decay'] *= batch_size * accumulate / nbs  # scale weight_decay
+    optimizer = smart_optimizer(model, opt.optimizer, hyp['lr0'], hyp['momentum'], hyp['weight_decay'])
+
+    # Scheduler
+    if opt.cos_lr:
+        lf = one_cycle(1, hyp['lrf'], epochs)  # cosine 1->hyp['lrf']
+    else:
+        lf = lambda x: (1 - x / epochs) * (1.0 - hyp['lrf']) + hyp['lrf']  # linear
+    scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)  # plot_lr_scheduler(optimizer, scheduler, epochs)
+
+    # EMA
+    ema = ModelEMA(model) if RANK in {-1, 0} else None
+
+    # Resume
+    best_fitness, start_epoch = 0.0, 0
+    if pretrained:
+        if resume:
+            best_fitness, start_epoch, epochs = smart_resume(ckpt, optimizer, ema, weights, epochs, resume)
+        del ckpt, csd
+
+    # DP mode
+    if cuda and RANK == -1 and torch.cuda.device_count() > 1:
+        LOGGER.warning('WARNING ⚠️ DP not recommended, use torch.distributed.run for best DDP Multi-GPU results.')
+        model = torch.nn.DataParallel(model)
+
+    # SyncBatchNorm
+    if opt.sync_bn and cuda and RANK != -1:
+        model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device)
+        LOGGER.info('Using SyncBatchNorm()')
+
+    # Trainloader
+    train_loader, dataset = create_dataloader(
+        train_path,
+        imgsz,
+        batch_size // WORLD_SIZE,
+        gs,
+        single_cls,
+        hyp=hyp,
+        augment=True,
+        cache=None if opt.cache == 'val' else opt.cache,
+        rect=opt.rect,
+        rank=LOCAL_RANK,
+        workers=workers,
+        image_weights=opt.image_weights,
+        close_mosaic=opt.close_mosaic != 0,
+        quad=opt.quad,
+        prefix=colorstr('train: '),
+        shuffle=True,
+        mask_downsample_ratio=mask_ratio,
+        overlap_mask=overlap,
+    )
+    labels = np.concatenate(dataset.labels, 0)
+    mlc = int(labels[:, 0].max())  # max label class
+    assert mlc < nc, f'Label class {mlc} exceeds nc={nc} in {data}. Possible class labels are 0-{nc - 1}'
+
+    # Process 0
+    if RANK in {-1, 0}:
+        val_loader = create_dataloader(val_path,
+                                       imgsz,
+                                       batch_size // WORLD_SIZE * 2,
+                                       gs,
+                                       single_cls,
+                                       hyp=hyp,
+                                       cache=None if noval else opt.cache,
+                                       rect=True,
+                                       rank=-1,
+                                       workers=workers * 2,
+                                       pad=0.5,
+                                       mask_downsample_ratio=mask_ratio,
+                                       overlap_mask=overlap,
+                                       prefix=colorstr('val: '))[0]
+
+        if not resume:
+            #if not opt.noautoanchor:
+            #    check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz)  # run AutoAnchor
+            model.half().float()  # pre-reduce anchor precision
+
+            if plots:
+                plot_labels(labels, names, save_dir)
+        # callbacks.run('on_pretrain_routine_end', labels, names)
+
+    # DDP mode
+    if cuda and RANK != -1:
+        model = smart_DDP(model)
+
+    # Model attributes
+    nl = de_parallel(model).model[-1].nl  # number of detection layers (to scale hyps)
+    #hyp['box'] *= 3 / nl  # scale to layers
+    #hyp['cls'] *= nc / 80 * 3 / nl  # scale to classes and layers
+    #hyp['obj'] *= (imgsz / 640) ** 2 * 3 / nl  # scale to image size and layers
+    hyp['label_smoothing'] = opt.label_smoothing
+    model.nc = nc  # attach number of classes to model
+    model.hyp = hyp  # attach hyperparameters to model
+    model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) * nc  # attach class weights
+    model.names = names
+
+    # Start training
+    t0 = time.time()
+    nb = len(train_loader)  # number of batches
+    nw = max(round(hyp['warmup_epochs'] * nb), 100)  # number of warmup iterations, max(3 epochs, 100 iterations)
+    # nw = min(nw, (epochs - start_epoch) / 2 * nb)  # limit warmup to < 1/2 of training
+    last_opt_step = -1
+    maps = np.zeros(nc)  # mAP per class
+    results = (0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0)  # P, R, [email protected], [email protected], val_loss(box, obj, cls)
+    scheduler.last_epoch = start_epoch - 1  # do not move
+    scaler = torch.cuda.amp.GradScaler(enabled=amp)
+    stopper, stop = EarlyStopping(patience=opt.patience), False
+    compute_loss = ComputeLoss(model, overlap=overlap)  # init loss class
+    # callbacks.run('on_train_start')
+    LOGGER.info(f'Image sizes {imgsz} train, {imgsz} val\n'
+                f'Using {train_loader.num_workers * WORLD_SIZE} dataloader workers\n'
+                f"Logging results to {colorstr('bold', save_dir)}\n"
+                f'Starting training for {epochs} epochs...')
+    for epoch in range(start_epoch, epochs):  # epoch ------------------------------------------------------------------
+        # callbacks.run('on_train_epoch_start')
+        model.train()
+
+        # Update image weights (optional, single-GPU only)
+        if opt.image_weights:
+            cw = model.class_weights.cpu().numpy() * (1 - maps) ** 2 / nc  # class weights
+            iw = labels_to_image_weights(dataset.labels, nc=nc, class_weights=cw)  # image weights
+            dataset.indices = random.choices(range(dataset.n), weights=iw, k=dataset.n)  # rand weighted idx
+        if epoch == (epochs - opt.close_mosaic):
+            LOGGER.info("Closing dataloader mosaic")
+            dataset.mosaic = False
+
+        # Update mosaic border (optional)
+        # b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs)
+        # dataset.mosaic_border = [b - imgsz, -b]  # height, width borders
+
+        mloss = torch.zeros(4, device=device)  # mean losses
+        if RANK != -1:
+            train_loader.sampler.set_epoch(epoch)
+        pbar = enumerate(train_loader)
+        LOGGER.info(('\n' + '%11s' * 8) %
+                    ('Epoch', 'GPU_mem', 'box_loss', 'seg_loss', 'cls_loss', 'dfl_loss', 'Instances', 'Size'))
+        if RANK in {-1, 0}:
+            pbar = tqdm(pbar, total=nb, bar_format=TQDM_BAR_FORMAT)  # progress bar
+        optimizer.zero_grad()
+        for i, (imgs, targets, paths, _, masks) in pbar:  # batch ------------------------------------------------------
+            # callbacks.run('on_train_batch_start')
+            ni = i + nb * epoch  # number integrated batches (since train start)
+            imgs = imgs.to(device, non_blocking=True).float() / 255  # uint8 to float32, 0-255 to 0.0-1.0
+
+            # Warmup
+            if ni <= nw:
+                xi = [0, nw]  # x interp
+                # compute_loss.gr = np.interp(ni, xi, [0.0, 1.0])  # iou loss ratio (obj_loss = 1.0 or iou)
+                accumulate = max(1, np.interp(ni, xi, [1, nbs / batch_size]).round())
+                for j, x in enumerate(optimizer.param_groups):
+                    # bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
+                    x['lr'] = np.interp(ni, xi, [hyp['warmup_bias_lr'] if j == 0 else 0.0, x['initial_lr'] * lf(epoch)])
+                    if 'momentum' in x:
+                        x['momentum'] = np.interp(ni, xi, [hyp['warmup_momentum'], hyp['momentum']])
+
+            # Multi-scale
+            if opt.multi_scale:
+                sz = random.randrange(imgsz * 0.5, imgsz * 1.5 + gs) // gs * gs  # size
+                sf = sz / max(imgs.shape[2:])  # scale factor
+                if sf != 1:
+                    ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]]  # new shape (stretched to gs-multiple)
+                    imgs = nn.functional.interpolate(imgs, size=ns, mode='bilinear', align_corners=False)
+
+            # Forward
+            with torch.cuda.amp.autocast(amp):
+                pred = model(imgs)  # forward
+                loss, loss_items = compute_loss(pred, targets.to(device), masks=masks.to(device).float())
+                if RANK != -1:
+                    loss *= WORLD_SIZE  # gradient averaged between devices in DDP mode
+                if opt.quad:
+                    loss *= 4.
+
+            # Backward
+            scaler.scale(loss).backward()
+
+            # Optimize - https://pytorch.org/docs/master/notes/amp_examples.html
+            if ni - last_opt_step >= accumulate:
+                scaler.unscale_(optimizer)  # unscale gradients
+                torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=10.0)  # clip gradients
+                scaler.step(optimizer)  # optimizer.step
+                scaler.update()
+                optimizer.zero_grad()
+                if ema:
+                    ema.update(model)
+                last_opt_step = ni
+
+            # Log
+            if RANK in {-1, 0}:
+                mloss = (mloss * i + loss_items) / (i + 1)  # update mean losses
+                mem = f'{torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0:.3g}G'  # (GB)
+                pbar.set_description(('%11s' * 2 + '%11.4g' * 6) %
+                                     (f'{epoch}/{epochs - 1}', mem, *mloss, targets.shape[0], imgs.shape[-1]))
+                # callbacks.run('on_train_batch_end', model, ni, imgs, targets, paths)
+                # if callbacks.stop_training:
+                #    return
+
+                # Mosaic plots
+                if plots:
+                    if ni < 3:
+                        plot_images_and_masks(imgs, targets, masks, paths, save_dir / f"train_batch{ni}.jpg")
+                    if ni == 10:
+                        files = sorted(save_dir.glob('train*.jpg'))
+                        logger.log_images(files, "Mosaics", epoch)
+            # end batch ------------------------------------------------------------------------------------------------
+
+        # Scheduler
+        lr = [x['lr'] for x in optimizer.param_groups]  # for loggers
+        scheduler.step()
+
+        if RANK in {-1, 0}:
+            # mAP
+            # callbacks.run('on_train_epoch_end', epoch=epoch)
+            ema.update_attr(model, include=['yaml', 'nc', 'hyp', 'names', 'stride', 'class_weights'])
+            final_epoch = (epoch + 1 == epochs) or stopper.possible_stop
+            if not noval or final_epoch:  # Calculate mAP
+                results, maps, _ = validate.run(data_dict,
+                                                batch_size=batch_size // WORLD_SIZE * 2,
+                                                imgsz=imgsz,
+                                                half=amp,
+                                                model=ema.ema,
+                                                single_cls=single_cls,
+                                                dataloader=val_loader,
+                                                save_dir=save_dir,
+                                                plots=False,
+                                                callbacks=callbacks,
+                                                compute_loss=compute_loss,
+                                                mask_downsample_ratio=mask_ratio,
+                                                overlap=overlap)
+
+            # Update best mAP
+            fi = fitness(np.array(results).reshape(1, -1))  # weighted combination of [P, R, [email protected], [email protected]]
+            stop = stopper(epoch=epoch, fitness=fi)  # early stop check
+            if fi > best_fitness:
+                best_fitness = fi
+            log_vals = list(mloss) + list(results) + lr
+            # callbacks.run('on_fit_epoch_end', log_vals, epoch, best_fitness, fi)
+            # Log val metrics and media
+            metrics_dict = dict(zip(KEYS, log_vals))
+            logger.log_metrics(metrics_dict, epoch)
+
+            # Save model
+            if (not nosave) or (final_epoch and not evolve):  # if save
+                ckpt = {
+                    'epoch': epoch,
+                    'best_fitness': best_fitness,
+                    'model': deepcopy(de_parallel(model)).half(),
+                    'ema': deepcopy(ema.ema).half(),
+                    'updates': ema.updates,
+                    'optimizer': optimizer.state_dict(),
+                    'opt': vars(opt),
+                    'git': GIT_INFO,  # {remote, branch, commit} if a git repo
+                    'date': datetime.now().isoformat()}
+
+                # Save last, best and delete
+                torch.save(ckpt, last)
+                if best_fitness == fi:
+                    torch.save(ckpt, best)
+                if opt.save_period > 0 and epoch % opt.save_period == 0:
+                    torch.save(ckpt, w / f'epoch{epoch}.pt')
+                    logger.log_model(w / f'epoch{epoch}.pt')
+                del ckpt
+                # callbacks.run('on_model_save', last, epoch, final_epoch, best_fitness, fi)
+
+        # EarlyStopping
+        if RANK != -1:  # if DDP training
+            broadcast_list = [stop if RANK == 0 else None]
+            dist.broadcast_object_list(broadcast_list, 0)  # broadcast 'stop' to all ranks
+            if RANK != 0:
+                stop = broadcast_list[0]
+        if stop:
+            break  # must break all DDP ranks
+
+        # end epoch ----------------------------------------------------------------------------------------------------
+    # end training -----------------------------------------------------------------------------------------------------
+    if RANK in {-1, 0}:
+        LOGGER.info(f'\n{epoch - start_epoch + 1} epochs completed in {(time.time() - t0) / 3600:.3f} hours.')
+        for f in last, best:
+            if f.exists():
+                strip_optimizer(f)  # strip optimizers
+                if f is best:
+                    LOGGER.info(f'\nValidating {f}...')
+                    results, _, _ = validate.run(
+                        data_dict,
+                        batch_size=batch_size // WORLD_SIZE * 2,
+                        imgsz=imgsz,
+                        model=attempt_load(f, device).half(),
+                        iou_thres=0.65 if is_coco else 0.60,  # best pycocotools at iou 0.65
+                        single_cls=single_cls,
+                        dataloader=val_loader,
+                        save_dir=save_dir,
+                        save_json=is_coco,
+                        verbose=True,
+                        plots=plots,
+                        callbacks=callbacks,
+                        compute_loss=compute_loss,
+                        mask_downsample_ratio=mask_ratio,
+                        overlap=overlap)  # val best model with plots
+                    if is_coco:
+                        # callbacks.run('on_fit_epoch_end', list(mloss) + list(results) + lr, epoch, best_fitness, fi)
+                        metrics_dict = dict(zip(KEYS, list(mloss) + list(results) + lr))
+                        logger.log_metrics(metrics_dict, epoch)
+
+        # callbacks.run('on_train_end', last, best, epoch, results)
+        # on train end callback using genericLogger
+        logger.log_metrics(dict(zip(KEYS[4:16], results)), epochs)
+        if not opt.evolve:
+            logger.log_model(best, epoch)
+        if plots:
+            plot_results_with_masks(file=save_dir / 'results.csv')  # save results.png
+            files = ['results.png', 'confusion_matrix.png', *(f'{x}_curve.png' for x in ('F1', 'PR', 'P', 'R'))]
+            files = [(save_dir / f) for f in files if (save_dir / f).exists()]  # filter
+            LOGGER.info(f"Results saved to {colorstr('bold', save_dir)}")
+            logger.log_images(files, "Results", epoch + 1)
+            logger.log_images(sorted(save_dir.glob('val*.jpg')), "Validation", epoch + 1)
+    torch.cuda.empty_cache()
+    return results
+
+
+def parse_opt(known=False):
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--weights', type=str, default=ROOT / 'yolo-seg.pt', help='initial weights path')
+    parser.add_argument('--cfg', type=str, default='', help='model.yaml path')
+    parser.add_argument('--data', type=str, default=ROOT / 'data/coco128-seg.yaml', help='dataset.yaml path')
+    parser.add_argument('--hyp', type=str, default=ROOT / 'data/hyps/hyp.scratch-low.yaml', help='hyperparameters path')
+    parser.add_argument('--epochs', type=int, default=100, help='total training epochs')
+    parser.add_argument('--batch-size', type=int, default=16, help='total batch size for all GPUs, -1 for autobatch')
+    parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=640, help='train, val image size (pixels)')
+    parser.add_argument('--rect', action='store_true', help='rectangular training')
+    parser.add_argument('--resume', nargs='?', const=True, default=False, help='resume most recent training')
+    parser.add_argument('--nosave', action='store_true', help='only save final checkpoint')
+    parser.add_argument('--noval', action='store_true', help='only validate final epoch')
+    parser.add_argument('--noautoanchor', action='store_true', help='disable AutoAnchor')
+    parser.add_argument('--noplots', action='store_true', help='save no plot files')
+    parser.add_argument('--evolve', type=int, nargs='?', const=300, help='evolve hyperparameters for x generations')
+    parser.add_argument('--bucket', type=str, default='', help='gsutil bucket')
+    parser.add_argument('--cache', type=str, nargs='?', const='ram', help='image --cache ram/disk')
+    parser.add_argument('--image-weights', action='store_true', help='use weighted image selection for training')
+    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
+    parser.add_argument('--multi-scale', action='store_true', help='vary img-size +/- 50%%')
+    parser.add_argument('--single-cls', action='store_true', help='train multi-class data as single-class')
+    parser.add_argument('--optimizer', type=str, choices=['SGD', 'Adam', 'AdamW', 'LION'], default='SGD', help='optimizer')
+    parser.add_argument('--sync-bn', action='store_true', help='use SyncBatchNorm, only available in DDP mode')
+    parser.add_argument('--workers', type=int, default=8, help='max dataloader workers (per RANK in DDP mode)')
+    parser.add_argument('--project', default=ROOT / 'runs/train-seg', help='save to project/name')
+    parser.add_argument('--name', default='exp', help='save to project/name')
+    parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')
+    parser.add_argument('--quad', action='store_true', help='quad dataloader')
+    parser.add_argument('--cos-lr', action='store_true', help='cosine LR scheduler')
+    parser.add_argument('--label-smoothing', type=float, default=0.0, help='Label smoothing epsilon')
+    parser.add_argument('--patience', type=int, default=100, help='EarlyStopping patience (epochs without improvement)')
+    parser.add_argument('--freeze', nargs='+', type=int, default=[0], help='Freeze layers: backbone=10, first3=0 1 2')
+    parser.add_argument('--save-period', type=int, default=-1, help='Save checkpoint every x epochs (disabled if < 1)')
+    parser.add_argument('--seed', type=int, default=0, help='Global training seed')
+    parser.add_argument('--local_rank', type=int, default=-1, help='Automatic DDP Multi-GPU argument, do not modify')
+    parser.add_argument('--close-mosaic', type=int, default=0, help='Experimental')
+
+    # Instance Segmentation Args
+    parser.add_argument('--mask-ratio', type=int, default=4, help='Downsample the truth masks to saving memory')
+    parser.add_argument('--no-overlap', action='store_true', help='Overlap masks train faster at slightly less mAP')
+
+    return parser.parse_known_args()[0] if known else parser.parse_args()
+
+
+def main(opt, callbacks=Callbacks()):
+    # Checks
+    if RANK in {-1, 0}:
+        print_args(vars(opt))
+        #check_git_status()
+        #check_requirements()
+
+    # Resume
+    if opt.resume and not opt.evolve:  # resume from specified or most recent last.pt
+        last = Path(check_file(opt.resume) if isinstance(opt.resume, str) else get_latest_run())
+        opt_yaml = last.parent.parent / 'opt.yaml'  # train options yaml
+        opt_data = opt.data  # original dataset
+        if opt_yaml.is_file():
+            with open(opt_yaml, errors='ignore') as f:
+                d = yaml.safe_load(f)
+        else:
+            d = torch.load(last, map_location='cpu')['opt']
+        opt = argparse.Namespace(**d)  # replace
+        opt.cfg, opt.weights, opt.resume = '', str(last), True  # reinstate
+        if is_url(opt_data):
+            opt.data = check_file(opt_data)  # avoid HUB resume auth timeout
+    else:
+        opt.data, opt.cfg, opt.hyp, opt.weights, opt.project = \
+            check_file(opt.data), check_yaml(opt.cfg), check_yaml(opt.hyp), str(opt.weights), str(opt.project)  # checks
+        assert len(opt.cfg) or len(opt.weights), 'either --cfg or --weights must be specified'
+        if opt.evolve:
+            if opt.project == str(ROOT / 'runs/train'):  # if default project name, rename to runs/evolve
+                opt.project = str(ROOT / 'runs/evolve')
+            opt.exist_ok, opt.resume = opt.resume, False  # pass resume to exist_ok and disable resume
+        if opt.name == 'cfg':
+            opt.name = Path(opt.cfg).stem  # use model.yaml as name
+        opt.save_dir = str(increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok))
+
+    # DDP mode
+    device = select_device(opt.device, batch_size=opt.batch_size)
+    if LOCAL_RANK != -1:
+        msg = 'is not compatible with YOLO Multi-GPU DDP training'
+        assert not opt.image_weights, f'--image-weights {msg}'
+        assert not opt.evolve, f'--evolve {msg}'
+        assert opt.batch_size != -1, f'AutoBatch with --batch-size -1 {msg}, please pass a valid --batch-size'
+        assert opt.batch_size % WORLD_SIZE == 0, f'--batch-size {opt.batch_size} must be multiple of WORLD_SIZE'
+        assert torch.cuda.device_count() > LOCAL_RANK, 'insufficient CUDA devices for DDP command'
+        torch.cuda.set_device(LOCAL_RANK)
+        device = torch.device('cuda', LOCAL_RANK)
+        dist.init_process_group(backend="nccl" if dist.is_nccl_available() else "gloo")
+
+    # Train
+    if not opt.evolve:
+        train(opt.hyp, opt, device, callbacks)
+
+    # Evolve hyperparameters (optional)
+    else:
+        # Hyperparameter evolution metadata (mutation scale 0-1, lower_limit, upper_limit)
+        meta = {
+            'lr0': (1, 1e-5, 1e-1),  # initial learning rate (SGD=1E-2, Adam=1E-3)
+            'lrf': (1, 0.01, 1.0),  # final OneCycleLR learning rate (lr0 * lrf)
+            'momentum': (0.3, 0.6, 0.98),  # SGD momentum/Adam beta1
+            'weight_decay': (1, 0.0, 0.001),  # optimizer weight decay
+            'warmup_epochs': (1, 0.0, 5.0),  # warmup epochs (fractions ok)
+            'warmup_momentum': (1, 0.0, 0.95),  # warmup initial momentum
+            'warmup_bias_lr': (1, 0.0, 0.2),  # warmup initial bias lr
+            'box': (1, 0.02, 0.2),  # box loss gain
+            'cls': (1, 0.2, 4.0),  # cls loss gain
+            'cls_pw': (1, 0.5, 2.0),  # cls BCELoss positive_weight
+            'obj': (1, 0.2, 4.0),  # obj loss gain (scale with pixels)
+            'obj_pw': (1, 0.5, 2.0),  # obj BCELoss positive_weight
+            'iou_t': (0, 0.1, 0.7),  # IoU training threshold
+            'anchor_t': (1, 2.0, 8.0),  # anchor-multiple threshold
+            'anchors': (2, 2.0, 10.0),  # anchors per output grid (0 to ignore)
+            'fl_gamma': (0, 0.0, 2.0),  # focal loss gamma (efficientDet default gamma=1.5)
+            'hsv_h': (1, 0.0, 0.1),  # image HSV-Hue augmentation (fraction)
+            'hsv_s': (1, 0.0, 0.9),  # image HSV-Saturation augmentation (fraction)
+            'hsv_v': (1, 0.0, 0.9),  # image HSV-Value augmentation (fraction)
+            'degrees': (1, 0.0, 45.0),  # image rotation (+/- deg)
+            'translate': (1, 0.0, 0.9),  # image translation (+/- fraction)
+            'scale': (1, 0.0, 0.9),  # image scale (+/- gain)
+            'shear': (1, 0.0, 10.0),  # image shear (+/- deg)
+            'perspective': (0, 0.0, 0.001),  # image perspective (+/- fraction), range 0-0.001
+            'flipud': (1, 0.0, 1.0),  # image flip up-down (probability)
+            'fliplr': (0, 0.0, 1.0),  # image flip left-right (probability)
+            'mosaic': (1, 0.0, 1.0),  # image mixup (probability)
+            'mixup': (1, 0.0, 1.0),  # image mixup (probability)
+            'copy_paste': (1, 0.0, 1.0)}  # segment copy-paste (probability)
+
+        with open(opt.hyp, errors='ignore') as f:
+            hyp = yaml.safe_load(f)  # load hyps dict
+            if 'anchors' not in hyp:  # anchors commented in hyp.yaml
+                hyp['anchors'] = 3
+        if opt.noautoanchor:
+            del hyp['anchors'], meta['anchors']
+        opt.noval, opt.nosave, save_dir = True, True, Path(opt.save_dir)  # only val/save final epoch
+        # ei = [isinstance(x, (int, float)) for x in hyp.values()]  # evolvable indices
+        evolve_yaml, evolve_csv = save_dir / 'hyp_evolve.yaml', save_dir / 'evolve.csv'
+        if opt.bucket:
+            os.system(f'gsutil cp gs://{opt.bucket}/evolve.csv {evolve_csv}')  # download evolve.csv if exists
+
+        for _ in range(opt.evolve):  # generations to evolve
+            if evolve_csv.exists():  # if evolve.csv exists: select best hyps and mutate
+                # Select parent(s)
+                parent = 'single'  # parent selection method: 'single' or 'weighted'
+                x = np.loadtxt(evolve_csv, ndmin=2, delimiter=',', skiprows=1)
+                n = min(5, len(x))  # number of previous results to consider
+                x = x[np.argsort(-fitness(x))][:n]  # top n mutations
+                w = fitness(x) - fitness(x).min() + 1E-6  # weights (sum > 0)
+                if parent == 'single' or len(x) == 1:
+                    # x = x[random.randint(0, n - 1)]  # random selection
+                    x = x[random.choices(range(n), weights=w)[0]]  # weighted selection
+                elif parent == 'weighted':
+                    x = (x * w.reshape(n, 1)).sum(0) / w.sum()  # weighted combination
+
+                # Mutate
+                mp, s = 0.8, 0.2  # mutation probability, sigma
+                npr = np.random
+                npr.seed(int(time.time()))
+                g = np.array([meta[k][0] for k in hyp.keys()])  # gains 0-1
+                ng = len(meta)
+                v = np.ones(ng)
+                while all(v == 1):  # mutate until a change occurs (prevent duplicates)
+                    v = (g * (npr.random(ng) < mp) * npr.randn(ng) * npr.random() * s + 1).clip(0.3, 3.0)
+                for i, k in enumerate(hyp.keys()):  # plt.hist(v.ravel(), 300)
+                    hyp[k] = float(x[i + 7] * v[i])  # mutate
+
+            # Constrain to limits
+            for k, v in meta.items():
+                hyp[k] = max(hyp[k], v[1])  # lower limit
+                hyp[k] = min(hyp[k], v[2])  # upper limit
+                hyp[k] = round(hyp[k], 5)  # significant digits
+
+            # Train mutation
+            results = train(hyp.copy(), opt, device, callbacks)
+            callbacks = Callbacks()
+            # Write mutation results
+            print_mutation(KEYS, results, hyp.copy(), save_dir, opt.bucket)
+
+        # Plot results
+        plot_evolve(evolve_csv)
+        LOGGER.info(f'Hyperparameter evolution finished {opt.evolve} generations\n'
+                    f"Results saved to {colorstr('bold', save_dir)}\n"
+                    f'Usage example: $ python train.py --hyp {evolve_yaml}')
+
+
+def run(**kwargs):
+    # Usage: import train; train.run(data='coco128.yaml', imgsz=320, weights='yolo.pt')
+    opt = parse_opt(True)
+    for k, v in kwargs.items():
+        setattr(opt, k, v)
+    main(opt)
+    return opt
+
+
+if __name__ == "__main__":
+    opt = parse_opt()
+    main(opt)

segment/val.py

@@ -0,0 +1,457 @@
+import argparse
+import json
+import os
+import sys
+from multiprocessing.pool import ThreadPool
+from pathlib import Path
+
+import numpy as np
+import torch
+from tqdm import tqdm
+
+FILE = Path(__file__).resolve()
+ROOT = FILE.parents[1]  # YOLO root directory
+if str(ROOT) not in sys.path:
+    sys.path.append(str(ROOT))  # add ROOT to PATH
+ROOT = Path(os.path.relpath(ROOT, Path.cwd()))  # relative
+
+import torch.nn.functional as F
+
+from models.common import DetectMultiBackend
+from models.yolo import SegmentationModel
+from utils.callbacks import Callbacks
+from utils.general import (LOGGER, NUM_THREADS, TQDM_BAR_FORMAT, Profile, check_dataset, check_img_size,
+                           check_requirements, check_yaml, coco80_to_coco91_class, colorstr, increment_path,
+                           non_max_suppression, print_args, scale_boxes, xywh2xyxy, xyxy2xywh)
+from utils.metrics import ConfusionMatrix, box_iou
+from utils.plots import output_to_target, plot_val_study
+from utils.segment.dataloaders import create_dataloader
+from utils.segment.general import mask_iou, process_mask, process_mask_upsample, scale_image
+from utils.segment.metrics import Metrics, ap_per_class_box_and_mask
+from utils.segment.plots import plot_images_and_masks
+from utils.torch_utils import de_parallel, select_device, smart_inference_mode
+
+
+def save_one_txt(predn, save_conf, shape, file):
+    # Save one txt result
+    gn = torch.tensor(shape)[[1, 0, 1, 0]]  # normalization gain whwh
+    for *xyxy, conf, cls in predn.tolist():
+        xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist()  # normalized xywh
+        line = (cls, *xywh, conf) if save_conf else (cls, *xywh)  # label format
+        with open(file, 'a') as f:
+            f.write(('%g ' * len(line)).rstrip() % line + '\n')
+
+
+def save_one_json(predn, jdict, path, class_map, pred_masks):
+    # Save one JSON result {"image_id": 42, "category_id": 18, "bbox": [258.15, 41.29, 348.26, 243.78], "score": 0.236}
+    from pycocotools.mask import encode
+
+    def single_encode(x):
+        rle = encode(np.asarray(x[:, :, None], order="F", dtype="uint8"))[0]
+        rle["counts"] = rle["counts"].decode("utf-8")
+        return rle
+
+    image_id = int(path.stem) if path.stem.isnumeric() else path.stem
+    box = xyxy2xywh(predn[:, :4])  # xywh
+    box[:, :2] -= box[:, 2:] / 2  # xy center to top-left corner
+    pred_masks = np.transpose(pred_masks, (2, 0, 1))
+    with ThreadPool(NUM_THREADS) as pool:
+        rles = pool.map(single_encode, pred_masks)
+    for i, (p, b) in enumerate(zip(predn.tolist(), box.tolist())):
+        jdict.append({
+            'image_id': image_id,
+            'category_id': class_map[int(p[5])],
+            'bbox': [round(x, 3) for x in b],
+            'score': round(p[4], 5),
+            'segmentation': rles[i]})
+
+
+def process_batch(detections, labels, iouv, pred_masks=None, gt_masks=None, overlap=False, masks=False):
+    """
+    Return correct prediction matrix
+    Arguments:
+        detections (array[N, 6]), x1, y1, x2, y2, conf, class
+        labels (array[M, 5]), class, x1, y1, x2, y2
+    Returns:
+        correct (array[N, 10]), for 10 IoU levels
+    """
+    if masks:
+        if overlap:
+            nl = len(labels)
+            index = torch.arange(nl, device=gt_masks.device).view(nl, 1, 1) + 1
+            gt_masks = gt_masks.repeat(nl, 1, 1)  # shape(1,640,640) -> (n,640,640)
+            gt_masks = torch.where(gt_masks == index, 1.0, 0.0)
+        if gt_masks.shape[1:] != pred_masks.shape[1:]:
+            gt_masks = F.interpolate(gt_masks[None], pred_masks.shape[1:], mode="bilinear", align_corners=False)[0]
+            gt_masks = gt_masks.gt_(0.5)
+        iou = mask_iou(gt_masks.view(gt_masks.shape[0], -1), pred_masks.view(pred_masks.shape[0], -1))
+    else:  # boxes
+        iou = box_iou(labels[:, 1:], detections[:, :4])
+
+    correct = np.zeros((detections.shape[0], iouv.shape[0])).astype(bool)
+    correct_class = labels[:, 0:1] == detections[:, 5]
+    for i in range(len(iouv)):
+        x = torch.where((iou >= iouv[i]) & correct_class)  # IoU > threshold and classes match
+        if x[0].shape[0]:
+            matches = torch.cat((torch.stack(x, 1), iou[x[0], x[1]][:, None]), 1).cpu().numpy()  # [label, detect, iou]
+            if x[0].shape[0] > 1:
+                matches = matches[matches[:, 2].argsort()[::-1]]
+                matches = matches[np.unique(matches[:, 1], return_index=True)[1]]
+                # matches = matches[matches[:, 2].argsort()[::-1]]
+                matches = matches[np.unique(matches[:, 0], return_index=True)[1]]
+            correct[matches[:, 1].astype(int), i] = True
+    return torch.tensor(correct, dtype=torch.bool, device=iouv.device)
+
+
+@smart_inference_mode()
+def run(
+        data,
+        weights=None,  # model.pt path(s)
+        batch_size=32,  # batch size
+        imgsz=640,  # inference size (pixels)
+        conf_thres=0.001,  # confidence threshold
+        iou_thres=0.6,  # NMS IoU threshold
+        max_det=300,  # maximum detections per image
+        task='val',  # train, val, test, speed or study
+        device='',  # cuda device, i.e. 0 or 0,1,2,3 or cpu
+        workers=8,  # max dataloader workers (per RANK in DDP mode)
+        single_cls=False,  # treat as single-class dataset
+        augment=False,  # augmented inference
+        verbose=False,  # verbose output
+        save_txt=False,  # save results to *.txt
+        save_hybrid=False,  # save label+prediction hybrid results to *.txt
+        save_conf=False,  # save confidences in --save-txt labels
+        save_json=False,  # save a COCO-JSON results file
+        project=ROOT / 'runs/val-seg',  # save to project/name
+        name='exp',  # save to project/name
+        exist_ok=False,  # existing project/name ok, do not increment
+        half=True,  # use FP16 half-precision inference
+        dnn=False,  # use OpenCV DNN for ONNX inference
+        model=None,
+        dataloader=None,
+        save_dir=Path(''),
+        plots=True,
+        overlap=False,
+        mask_downsample_ratio=1,
+        compute_loss=None,
+        callbacks=Callbacks(),
+):
+    if save_json:
+        check_requirements(['pycocotools'])
+        process = process_mask_upsample  # more accurate
+    else:
+        process = process_mask  # faster
+
+    # Initialize/load model and set device
+    training = model is not None
+    if training:  # called by train.py
+        device, pt, jit, engine = next(model.parameters()).device, True, False, False  # get model device, PyTorch model
+        half &= device.type != 'cpu'  # half precision only supported on CUDA
+        model.half() if half else model.float()
+        nm = de_parallel(model).model[-1].nm  # number of masks
+    else:  # called directly
+        device = select_device(device, batch_size=batch_size)
+
+        # Directories
+        save_dir = increment_path(Path(project) / name, exist_ok=exist_ok)  # increment run
+        (save_dir / 'labels' if save_txt else save_dir).mkdir(parents=True, exist_ok=True)  # make dir
+
+        # Load model
+        model = DetectMultiBackend(weights, device=device, dnn=dnn, data=data, fp16=half)
+        stride, pt, jit, engine = model.stride, model.pt, model.jit, model.engine
+        imgsz = check_img_size(imgsz, s=stride)  # check image size
+        half = model.fp16  # FP16 supported on limited backends with CUDA
+        nm = de_parallel(model).model.model[-1].nm if isinstance(model, SegmentationModel) else 32  # number of masks
+        if engine:
+            batch_size = model.batch_size
+        else:
+            device = model.device
+            if not (pt or jit):
+                batch_size = 1  # export.py models default to batch-size 1
+                LOGGER.info(f'Forcing --batch-size 1 square inference (1,3,{imgsz},{imgsz}) for non-PyTorch models')
+
+        # Data
+        data = check_dataset(data)  # check
+
+    # Configure
+    model.eval()
+    cuda = device.type != 'cpu'
+    #is_coco = isinstance(data.get('val'), str) and data['val'].endswith(f'coco{os.sep}val2017.txt')  # COCO dataset
+    is_coco = isinstance(data.get('val'), str) and data['val'].endswith(f'val2017.txt')  # COCO dataset
+    nc = 1 if single_cls else int(data['nc'])  # number of classes
+    iouv = torch.linspace(0.5, 0.95, 10, device=device)  # iou vector for [email protected]:0.95
+    niou = iouv.numel()
+
+    # Dataloader
+    if not training:
+        if pt and not single_cls:  # check --weights are trained on --data
+            ncm = model.model.nc
+            assert ncm == nc, f'{weights} ({ncm} classes) trained on different --data than what you passed ({nc} ' \
+                              f'classes). Pass correct combination of --weights and --data that are trained together.'
+        model.warmup(imgsz=(1 if pt else batch_size, 3, imgsz, imgsz))  # warmup
+        pad, rect = (0.0, False) if task == 'speed' else (0.5, pt)  # square inference for benchmarks
+        task = task if task in ('train', 'val', 'test') else 'val'  # path to train/val/test images
+        dataloader = create_dataloader(data[task],
+                                       imgsz,
+                                       batch_size,
+                                       stride,
+                                       single_cls,
+                                       pad=pad,
+                                       rect=rect,
+                                       workers=workers,
+                                       prefix=colorstr(f'{task}: '),
+                                       overlap_mask=overlap,
+                                       mask_downsample_ratio=mask_downsample_ratio)[0]
+
+    seen = 0
+    confusion_matrix = ConfusionMatrix(nc=nc)
+    names = model.names if hasattr(model, 'names') else model.module.names  # get class names
+    if isinstance(names, (list, tuple)):  # old format
+        names = dict(enumerate(names))
+    class_map = coco80_to_coco91_class() if is_coco else list(range(1000))
+    s = ('%22s' + '%11s' * 10) % ('Class', 'Images', 'Instances', 'Box(P', "R", "mAP50", "mAP50-95)", "Mask(P", "R",
+                                  "mAP50", "mAP50-95)")
+    dt = Profile(), Profile(), Profile()
+    metrics = Metrics()
+    loss = torch.zeros(4, device=device)
+    jdict, stats = [], []
+    # callbacks.run('on_val_start')
+    pbar = tqdm(dataloader, desc=s, bar_format=TQDM_BAR_FORMAT)  # progress bar
+    for batch_i, (im, targets, paths, shapes, masks) in enumerate(pbar):
+        # callbacks.run('on_val_batch_start')
+        with dt[0]:
+            if cuda:
+                im = im.to(device, non_blocking=True)
+                targets = targets.to(device)
+                masks = masks.to(device)
+            masks = masks.float()
+            im = im.half() if half else im.float()  # uint8 to fp16/32
+            im /= 255  # 0 - 255 to 0.0 - 1.0
+            nb, _, height, width = im.shape  # batch size, channels, height, width
+
+        # Inference
+        with dt[1]:
+            preds, train_out = model(im)# if compute_loss else (*model(im, augment=augment)[:2], None)
+            #train_out, preds, protos = p if len(p) == 3 else p[1]
+            #preds = p
+            #train_out = p[1][0] if len(p[1]) == 3 else p[0]
+            protos = train_out[-1]
+            #print(preds.shape)
+            #print(train_out[0].shape)
+            #print(train_out[1].shape)
+            #print(train_out[2].shape)
+
+        # Loss
+        if compute_loss:
+            loss += compute_loss(train_out, targets, masks)[1]  # box, obj, cls
+
+        # NMS
+        targets[:, 2:] *= torch.tensor((width, height, width, height), device=device)  # to pixels
+        lb = [targets[targets[:, 0] == i, 1:] for i in range(nb)] if save_hybrid else []  # for autolabelling
+        with dt[2]:
+            preds = non_max_suppression(preds,
+                                        conf_thres,
+                                        iou_thres,
+                                        labels=lb,
+                                        multi_label=True,
+                                        agnostic=single_cls,
+                                        max_det=max_det,
+                                        nm=nm)
+
+        # Metrics
+        plot_masks = []  # masks for plotting
+        for si, (pred, proto) in enumerate(zip(preds, protos)):
+            labels = targets[targets[:, 0] == si, 1:]
+            nl, npr = labels.shape[0], pred.shape[0]  # number of labels, predictions
+            path, shape = Path(paths[si]), shapes[si][0]
+            correct_masks = torch.zeros(npr, niou, dtype=torch.bool, device=device)  # init
+            correct_bboxes = torch.zeros(npr, niou, dtype=torch.bool, device=device)  # init
+            seen += 1
+
+            if npr == 0:
+                if nl:
+                    stats.append((correct_masks, correct_bboxes, *torch.zeros((2, 0), device=device), labels[:, 0]))
+                    if plots:
+                        confusion_matrix.process_batch(detections=None, labels=labels[:, 0])
+                continue
+
+            # Masks
+            midx = [si] if overlap else targets[:, 0] == si
+            gt_masks = masks[midx]
+            pred_masks = process(proto, pred[:, 6:], pred[:, :4], shape=im[si].shape[1:])
+
+            # Predictions
+            if single_cls:
+                pred[:, 5] = 0
+            predn = pred.clone()
+            scale_boxes(im[si].shape[1:], predn[:, :4], shape, shapes[si][1])  # native-space pred
+
+            # Evaluate
+            if nl:
+                tbox = xywh2xyxy(labels[:, 1:5])  # target boxes
+                scale_boxes(im[si].shape[1:], tbox, shape, shapes[si][1])  # native-space labels
+                labelsn = torch.cat((labels[:, 0:1], tbox), 1)  # native-space labels
+                correct_bboxes = process_batch(predn, labelsn, iouv)
+                correct_masks = process_batch(predn, labelsn, iouv, pred_masks, gt_masks, overlap=overlap, masks=True)
+                if plots:
+                    confusion_matrix.process_batch(predn, labelsn)
+            stats.append((correct_masks, correct_bboxes, pred[:, 4], pred[:, 5], labels[:, 0]))  # (conf, pcls, tcls)
+
+            pred_masks = torch.as_tensor(pred_masks, dtype=torch.uint8)
+            if plots and batch_i < 3:
+                plot_masks.append(pred_masks[:15].cpu())  # filter top 15 to plot
+
+            # Save/log
+            if save_txt:
+                save_one_txt(predn, save_conf, shape, file=save_dir / 'labels' / f'{path.stem}.txt')
+            if save_json:
+                pred_masks = scale_image(im[si].shape[1:],
+                                         pred_masks.permute(1, 2, 0).contiguous().cpu().numpy(), shape, shapes[si][1])
+                save_one_json(predn, jdict, path, class_map, pred_masks)  # append to COCO-JSON dictionary
+            # callbacks.run('on_val_image_end', pred, predn, path, names, im[si])
+
+        # Plot images
+        if plots and batch_i < 3:
+            if len(plot_masks):
+                plot_masks = torch.cat(plot_masks, dim=0)
+            plot_images_and_masks(im, targets, masks, paths, save_dir / f'val_batch{batch_i}_labels.jpg', names)
+            plot_images_and_masks(im, output_to_target(preds, max_det=15), plot_masks, paths,
+                                  save_dir / f'val_batch{batch_i}_pred.jpg', names)  # pred
+
+        # callbacks.run('on_val_batch_end')
+
+    # Compute metrics
+    stats = [torch.cat(x, 0).cpu().numpy() for x in zip(*stats)]  # to numpy
+    if len(stats) and stats[0].any():
+        results = ap_per_class_box_and_mask(*stats, plot=plots, save_dir=save_dir, names=names)
+        metrics.update(results)
+    nt = np.bincount(stats[4].astype(int), minlength=nc)  # number of targets per class
+
+    # Print results
+    pf = '%22s' + '%11i' * 2 + '%11.3g' * 8  # print format
+    LOGGER.info(pf % ("all", seen, nt.sum(), *metrics.mean_results()))
+    if nt.sum() == 0:
+        LOGGER.warning(f'WARNING ⚠️ no labels found in {task} set, can not compute metrics without labels')
+
+    # Print results per class
+    if (verbose or (nc < 50 and not training)) and nc > 1 and len(stats):
+        for i, c in enumerate(metrics.ap_class_index):
+            LOGGER.info(pf % (names[c], seen, nt[c], *metrics.class_result(i)))
+
+    # Print speeds
+    t = tuple(x.t / seen * 1E3 for x in dt)  # speeds per image
+    if not training:
+        shape = (batch_size, 3, imgsz, imgsz)
+        LOGGER.info(f'Speed: %.1fms pre-process, %.1fms inference, %.1fms NMS per image at shape {shape}' % t)
+
+    # Plots
+    if plots:
+        confusion_matrix.plot(save_dir=save_dir, names=list(names.values()))
+    # callbacks.run('on_val_end')
+
+    mp_bbox, mr_bbox, map50_bbox, map_bbox, mp_mask, mr_mask, map50_mask, map_mask = metrics.mean_results()
+
+    # Save JSON
+    if save_json and len(jdict):
+        w = Path(weights[0] if isinstance(weights, list) else weights).stem if weights is not None else ''  # weights
+        anno_json = str(Path(data.get('path', '../coco')) / 'annotations/instances_val2017.json')  # annotations json
+        pred_json = str(save_dir / f"{w}_predictions.json")  # predictions json
+        LOGGER.info(f'\nEvaluating pycocotools mAP... saving {pred_json}...')
+        with open(pred_json, 'w') as f:
+            json.dump(jdict, f)
+
+        try:  # https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocoEvalDemo.ipynb
+            from pycocotools.coco import COCO
+            from pycocotools.cocoeval import COCOeval
+
+            anno = COCO(anno_json)  # init annotations api
+            pred = anno.loadRes(pred_json)  # init predictions api
+            results = []
+            for eval in COCOeval(anno, pred, 'bbox'), COCOeval(anno, pred, 'segm'):
+                if is_coco:
+                    eval.params.imgIds = [int(Path(x).stem) for x in dataloader.dataset.im_files]  # img ID to evaluate
+                eval.evaluate()
+                eval.accumulate()
+                eval.summarize()
+                results.extend(eval.stats[:2])  # update results ([email protected]:0.95, [email protected])
+            map_bbox, map50_bbox, map_mask, map50_mask = results
+        except Exception as e:
+            LOGGER.info(f'pycocotools unable to run: {e}')
+
+    # Return results
+    model.float()  # for training
+    if not training:
+        s = f"\n{len(list(save_dir.glob('labels/*.txt')))} labels saved to {save_dir / 'labels'}" if save_txt else ''
+        LOGGER.info(f"Results saved to {colorstr('bold', save_dir)}{s}")
+    final_metric = mp_bbox, mr_bbox, map50_bbox, map_bbox, mp_mask, mr_mask, map50_mask, map_mask
+    return (*final_metric, *(loss.cpu() / len(dataloader)).tolist()), metrics.get_maps(nc), t
+
+
+def parse_opt():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--data', type=str, default=ROOT / 'data/coco128-seg.yaml', help='dataset.yaml path')
+    parser.add_argument('--weights', nargs='+', type=str, default=ROOT / 'yolo-seg.pt', help='model path(s)')
+    parser.add_argument('--batch-size', type=int, default=32, help='batch size')
+    parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=640, help='inference size (pixels)')
+    parser.add_argument('--conf-thres', type=float, default=0.001, help='confidence threshold')
+    parser.add_argument('--iou-thres', type=float, default=0.6, help='NMS IoU threshold')
+    parser.add_argument('--max-det', type=int, default=300, help='maximum detections per image')
+    parser.add_argument('--task', default='val', help='train, val, test, speed or study')
+    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
+    parser.add_argument('--workers', type=int, default=8, help='max dataloader workers (per RANK in DDP mode)')
+    parser.add_argument('--single-cls', action='store_true', help='treat as single-class dataset')
+    parser.add_argument('--augment', action='store_true', help='augmented inference')
+    parser.add_argument('--verbose', action='store_true', help='report mAP by class')
+    parser.add_argument('--save-txt', action='store_true', help='save results to *.txt')
+    parser.add_argument('--save-hybrid', action='store_true', help='save label+prediction hybrid results to *.txt')
+    parser.add_argument('--save-conf', action='store_true', help='save confidences in --save-txt labels')
+    parser.add_argument('--save-json', action='store_true', help='save a COCO-JSON results file')
+    parser.add_argument('--project', default=ROOT / 'runs/val-seg', help='save results to project/name')
+    parser.add_argument('--name', default='exp', help='save to project/name')
+    parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')
+    parser.add_argument('--half', action='store_true', help='use FP16 half-precision inference')
+    parser.add_argument('--dnn', action='store_true', help='use OpenCV DNN for ONNX inference')
+    opt = parser.parse_args()
+    opt.data = check_yaml(opt.data)  # check YAML
+    # opt.save_json |= opt.data.endswith('coco.yaml')
+    opt.save_txt |= opt.save_hybrid
+    print_args(vars(opt))
+    return opt
+
+
+def main(opt):
+    #check_requirements(requirements=ROOT / 'requirements.txt', exclude=('tensorboard', 'thop'))
+
+    if opt.task in ('train', 'val', 'test'):  # run normally
+        if opt.conf_thres > 0.001:  # https://github.com/ultralytics/yolov5/issues/1466
+            LOGGER.warning(f'WARNING ⚠️ confidence threshold {opt.conf_thres} > 0.001 produces invalid results')
+        if opt.save_hybrid:
+            LOGGER.warning('WARNING ⚠️ --save-hybrid returns high mAP from hybrid labels, not from predictions alone')
+        run(**vars(opt))
+
+    else:
+        weights = opt.weights if isinstance(opt.weights, list) else [opt.weights]
+        opt.half = torch.cuda.is_available() and opt.device != 'cpu'  # FP16 for fastest results
+        if opt.task == 'speed':  # speed benchmarks
+            # python val.py --task speed --data coco.yaml --batch 1 --weights yolo.pt...
+            opt.conf_thres, opt.iou_thres, opt.save_json = 0.25, 0.45, False
+            for opt.weights in weights:
+                run(**vars(opt), plots=False)
+
+        elif opt.task == 'study':  # speed vs mAP benchmarks
+            # python val.py --task study --data coco.yaml --iou 0.7 --weights yolo.pt...
+            for opt.weights in weights:
+                f = f'study_{Path(opt.data).stem}_{Path(opt.weights).stem}.txt'  # filename to save to
+                x, y = list(range(256, 1536 + 128, 128)), []  # x axis (image sizes), y axis
+                for opt.imgsz in x:  # img-size
+                    LOGGER.info(f'\nRunning {f} --imgsz {opt.imgsz}...')
+                    r, _, t = run(**vars(opt), plots=False)
+                    y.append(r + t)  # results and times
+                np.savetxt(f, y, fmt='%10.4g')  # save
+            os.system('zip -r study.zip study_*.txt')
+            plot_val_study(x=x)  # plot
+
+
+if __name__ == "__main__":
+    opt = parse_opt()
+    main(opt)

segment/val_dual.py

@@ -0,0 +1,458 @@
+import argparse
+import json
+import os
+import sys
+from multiprocessing.pool import ThreadPool
+from pathlib import Path
+
+import numpy as np
+import torch
+from tqdm import tqdm
+
+FILE = Path(__file__).resolve()
+ROOT = FILE.parents[1]  # YOLO root directory
+if str(ROOT) not in sys.path:
+    sys.path.append(str(ROOT))  # add ROOT to PATH
+ROOT = Path(os.path.relpath(ROOT, Path.cwd()))  # relative
+
+import torch.nn.functional as F
+
+from models.common import DetectMultiBackend
+from models.yolo import SegmentationModel
+from utils.callbacks import Callbacks
+from utils.general import (LOGGER, NUM_THREADS, TQDM_BAR_FORMAT, Profile, check_dataset, check_img_size,
+                           check_requirements, check_yaml, coco80_to_coco91_class, colorstr, increment_path,
+                           non_max_suppression, print_args, scale_boxes, xywh2xyxy, xyxy2xywh)
+from utils.metrics import ConfusionMatrix, box_iou
+from utils.plots import output_to_target, plot_val_study
+from utils.segment.dataloaders import create_dataloader
+from utils.segment.general import mask_iou, process_mask, process_mask_upsample, scale_image
+from utils.segment.metrics import Metrics, ap_per_class_box_and_mask
+from utils.segment.plots import plot_images_and_masks
+from utils.torch_utils import de_parallel, select_device, smart_inference_mode
+
+
+def save_one_txt(predn, save_conf, shape, file):
+    # Save one txt result
+    gn = torch.tensor(shape)[[1, 0, 1, 0]]  # normalization gain whwh
+    for *xyxy, conf, cls in predn.tolist():
+        xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist()  # normalized xywh
+        line = (cls, *xywh, conf) if save_conf else (cls, *xywh)  # label format
+        with open(file, 'a') as f:
+            f.write(('%g ' * len(line)).rstrip() % line + '\n')
+
+
+def save_one_json(predn, jdict, path, class_map, pred_masks):
+    # Save one JSON result {"image_id": 42, "category_id": 18, "bbox": [258.15, 41.29, 348.26, 243.78], "score": 0.236}
+    from pycocotools.mask import encode
+
+    def single_encode(x):
+        rle = encode(np.asarray(x[:, :, None], order="F", dtype="uint8"))[0]
+        rle["counts"] = rle["counts"].decode("utf-8")
+        return rle
+
+    image_id = int(path.stem) if path.stem.isnumeric() else path.stem
+    box = xyxy2xywh(predn[:, :4])  # xywh
+    box[:, :2] -= box[:, 2:] / 2  # xy center to top-left corner
+    pred_masks = np.transpose(pred_masks, (2, 0, 1))
+    with ThreadPool(NUM_THREADS) as pool:
+        rles = pool.map(single_encode, pred_masks)
+    for i, (p, b) in enumerate(zip(predn.tolist(), box.tolist())):
+        jdict.append({
+            'image_id': image_id,
+            'category_id': class_map[int(p[5])],
+            'bbox': [round(x, 3) for x in b],
+            'score': round(p[4], 5),
+            'segmentation': rles[i]})
+
+
+def process_batch(detections, labels, iouv, pred_masks=None, gt_masks=None, overlap=False, masks=False):
+    """
+    Return correct prediction matrix
+    Arguments:
+        detections (array[N, 6]), x1, y1, x2, y2, conf, class
+        labels (array[M, 5]), class, x1, y1, x2, y2
+    Returns:
+        correct (array[N, 10]), for 10 IoU levels
+    """
+    if masks:
+        if overlap:
+            nl = len(labels)
+            index = torch.arange(nl, device=gt_masks.device).view(nl, 1, 1) + 1
+            gt_masks = gt_masks.repeat(nl, 1, 1)  # shape(1,640,640) -> (n,640,640)
+            gt_masks = torch.where(gt_masks == index, 1.0, 0.0)
+        if gt_masks.shape[1:] != pred_masks.shape[1:]:
+            gt_masks = F.interpolate(gt_masks[None], pred_masks.shape[1:], mode="bilinear", align_corners=False)[0]
+            gt_masks = gt_masks.gt_(0.5)
+        iou = mask_iou(gt_masks.view(gt_masks.shape[0], -1), pred_masks.view(pred_masks.shape[0], -1))
+    else:  # boxes
+        iou = box_iou(labels[:, 1:], detections[:, :4])
+
+    correct = np.zeros((detections.shape[0], iouv.shape[0])).astype(bool)
+    correct_class = labels[:, 0:1] == detections[:, 5]
+    for i in range(len(iouv)):
+        x = torch.where((iou >= iouv[i]) & correct_class)  # IoU > threshold and classes match
+        if x[0].shape[0]:
+            matches = torch.cat((torch.stack(x, 1), iou[x[0], x[1]][:, None]), 1).cpu().numpy()  # [label, detect, iou]
+            if x[0].shape[0] > 1:
+                matches = matches[matches[:, 2].argsort()[::-1]]
+                matches = matches[np.unique(matches[:, 1], return_index=True)[1]]
+                # matches = matches[matches[:, 2].argsort()[::-1]]
+                matches = matches[np.unique(matches[:, 0], return_index=True)[1]]
+            correct[matches[:, 1].astype(int), i] = True
+    return torch.tensor(correct, dtype=torch.bool, device=iouv.device)
+
+
+@smart_inference_mode()
+def run(
+        data,
+        weights=None,  # model.pt path(s)
+        batch_size=32,  # batch size
+        imgsz=640,  # inference size (pixels)
+        conf_thres=0.001,  # confidence threshold
+        iou_thres=0.6,  # NMS IoU threshold
+        max_det=300,  # maximum detections per image
+        task='val',  # train, val, test, speed or study
+        device='',  # cuda device, i.e. 0 or 0,1,2,3 or cpu
+        workers=8,  # max dataloader workers (per RANK in DDP mode)
+        single_cls=False,  # treat as single-class dataset
+        augment=False,  # augmented inference
+        verbose=False,  # verbose output
+        save_txt=False,  # save results to *.txt
+        save_hybrid=False,  # save label+prediction hybrid results to *.txt
+        save_conf=False,  # save confidences in --save-txt labels
+        save_json=False,  # save a COCO-JSON results file
+        project=ROOT / 'runs/val-seg',  # save to project/name
+        name='exp',  # save to project/name
+        exist_ok=False,  # existing project/name ok, do not increment
+        half=True,  # use FP16 half-precision inference
+        dnn=False,  # use OpenCV DNN for ONNX inference
+        model=None,
+        dataloader=None,
+        save_dir=Path(''),
+        plots=True,
+        overlap=False,
+        mask_downsample_ratio=1,
+        compute_loss=None,
+        callbacks=Callbacks(),
+):
+    if save_json:
+        check_requirements(['pycocotools'])
+        process = process_mask_upsample  # more accurate
+    else:
+        process = process_mask  # faster
+
+    # Initialize/load model and set device
+    training = model is not None
+    if training:  # called by train.py
+        device, pt, jit, engine = next(model.parameters()).device, True, False, False  # get model device, PyTorch model
+        half &= device.type != 'cpu'  # half precision only supported on CUDA
+        model.half() if half else model.float()
+        nm = de_parallel(model).model[-1].nm  # number of masks
+    else:  # called directly
+        device = select_device(device, batch_size=batch_size)
+
+        # Directories
+        save_dir = increment_path(Path(project) / name, exist_ok=exist_ok)  # increment run
+        (save_dir / 'labels' if save_txt else save_dir).mkdir(parents=True, exist_ok=True)  # make dir
+
+        # Load model
+        model = DetectMultiBackend(weights, device=device, dnn=dnn, data=data, fp16=half)
+        stride, pt, jit, engine = model.stride, model.pt, model.jit, model.engine
+        imgsz = check_img_size(imgsz, s=stride)  # check image size
+        half = model.fp16  # FP16 supported on limited backends with CUDA
+        nm = de_parallel(model).model.model[-1].nm if isinstance(model, SegmentationModel) else 32  # number of masks
+        if engine:
+            batch_size = model.batch_size
+        else:
+            device = model.device
+            if not (pt or jit):
+                batch_size = 1  # export.py models default to batch-size 1
+                LOGGER.info(f'Forcing --batch-size 1 square inference (1,3,{imgsz},{imgsz}) for non-PyTorch models')
+
+        # Data
+        data = check_dataset(data)  # check
+
+    # Configure
+    model.eval()
+    cuda = device.type != 'cpu'
+    #is_coco = isinstance(data.get('val'), str) and data['val'].endswith(f'coco{os.sep}val2017.txt')  # COCO dataset
+    is_coco = isinstance(data.get('val'), str) and data['val'].endswith(f'val2017.txt')  # COCO dataset
+    nc = 1 if single_cls else int(data['nc'])  # number of classes
+    iouv = torch.linspace(0.5, 0.95, 10, device=device)  # iou vector for [email protected]:0.95
+    niou = iouv.numel()
+
+    # Dataloader
+    if not training:
+        if pt and not single_cls:  # check --weights are trained on --data
+            ncm = model.model.nc
+            assert ncm == nc, f'{weights} ({ncm} classes) trained on different --data than what you passed ({nc} ' \
+                              f'classes). Pass correct combination of --weights and --data that are trained together.'
+        model.warmup(imgsz=(1 if pt else batch_size, 3, imgsz, imgsz))  # warmup
+        pad, rect = (0.0, False) if task == 'speed' else (0.5, pt)  # square inference for benchmarks
+        task = task if task in ('train', 'val', 'test') else 'val'  # path to train/val/test images
+        dataloader = create_dataloader(data[task],
+                                       imgsz,
+                                       batch_size,
+                                       stride,
+                                       single_cls,
+                                       pad=pad,
+                                       rect=rect,
+                                       workers=workers,
+                                       prefix=colorstr(f'{task}: '),
+                                       overlap_mask=overlap,
+                                       mask_downsample_ratio=mask_downsample_ratio)[0]
+
+    seen = 0
+    confusion_matrix = ConfusionMatrix(nc=nc)
+    names = model.names if hasattr(model, 'names') else model.module.names  # get class names
+    if isinstance(names, (list, tuple)):  # old format
+        names = dict(enumerate(names))
+    class_map = coco80_to_coco91_class() if is_coco else list(range(1000))
+    s = ('%22s' + '%11s' * 10) % ('Class', 'Images', 'Instances', 'Box(P', "R", "mAP50", "mAP50-95)", "Mask(P", "R",
+                                  "mAP50", "mAP50-95)")
+    dt = Profile(), Profile(), Profile()
+    metrics = Metrics()
+    loss = torch.zeros(4, device=device)
+    jdict, stats = [], []
+    # callbacks.run('on_val_start')
+    pbar = tqdm(dataloader, desc=s, bar_format=TQDM_BAR_FORMAT)  # progress bar
+    for batch_i, (im, targets, paths, shapes, masks) in enumerate(pbar):
+        # callbacks.run('on_val_batch_start')
+        with dt[0]:
+            if cuda:
+                im = im.to(device, non_blocking=True)
+                targets = targets.to(device)
+                masks = masks.to(device)
+            masks = masks.float()
+            im = im.half() if half else im.float()  # uint8 to fp16/32
+            im /= 255  # 0 - 255 to 0.0 - 1.0
+            nb, _, height, width = im.shape  # batch size, channels, height, width
+
+        # Inference
+        with dt[1]:
+            preds, train_out = model(im)# if compute_loss else (*model(im, augment=augment)[:2], None)
+            #preds = preds[1]
+            #train_out, preds, protos = p if len(p) == 3 else p[1]
+            #preds = p
+            #train_out = p[1][0] if len(p[1]) == 3 else p[0]
+            protos = train_out[-1]
+            #print(preds.shape)
+            #print(train_out[0].shape)
+            #print(train_out[1].shape)
+            #print(train_out[2].shape)
+
+        # Loss
+        #if compute_loss:
+        #    loss += compute_loss(train_out, targets, masks)[1]  # box, obj, cls
+
+        # NMS
+        targets[:, 2:] *= torch.tensor((width, height, width, height), device=device)  # to pixels
+        lb = [targets[targets[:, 0] == i, 1:] for i in range(nb)] if save_hybrid else []  # for autolabelling
+        with dt[2]:
+            preds = non_max_suppression(preds,
+                                        conf_thres,
+                                        iou_thres,
+                                        labels=lb,
+                                        multi_label=True,
+                                        agnostic=single_cls,
+                                        max_det=max_det,
+                                        nm=nm)
+
+        # Metrics
+        plot_masks = []  # masks for plotting
+        for si, (pred, proto) in enumerate(zip(preds, protos)):
+            labels = targets[targets[:, 0] == si, 1:]
+            nl, npr = labels.shape[0], pred.shape[0]  # number of labels, predictions
+            path, shape = Path(paths[si]), shapes[si][0]
+            correct_masks = torch.zeros(npr, niou, dtype=torch.bool, device=device)  # init
+            correct_bboxes = torch.zeros(npr, niou, dtype=torch.bool, device=device)  # init
+            seen += 1
+
+            if npr == 0:
+                if nl:
+                    stats.append((correct_masks, correct_bboxes, *torch.zeros((2, 0), device=device), labels[:, 0]))
+                    if plots:
+                        confusion_matrix.process_batch(detections=None, labels=labels[:, 0])
+                continue
+
+            # Masks
+            midx = [si] if overlap else targets[:, 0] == si
+            gt_masks = masks[midx]
+            pred_masks = process(proto, pred[:, 6:], pred[:, :4], shape=im[si].shape[1:])
+
+            # Predictions
+            if single_cls:
+                pred[:, 5] = 0
+            predn = pred.clone()
+            scale_boxes(im[si].shape[1:], predn[:, :4], shape, shapes[si][1])  # native-space pred
+
+            # Evaluate
+            if nl:
+                tbox = xywh2xyxy(labels[:, 1:5])  # target boxes
+                scale_boxes(im[si].shape[1:], tbox, shape, shapes[si][1])  # native-space labels
+                labelsn = torch.cat((labels[:, 0:1], tbox), 1)  # native-space labels
+                correct_bboxes = process_batch(predn, labelsn, iouv)
+                correct_masks = process_batch(predn, labelsn, iouv, pred_masks, gt_masks, overlap=overlap, masks=True)
+                if plots:
+                    confusion_matrix.process_batch(predn, labelsn)
+            stats.append((correct_masks, correct_bboxes, pred[:, 4], pred[:, 5], labels[:, 0]))  # (conf, pcls, tcls)
+
+            pred_masks = torch.as_tensor(pred_masks, dtype=torch.uint8)
+            if plots and batch_i < 3:
+                plot_masks.append(pred_masks[:15].cpu())  # filter top 15 to plot
+
+            # Save/log
+            if save_txt:
+                save_one_txt(predn, save_conf, shape, file=save_dir / 'labels' / f'{path.stem}.txt')
+            if save_json:
+                pred_masks = scale_image(im[si].shape[1:],
+                                         pred_masks.permute(1, 2, 0).contiguous().cpu().numpy(), shape, shapes[si][1])
+                save_one_json(predn, jdict, path, class_map, pred_masks)  # append to COCO-JSON dictionary
+            # callbacks.run('on_val_image_end', pred, predn, path, names, im[si])
+
+        # Plot images
+        if plots and batch_i < 3:
+            if len(plot_masks):
+                plot_masks = torch.cat(plot_masks, dim=0)
+            plot_images_and_masks(im, targets, masks, paths, save_dir / f'val_batch{batch_i}_labels.jpg', names)
+            plot_images_and_masks(im, output_to_target(preds, max_det=15), plot_masks, paths,
+                                  save_dir / f'val_batch{batch_i}_pred.jpg', names)  # pred
+
+        # callbacks.run('on_val_batch_end')
+
+    # Compute metrics
+    stats = [torch.cat(x, 0).cpu().numpy() for x in zip(*stats)]  # to numpy
+    if len(stats) and stats[0].any():
+        results = ap_per_class_box_and_mask(*stats, plot=plots, save_dir=save_dir, names=names)
+        metrics.update(results)
+    nt = np.bincount(stats[4].astype(int), minlength=nc)  # number of targets per class
+
+    # Print results
+    pf = '%22s' + '%11i' * 2 + '%11.3g' * 8  # print format
+    LOGGER.info(pf % ("all", seen, nt.sum(), *metrics.mean_results()))
+    if nt.sum() == 0:
+        LOGGER.warning(f'WARNING ⚠️ no labels found in {task} set, can not compute metrics without labels')
+
+    # Print results per class
+    if (verbose or (nc < 50 and not training)) and nc > 1 and len(stats):
+        for i, c in enumerate(metrics.ap_class_index):
+            LOGGER.info(pf % (names[c], seen, nt[c], *metrics.class_result(i)))
+
+    # Print speeds
+    t = tuple(x.t / seen * 1E3 for x in dt)  # speeds per image
+    if not training:
+        shape = (batch_size, 3, imgsz, imgsz)
+        LOGGER.info(f'Speed: %.1fms pre-process, %.1fms inference, %.1fms NMS per image at shape {shape}' % t)
+
+    # Plots
+    if plots:
+        confusion_matrix.plot(save_dir=save_dir, names=list(names.values()))
+    # callbacks.run('on_val_end')
+
+    mp_bbox, mr_bbox, map50_bbox, map_bbox, mp_mask, mr_mask, map50_mask, map_mask = metrics.mean_results()
+
+    # Save JSON
+    if save_json and len(jdict):
+        w = Path(weights[0] if isinstance(weights, list) else weights).stem if weights is not None else ''  # weights
+        anno_json = str(Path(data.get('path', '../coco')) / 'annotations/instances_val2017.json')  # annotations json
+        pred_json = str(save_dir / f"{w}_predictions.json")  # predictions json
+        LOGGER.info(f'\nEvaluating pycocotools mAP... saving {pred_json}...')
+        with open(pred_json, 'w') as f:
+            json.dump(jdict, f)
+
+        try:  # https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocoEvalDemo.ipynb
+            from pycocotools.coco import COCO
+            from pycocotools.cocoeval import COCOeval
+
+            anno = COCO(anno_json)  # init annotations api
+            pred = anno.loadRes(pred_json)  # init predictions api
+            results = []
+            for eval in COCOeval(anno, pred, 'bbox'), COCOeval(anno, pred, 'segm'):
+                if is_coco:
+                    eval.params.imgIds = [int(Path(x).stem) for x in dataloader.dataset.im_files]  # img ID to evaluate
+                eval.evaluate()
+                eval.accumulate()
+                eval.summarize()
+                results.extend(eval.stats[:2])  # update results ([email protected]:0.95, [email protected])
+            map_bbox, map50_bbox, map_mask, map50_mask = results
+        except Exception as e:
+            LOGGER.info(f'pycocotools unable to run: {e}')
+
+    # Return results
+    model.float()  # for training
+    if not training:
+        s = f"\n{len(list(save_dir.glob('labels/*.txt')))} labels saved to {save_dir / 'labels'}" if save_txt else ''
+        LOGGER.info(f"Results saved to {colorstr('bold', save_dir)}{s}")
+    final_metric = mp_bbox, mr_bbox, map50_bbox, map_bbox, mp_mask, mr_mask, map50_mask, map_mask
+    return (*final_metric, *(loss.cpu() / len(dataloader)).tolist()), metrics.get_maps(nc), t
+
+
+def parse_opt():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--data', type=str, default=ROOT / 'data/coco128-seg.yaml', help='dataset.yaml path')
+    parser.add_argument('--weights', nargs='+', type=str, default=ROOT / 'yolo-seg.pt', help='model path(s)')
+    parser.add_argument('--batch-size', type=int, default=32, help='batch size')
+    parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=640, help='inference size (pixels)')
+    parser.add_argument('--conf-thres', type=float, default=0.001, help='confidence threshold')
+    parser.add_argument('--iou-thres', type=float, default=0.6, help='NMS IoU threshold')
+    parser.add_argument('--max-det', type=int, default=300, help='maximum detections per image')
+    parser.add_argument('--task', default='val', help='train, val, test, speed or study')
+    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
+    parser.add_argument('--workers', type=int, default=8, help='max dataloader workers (per RANK in DDP mode)')
+    parser.add_argument('--single-cls', action='store_true', help='treat as single-class dataset')
+    parser.add_argument('--augment', action='store_true', help='augmented inference')
+    parser.add_argument('--verbose', action='store_true', help='report mAP by class')
+    parser.add_argument('--save-txt', action='store_true', help='save results to *.txt')
+    parser.add_argument('--save-hybrid', action='store_true', help='save label+prediction hybrid results to *.txt')
+    parser.add_argument('--save-conf', action='store_true', help='save confidences in --save-txt labels')
+    parser.add_argument('--save-json', action='store_true', help='save a COCO-JSON results file')
+    parser.add_argument('--project', default=ROOT / 'runs/val-seg', help='save results to project/name')
+    parser.add_argument('--name', default='exp', help='save to project/name')
+    parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')
+    parser.add_argument('--half', action='store_true', help='use FP16 half-precision inference')
+    parser.add_argument('--dnn', action='store_true', help='use OpenCV DNN for ONNX inference')
+    opt = parser.parse_args()
+    opt.data = check_yaml(opt.data)  # check YAML
+    # opt.save_json |= opt.data.endswith('coco.yaml')
+    opt.save_txt |= opt.save_hybrid
+    print_args(vars(opt))
+    return opt
+
+
+def main(opt):
+    #check_requirements(requirements=ROOT / 'requirements.txt', exclude=('tensorboard', 'thop'))
+
+    if opt.task in ('train', 'val', 'test'):  # run normally
+        if opt.conf_thres > 0.001:  # https://github.com/ultralytics/yolov5/issues/1466
+            LOGGER.warning(f'WARNING ⚠️ confidence threshold {opt.conf_thres} > 0.001 produces invalid results')
+        if opt.save_hybrid:
+            LOGGER.warning('WARNING ⚠️ --save-hybrid returns high mAP from hybrid labels, not from predictions alone')
+        run(**vars(opt))
+
+    else:
+        weights = opt.weights if isinstance(opt.weights, list) else [opt.weights]
+        opt.half = torch.cuda.is_available() and opt.device != 'cpu'  # FP16 for fastest results
+        if opt.task == 'speed':  # speed benchmarks
+            # python val.py --task speed --data coco.yaml --batch 1 --weights yolo.pt...
+            opt.conf_thres, opt.iou_thres, opt.save_json = 0.25, 0.45, False
+            for opt.weights in weights:
+                run(**vars(opt), plots=False)
+
+        elif opt.task == 'study':  # speed vs mAP benchmarks
+            # python val.py --task study --data coco.yaml --iou 0.7 --weights yolo.pt...
+            for opt.weights in weights:
+                f = f'study_{Path(opt.data).stem}_{Path(opt.weights).stem}.txt'  # filename to save to
+                x, y = list(range(256, 1536 + 128, 128)), []  # x axis (image sizes), y axis
+                for opt.imgsz in x:  # img-size
+                    LOGGER.info(f'\nRunning {f} --imgsz {opt.imgsz}...')
+                    r, _, t = run(**vars(opt), plots=False)
+                    y.append(r + t)  # results and times
+                np.savetxt(f, y, fmt='%10.4g')  # save
+            os.system('zip -r study.zip study_*.txt')
+            plot_val_study(x=x)  # plot
+
+
+if __name__ == "__main__":
+    opt = parse_opt()
+    main(opt)

tools/reparameterization.ipynb

@@ -0,0 +1,450 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "4beac401",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import torch\n",
+    "from models.yolo import Model"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "d1a8399f",
+   "metadata": {},
+   "source": [
+    "## Convert YOLOv9-S"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "c7a40f10",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "device = torch.device(\"cpu\")\n",
+    "cfg = \"./models/detect/gelan-s.yaml\"\n",
+    "model = Model(cfg, ch=3, nc=80, anchors=3)\n",
+    "#model = model.half()\n",
+    "model = model.to(device)\n",
+    "_ = model.eval()\n",
+    "ckpt = torch.load('./yolov9-s.pt', map_location='cpu')\n",
+    "model.names = ckpt['model'].names\n",
+    "model.nc = ckpt['model'].nc"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "3b046bb2",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "idx = 0\n",
+    "for k, v in model.state_dict().items():\n",
+    "    if \"model.{}.\".format(idx) in k:\n",
+    "        if idx < 22:\n",
+    "            kr = k.replace(\"model.{}.\".format(idx), \"model.{}.\".format(idx))\n",
+    "            model.state_dict()[k] -= model.state_dict()[k]\n",
+    "            model.state_dict()[k] += ckpt['model'].state_dict()[kr]\n",
+    "            print(k, \"perfectly matched!!\")\n",
+    "        elif \"model.{}.cv2.\".format(idx) in k:\n",
+    "            kr = k.replace(\"model.{}.cv2.\".format(idx), \"model.{}.cv4.\".format(idx+7))\n",
+    "            model.state_dict()[k] -= model.state_dict()[k]\n",
+    "            model.state_dict()[k] += ckpt['model'].state_dict()[kr]\n",
+    "            print(k, \"perfectly matched!!\")\n",
+    "        elif \"model.{}.cv3.\".format(idx) in k:\n",
+    "            kr = k.replace(\"model.{}.cv3.\".format(idx), \"model.{}.cv5.\".format(idx+7))\n",
+    "            model.state_dict()[k] -= model.state_dict()[k]\n",
+    "            model.state_dict()[k] += ckpt['model'].state_dict()[kr]\n",
+    "            print(k, \"perfectly matched!!\")\n",
+    "        elif \"model.{}.dfl.\".format(idx) in k:\n",
+    "            kr = k.replace(\"model.{}.dfl.\".format(idx), \"model.{}.dfl2.\".format(idx+7))\n",
+    "            model.state_dict()[k] -= model.state_dict()[k]\n",
+    "            model.state_dict()[k] += ckpt['model'].state_dict()[kr]\n",
+    "            print(k, \"perfectly matched!!\")\n",
+    "    else:\n",
+    "        while True:\n",
+    "            idx += 1\n",
+    "            if \"model.{}.\".format(idx) in k:\n",
+    "                break\n",
+    "        if idx < 22:\n",
+    "            kr = k.replace(\"model.{}.\".format(idx), \"model.{}.\".format(idx))\n",
+    "            model.state_dict()[k] -= model.state_dict()[k]\n",
+    "            model.state_dict()[k] += ckpt['model'].state_dict()[kr]\n",
+    "            print(k, \"perfectly matched!!\")\n",
+    "        elif \"model.{}.cv2.\".format(idx) in k:\n",
+    "            kr = k.replace(\"model.{}.cv2.\".format(idx), \"model.{}.cv4.\".format(idx+7))\n",
+    "            model.state_dict()[k] -= model.state_dict()[k]\n",
+    "            model.state_dict()[k] += ckpt['model'].state_dict()[kr]\n",
+    "            print(k, \"perfectly matched!!\")\n",
+    "        elif \"model.{}.cv3.\".format(idx) in k:\n",
+    "            kr = k.replace(\"model.{}.cv3.\".format(idx), \"model.{}.cv5.\".format(idx+7))\n",
+    "            model.state_dict()[k] -= model.state_dict()[k]\n",
+    "            model.state_dict()[k] += ckpt['model'].state_dict()[kr]\n",
+    "            print(k, \"perfectly matched!!\")\n",
+    "        elif \"model.{}.dfl.\".format(idx) in k:\n",
+    "            kr = k.replace(\"model.{}.dfl.\".format(idx), \"model.{}.dfl2.\".format(idx+7))\n",
+    "            model.state_dict()[k] -= model.state_dict()[k]\n",
+    "            model.state_dict()[k] += ckpt['model'].state_dict()[kr]\n",
+    "            print(k, \"perfectly matched!!\")\n",
+    "_ = model.eval()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "07eb0cde",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "m_ckpt = {'model': model.half(),\n",
+    "          'optimizer': None,\n",
+    "          'best_fitness': None,\n",
+    "          'ema': None,\n",
+    "          'updates': None,\n",
+    "          'opt': None,\n",
+    "          'git': None,\n",
+    "          'date': None,\n",
+    "          'epoch': -1}\n",
+    "torch.save(m_ckpt, \"./yolov9-s-converted.pt\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "ba87d10f",
+   "metadata": {},
+   "source": [
+    "## Convert YOLOv9-M"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "cc41b027",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "device = torch.device(\"cpu\")\n",
+    "cfg = \"./models/detect/gelan-m.yaml\"\n",
+    "model = Model(cfg, ch=3, nc=80, anchors=3)\n",
+    "#model = model.half()\n",
+    "model = model.to(device)\n",
+    "_ = model.eval()\n",
+    "ckpt = torch.load('./yolov9-m.pt', map_location='cpu')\n",
+    "model.names = ckpt['model'].names\n",
+    "model.nc = ckpt['model'].nc"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "cf7c3978",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "idx = 0\n",
+    "for k, v in model.state_dict().items():\n",
+    "    if \"model.{}.\".format(idx) in k:\n",
+    "        if idx < 22:\n",
+    "            kr = k.replace(\"model.{}.\".format(idx), \"model.{}.\".format(idx+1))\n",
+    "            model.state_dict()[k] -= model.state_dict()[k]\n",
+    "            model.state_dict()[k] += ckpt['model'].state_dict()[kr]\n",
+    "            print(k, \"perfectly matched!!\")\n",
+    "        elif \"model.{}.cv2.\".format(idx) in k:\n",
+    "            kr = k.replace(\"model.{}.cv2.\".format(idx), \"model.{}.cv4.\".format(idx+16))\n",
+    "            model.state_dict()[k] -= model.state_dict()[k]\n",
+    "            model.state_dict()[k] += ckpt['model'].state_dict()[kr]\n",
+    "            print(k, \"perfectly matched!!\")\n",
+    "        elif \"model.{}.cv3.\".format(idx) in k:\n",
+    "            kr = k.replace(\"model.{}.cv3.\".format(idx), \"model.{}.cv5.\".format(idx+16))\n",
+    "            model.state_dict()[k] -= model.state_dict()[k]\n",
+    "            model.state_dict()[k] += ckpt['model'].state_dict()[kr]\n",
+    "            print(k, \"perfectly matched!!\")\n",
+    "        elif \"model.{}.dfl.\".format(idx) in k:\n",
+    "            kr = k.replace(\"model.{}.dfl.\".format(idx), \"model.{}.dfl2.\".format(idx+16))\n",
+    "            model.state_dict()[k] -= model.state_dict()[k]\n",
+    "            model.state_dict()[k] += ckpt['model'].state_dict()[kr]\n",
+    "            print(k, \"perfectly matched!!\")\n",
+    "    else:\n",
+    "        while True:\n",
+    "            idx += 1\n",
+    "            if \"model.{}.\".format(idx) in k:\n",
+    "                break\n",
+    "        if idx < 22:\n",
+    "            kr = k.replace(\"model.{}.\".format(idx), \"model.{}.\".format(idx+1))\n",
+    "            model.state_dict()[k] -= model.state_dict()[k]\n",
+    "            model.state_dict()[k] += ckpt['model'].state_dict()[kr]\n",
+    "            print(k, \"perfectly matched!!\")\n",
+    "        elif \"model.{}.cv2.\".format(idx) in k:\n",
+    "            kr = k.replace(\"model.{}.cv2.\".format(idx), \"model.{}.cv4.\".format(idx+16))\n",
+    "            model.state_dict()[k] -= model.state_dict()[k]\n",
+    "            model.state_dict()[k] += ckpt['model'].state_dict()[kr]\n",
+    "            print(k, \"perfectly matched!!\")\n",
+    "        elif \"model.{}.cv3.\".format(idx) in k:\n",
+    "            kr = k.replace(\"model.{}.cv3.\".format(idx), \"model.{}.cv5.\".format(idx+16))\n",
+    "            model.state_dict()[k] -= model.state_dict()[k]\n",
+    "            model.state_dict()[k] += ckpt['model'].state_dict()[kr]\n",
+    "            print(k, \"perfectly matched!!\")\n",
+    "        elif \"model.{}.dfl.\".format(idx) in k:\n",
+    "            kr = k.replace(\"model.{}.dfl.\".format(idx), \"model.{}.dfl2.\".format(idx+16))\n",
+    "            model.state_dict()[k] -= model.state_dict()[k]\n",
+    "            model.state_dict()[k] += ckpt['model'].state_dict()[kr]\n",
+    "            print(k, \"perfectly matched!!\")\n",
+    "_ = model.eval()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "00a92a45",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "m_ckpt = {'model': model.half(),\n",
+    "          'optimizer': None,\n",
+    "          'best_fitness': None,\n",
+    "          'ema': None,\n",
+    "          'updates': None,\n",
+    "          'opt': None,\n",
+    "          'git': None,\n",
+    "          'date': None,\n",
+    "          'epoch': -1}\n",
+    "torch.save(m_ckpt, \"./yolov9-m-converted.pt\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "8680f822",
+   "metadata": {},
+   "source": [
+    "## Convert YOLOv9-C"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "59f0198d",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "device = torch.device(\"cpu\")\n",
+    "cfg = \"./models/detect/gelan-c.yaml\"\n",
+    "model = Model(cfg, ch=3, nc=80, anchors=3)\n",
+    "#model = model.half()\n",
+    "model = model.to(device)\n",
+    "_ = model.eval()\n",
+    "ckpt = torch.load('./yolov9-c.pt', map_location='cpu')\n",
+    "model.names = ckpt['model'].names\n",
+    "model.nc = ckpt['model'].nc"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "2de7e1be",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "idx = 0\n",
+    "for k, v in model.state_dict().items():\n",
+    "    if \"model.{}.\".format(idx) in k:\n",
+    "        if idx < 22:\n",
+    "            kr = k.replace(\"model.{}.\".format(idx), \"model.{}.\".format(idx+1))\n",
+    "            model.state_dict()[k] -= model.state_dict()[k]\n",
+    "            model.state_dict()[k] += ckpt['model'].state_dict()[kr]\n",
+    "        elif \"model.{}.cv2.\".format(idx) in k:\n",
+    "            kr = k.replace(\"model.{}.cv2.\".format(idx), \"model.{}.cv4.\".format(idx+16))\n",
+    "            model.state_dict()[k] -= model.state_dict()[k]\n",
+    "            model.state_dict()[k] += ckpt['model'].state_dict()[kr]\n",
+    "        elif \"model.{}.cv3.\".format(idx) in k:\n",
+    "            kr = k.replace(\"model.{}.cv3.\".format(idx), \"model.{}.cv5.\".format(idx+16))\n",
+    "            model.state_dict()[k] -= model.state_dict()[k]\n",
+    "            model.state_dict()[k] += ckpt['model'].state_dict()[kr]\n",
+    "        elif \"model.{}.dfl.\".format(idx) in k:\n",
+    "            kr = k.replace(\"model.{}.dfl.\".format(idx), \"model.{}.dfl2.\".format(idx+16))\n",
+    "            model.state_dict()[k] -= model.state_dict()[k]\n",
+    "            model.state_dict()[k] += ckpt['model'].state_dict()[kr]\n",
+    "    else:\n",
+    "        while True:\n",
+    "            idx += 1\n",
+    "            if \"model.{}.\".format(idx) in k:\n",
+    "                break\n",
+    "        if idx < 22:\n",
+    "            kr = k.replace(\"model.{}.\".format(idx), \"model.{}.\".format(idx+1))\n",
+    "            model.state_dict()[k] -= model.state_dict()[k]\n",
+    "            model.state_dict()[k] += ckpt['model'].state_dict()[kr]\n",
+    "        elif \"model.{}.cv2.\".format(idx) in k:\n",
+    "            kr = k.replace(\"model.{}.cv2.\".format(idx), \"model.{}.cv4.\".format(idx+16))\n",
+    "            model.state_dict()[k] -= model.state_dict()[k]\n",
+    "            model.state_dict()[k] += ckpt['model'].state_dict()[kr]\n",
+    "        elif \"model.{}.cv3.\".format(idx) in k:\n",
+    "            kr = k.replace(\"model.{}.cv3.\".format(idx), \"model.{}.cv5.\".format(idx+16))\n",
+    "            model.state_dict()[k] -= model.state_dict()[k]\n",
+    "            model.state_dict()[k] += ckpt['model'].state_dict()[kr]\n",
+    "        elif \"model.{}.dfl.\".format(idx) in k:\n",
+    "            kr = k.replace(\"model.{}.dfl.\".format(idx), \"model.{}.dfl2.\".format(idx+16))\n",
+    "            model.state_dict()[k] -= model.state_dict()[k]\n",
+    "            model.state_dict()[k] += ckpt['model'].state_dict()[kr]\n",
+    "_ = model.eval()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "960796e3",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "m_ckpt = {'model': model.half(),\n",
+    "          'optimizer': None,\n",
+    "          'best_fitness': None,\n",
+    "          'ema': None,\n",
+    "          'updates': None,\n",
+    "          'opt': None,\n",
+    "          'git': None,\n",
+    "          'date': None,\n",
+    "          'epoch': -1}\n",
+    "torch.save(m_ckpt, \"./yolov9-c-converted.pt\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "47c6e6ae",
+   "metadata": {},
+   "source": [
+    "## Convert YOLOv9-E"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "801a1b7c",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "device = torch.device(\"cpu\")\n",
+    "cfg = \"./models/detect/gelan-e.yaml\"\n",
+    "model = Model(cfg, ch=3, nc=80, anchors=3)\n",
+    "#model = model.half()\n",
+    "model = model.to(device)\n",
+    "_ = model.eval()\n",
+    "ckpt = torch.load('./yolov9-e.pt', map_location='cpu')\n",
+    "model.names = ckpt['model'].names\n",
+    "model.nc = ckpt['model'].nc"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "a2ef4fe6",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "idx = 0\n",
+    "for k, v in model.state_dict().items():\n",
+    "    if \"model.{}.\".format(idx) in k:\n",
+    "        if idx < 29:\n",
+    "            kr = k.replace(\"model.{}.\".format(idx), \"model.{}.\".format(idx))\n",
+    "            model.state_dict()[k] -= model.state_dict()[k]\n",
+    "            model.state_dict()[k] += ckpt['model'].state_dict()[kr]\n",
+    "            print(k, \"perfectly matched!!\")\n",
+    "        elif idx < 42:\n",
+    "            kr = k.replace(\"model.{}.\".format(idx), \"model.{}.\".format(idx+7))\n",
+    "            model.state_dict()[k] -= model.state_dict()[k]\n",
+    "            model.state_dict()[k] += ckpt['model'].state_dict()[kr]\n",
+    "            print(k, \"perfectly matched!!\")\n",
+    "        elif \"model.{}.cv2.\".format(idx) in k:\n",
+    "            kr = k.replace(\"model.{}.cv2.\".format(idx), \"model.{}.cv4.\".format(idx+7))\n",
+    "            model.state_dict()[k] -= model.state_dict()[k]\n",
+    "            model.state_dict()[k] += ckpt['model'].state_dict()[kr]\n",
+    "            print(k, \"perfectly matched!!\")\n",
+    "        elif \"model.{}.cv3.\".format(idx) in k:\n",
+    "            kr = k.replace(\"model.{}.cv3.\".format(idx), \"model.{}.cv5.\".format(idx+7))\n",
+    "            model.state_dict()[k] -= model.state_dict()[k]\n",
+    "            model.state_dict()[k] += ckpt['model'].state_dict()[kr]\n",
+    "            print(k, \"perfectly matched!!\")\n",
+    "        elif \"model.{}.dfl.\".format(idx) in k:\n",
+    "            kr = k.replace(\"model.{}.dfl.\".format(idx), \"model.{}.dfl2.\".format(idx+7))\n",
+    "            model.state_dict()[k] -= model.state_dict()[k]\n",
+    "            model.state_dict()[k] += ckpt['model'].state_dict()[kr]\n",
+    "            print(k, \"perfectly matched!!\")\n",
+    "    else:\n",
+    "        while True:\n",
+    "            idx += 1\n",
+    "            if \"model.{}.\".format(idx) in k:\n",
+    "                break\n",
+    "        if idx < 29:\n",
+    "            kr = k.replace(\"model.{}.\".format(idx), \"model.{}.\".format(idx))\n",
+    "            model.state_dict()[k] -= model.state_dict()[k]\n",
+    "            model.state_dict()[k] += ckpt['model'].state_dict()[kr]\n",
+    "            print(k, \"perfectly matched!!\")\n",
+    "        elif idx < 42:\n",
+    "            kr = k.replace(\"model.{}.\".format(idx), \"model.{}.\".format(idx+7))\n",
+    "            model.state_dict()[k] -= model.state_dict()[k]\n",
+    "            model.state_dict()[k] += ckpt['model'].state_dict()[kr]\n",
+    "            print(k, \"perfectly matched!!\")\n",
+    "        elif \"model.{}.cv2.\".format(idx) in k:\n",
+    "            kr = k.replace(\"model.{}.cv2.\".format(idx), \"model.{}.cv4.\".format(idx+7))\n",
+    "            model.state_dict()[k] -= model.state_dict()[k]\n",
+    "            model.state_dict()[k] += ckpt['model'].state_dict()[kr]\n",
+    "            print(k, \"perfectly matched!!\")\n",
+    "        elif \"model.{}.cv3.\".format(idx) in k:\n",
+    "            kr = k.replace(\"model.{}.cv3.\".format(idx), \"model.{}.cv5.\".format(idx+7))\n",
+    "            model.state_dict()[k] -= model.state_dict()[k]\n",
+    "            model.state_dict()[k] += ckpt['model'].state_dict()[kr]\n",
+    "            print(k, \"perfectly matched!!\")\n",
+    "        elif \"model.{}.dfl.\".format(idx) in k:\n",
+    "            kr = k.replace(\"model.{}.dfl.\".format(idx), \"model.{}.dfl2.\".format(idx+7))\n",
+    "            model.state_dict()[k] -= model.state_dict()[k]\n",
+    "            model.state_dict()[k] += ckpt['model'].state_dict()[kr]\n",
+    "            print(k, \"perfectly matched!!\")\n",
+    "_ = model.eval()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "27bc1869",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "m_ckpt = {'model': model.half(),\n",
+    "          'optimizer': None,\n",
+    "          'best_fitness': None,\n",
+    "          'ema': None,\n",
+    "          'updates': None,\n",
+    "          'opt': None,\n",
+    "          'git': None,\n",
+    "          'date': None,\n",
+    "          'epoch': -1}\n",
+    "torch.save(m_ckpt, \"./yolov9-e-converted.pt\")"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3 (ipykernel)",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.8.12"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

train.py

@@ -0,0 +1,634 @@
+import argparse
+import math
+import os
+import random
+import sys
+import time
+from copy import deepcopy
+from datetime import datetime
+from pathlib import Path
+
+import numpy as np
+import torch
+import torch.distributed as dist
+import torch.nn as nn
+import yaml
+from torch.optim import lr_scheduler
+from tqdm import tqdm
+
+FILE = Path(__file__).resolve()
+ROOT = FILE.parents[0]  # root directory
+if str(ROOT) not in sys.path:
+    sys.path.append(str(ROOT))  # add ROOT to PATH
+ROOT = Path(os.path.relpath(ROOT, Path.cwd()))  # relative
+
+import val as validate  # for end-of-epoch mAP
+from models.experimental import attempt_load
+from models.yolo import Model
+from utils.autoanchor import check_anchors
+from utils.autobatch import check_train_batch_size
+from utils.callbacks import Callbacks
+from utils.dataloaders import create_dataloader
+from utils.downloads import attempt_download, is_url
+from utils.general import (LOGGER, TQDM_BAR_FORMAT, check_amp, check_dataset, check_file, check_img_size,
+                           check_suffix, check_yaml, colorstr, get_latest_run, increment_path, init_seeds,
+                           intersect_dicts, labels_to_class_weights, labels_to_image_weights, methods,
+                           one_cycle, one_flat_cycle, print_args, print_mutation, strip_optimizer, yaml_save)
+from utils.loggers import Loggers
+from utils.loggers.comet.comet_utils import check_comet_resume
+from utils.loss_tal import ComputeLoss
+from utils.metrics import fitness
+from utils.plots import plot_evolve
+from utils.torch_utils import (EarlyStopping, ModelEMA, de_parallel, select_device, smart_DDP,
+                               smart_optimizer, smart_resume, torch_distributed_zero_first)
+
+LOCAL_RANK = int(os.getenv('LOCAL_RANK', -1))  # https://pytorch.org/docs/stable/elastic/run.html
+RANK = int(os.getenv('RANK', -1))
+WORLD_SIZE = int(os.getenv('WORLD_SIZE', 1))
+GIT_INFO = None
+
+
+def train(hyp, opt, device, callbacks):  # hyp is path/to/hyp.yaml or hyp dictionary
+    save_dir, epochs, batch_size, weights, single_cls, evolve, data, cfg, resume, noval, nosave, workers, freeze = \
+        Path(opt.save_dir), opt.epochs, opt.batch_size, opt.weights, opt.single_cls, opt.evolve, opt.data, opt.cfg, \
+        opt.resume, opt.noval, opt.nosave, opt.workers, opt.freeze
+    callbacks.run('on_pretrain_routine_start')
+
+    # Directories
+    w = save_dir / 'weights'  # weights dir
+    (w.parent if evolve else w).mkdir(parents=True, exist_ok=True)  # make dir
+    last, best = w / 'last.pt', w / 'best.pt'
+    last_striped, best_striped = w / 'last_striped.pt', w / 'best_striped.pt'
+
+    # Hyperparameters
+    if isinstance(hyp, str):
+        with open(hyp, errors='ignore') as f:
+            hyp = yaml.safe_load(f)  # load hyps dict
+    LOGGER.info(colorstr('hyperparameters: ') + ', '.join(f'{k}={v}' for k, v in hyp.items()))
+    hyp['anchor_t'] = 5.0
+    opt.hyp = hyp.copy()  # for saving hyps to checkpoints
+
+    # Save run settings
+    if not evolve:
+        yaml_save(save_dir / 'hyp.yaml', hyp)
+        yaml_save(save_dir / 'opt.yaml', vars(opt))
+
+    # Loggers
+    data_dict = None
+    if RANK in {-1, 0}:
+        loggers = Loggers(save_dir, weights, opt, hyp, LOGGER)  # loggers instance
+
+        # Register actions
+        for k in methods(loggers):
+            callbacks.register_action(k, callback=getattr(loggers, k))
+
+        # Process custom dataset artifact link
+        data_dict = loggers.remote_dataset
+        if resume:  # If resuming runs from remote artifact
+            weights, epochs, hyp, batch_size = opt.weights, opt.epochs, opt.hyp, opt.batch_size
+
+    # Config
+    plots = not evolve and not opt.noplots  # create plots
+    cuda = device.type != 'cpu'
+    init_seeds(opt.seed + 1 + RANK, deterministic=True)
+    with torch_distributed_zero_first(LOCAL_RANK):
+        data_dict = data_dict or check_dataset(data)  # check if None
+    train_path, val_path = data_dict['train'], data_dict['val']
+    nc = 1 if single_cls else int(data_dict['nc'])  # number of classes
+    names = {0: 'item'} if single_cls and len(data_dict['names']) != 1 else data_dict['names']  # class names
+    #is_coco = isinstance(val_path, str) and val_path.endswith('coco/val2017.txt')  # COCO dataset
+    is_coco = isinstance(val_path, str) and val_path.endswith('val2017.txt')  # COCO dataset
+
+    # Model
+    check_suffix(weights, '.pt')  # check weights
+    pretrained = weights.endswith('.pt')
+    if pretrained:
+        with torch_distributed_zero_first(LOCAL_RANK):
+            weights = attempt_download(weights)  # download if not found locally
+        ckpt = torch.load(weights, map_location='cpu')  # load checkpoint to CPU to avoid CUDA memory leak
+        model = Model(cfg or ckpt['model'].yaml, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device)  # create
+        exclude = ['anchor'] if (cfg or hyp.get('anchors')) and not resume else []  # exclude keys
+        csd = ckpt['model'].float().state_dict()  # checkpoint state_dict as FP32
+        csd = intersect_dicts(csd, model.state_dict(), exclude=exclude)  # intersect
+        model.load_state_dict(csd, strict=False)  # load
+        LOGGER.info(f'Transferred {len(csd)}/{len(model.state_dict())} items from {weights}')  # report
+    else:
+        model = Model(cfg, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device)  # create
+    amp = check_amp(model)  # check AMP
+
+    # Freeze
+    freeze = [f'model.{x}.' for x in (freeze if len(freeze) > 1 else range(freeze[0]))]  # layers to freeze
+    for k, v in model.named_parameters():
+        # v.requires_grad = True  # train all layers TODO: uncomment this line as in master
+        # v.register_hook(lambda x: torch.nan_to_num(x))  # NaN to 0 (commented for erratic training results)
+        if any(x in k for x in freeze):
+            LOGGER.info(f'freezing {k}')
+            v.requires_grad = False
+
+    # Image size
+    gs = max(int(model.stride.max()), 32)  # grid size (max stride)
+    imgsz = check_img_size(opt.imgsz, gs, floor=gs * 2)  # verify imgsz is gs-multiple
+
+    # Batch size
+    if RANK == -1 and batch_size == -1:  # single-GPU only, estimate best batch size
+        batch_size = check_train_batch_size(model, imgsz, amp)
+        loggers.on_params_update({"batch_size": batch_size})
+
+    # Optimizer
+    nbs = 64  # nominal batch size
+    accumulate = max(round(nbs / batch_size), 1)  # accumulate loss before optimizing
+    hyp['weight_decay'] *= batch_size * accumulate / nbs  # scale weight_decay
+    optimizer = smart_optimizer(model, opt.optimizer, hyp['lr0'], hyp['momentum'], hyp['weight_decay'])
+
+    # Scheduler
+    if opt.cos_lr:
+        lf = one_cycle(1, hyp['lrf'], epochs)  # cosine 1->hyp['lrf']
+    elif opt.flat_cos_lr:
+        lf = one_flat_cycle(1, hyp['lrf'], epochs)  # flat cosine 1->hyp['lrf']        
+    elif opt.fixed_lr:
+        lf = lambda x: 1.0
+    else:
+        lf = lambda x: (1 - x / epochs) * (1.0 - hyp['lrf']) + hyp['lrf']  # linear
+
+    scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
+    # from utils.plots import plot_lr_scheduler; plot_lr_scheduler(optimizer, scheduler, epochs)
+
+    # EMA
+    ema = ModelEMA(model) if RANK in {-1, 0} else None
+
+    # Resume
+    best_fitness, start_epoch = 0.0, 0
+    if pretrained:
+        if resume:
+            best_fitness, start_epoch, epochs = smart_resume(ckpt, optimizer, ema, weights, epochs, resume)
+        del ckpt, csd
+
+    # DP mode
+    if cuda and RANK == -1 and torch.cuda.device_count() > 1:
+        LOGGER.warning('WARNING ⚠️ DP not recommended, use torch.distributed.run for best DDP Multi-GPU results.')
+        model = torch.nn.DataParallel(model)
+
+    # SyncBatchNorm
+    if opt.sync_bn and cuda and RANK != -1:
+        model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device)
+        LOGGER.info('Using SyncBatchNorm()')
+
+    # Trainloader
+    train_loader, dataset = create_dataloader(train_path,
+                                              imgsz,
+                                              batch_size // WORLD_SIZE,
+                                              gs,
+                                              single_cls,
+                                              hyp=hyp,
+                                              augment=True,
+                                              cache=None if opt.cache == 'val' else opt.cache,
+                                              rect=opt.rect,
+                                              rank=LOCAL_RANK,
+                                              workers=workers,
+                                              image_weights=opt.image_weights,
+                                              close_mosaic=opt.close_mosaic != 0,
+                                              quad=opt.quad,
+                                              prefix=colorstr('train: '),
+                                              shuffle=True,
+                                              min_items=opt.min_items)
+    labels = np.concatenate(dataset.labels, 0)
+    mlc = int(labels[:, 0].max())  # max label class
+    assert mlc < nc, f'Label class {mlc} exceeds nc={nc} in {data}. Possible class labels are 0-{nc - 1}'
+
+    # Process 0
+    if RANK in {-1, 0}:
+        val_loader = create_dataloader(val_path,
+                                       imgsz,
+                                       batch_size // WORLD_SIZE * 2,
+                                       gs,
+                                       single_cls,
+                                       hyp=hyp,
+                                       cache=None if noval else opt.cache,
+                                       rect=True,
+                                       rank=-1,
+                                       workers=workers * 2,
+                                       pad=0.5,
+                                       prefix=colorstr('val: '))[0]
+
+        if not resume:
+            # if not opt.noautoanchor:
+            #     check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz)  # run AutoAnchor
+            model.half().float()  # pre-reduce anchor precision
+
+        callbacks.run('on_pretrain_routine_end', labels, names)
+
+    # DDP mode
+    if cuda and RANK != -1:
+        model = smart_DDP(model)
+
+    # Model attributes
+    nl = de_parallel(model).model[-1].nl  # number of detection layers (to scale hyps)
+    #hyp['box'] *= 3 / nl  # scale to layers
+    #hyp['cls'] *= nc / 80 * 3 / nl  # scale to classes and layers
+    #hyp['obj'] *= (imgsz / 640) ** 2 * 3 / nl  # scale to image size and layers
+    hyp['label_smoothing'] = opt.label_smoothing
+    model.nc = nc  # attach number of classes to model
+    model.hyp = hyp  # attach hyperparameters to model
+    model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) * nc  # attach class weights
+    model.names = names
+
+    # Start training
+    t0 = time.time()
+    nb = len(train_loader)  # number of batches
+    nw = max(round(hyp['warmup_epochs'] * nb), 100)  # number of warmup iterations, max(3 epochs, 100 iterations)
+    # nw = min(nw, (epochs - start_epoch) / 2 * nb)  # limit warmup to < 1/2 of training
+    last_opt_step = -1
+    maps = np.zeros(nc)  # mAP per class
+    results = (0, 0, 0, 0, 0, 0, 0)  # P, R, [email protected], [email protected], val_loss(box, obj, cls)
+    scheduler.last_epoch = start_epoch - 1  # do not move
+    scaler = torch.cuda.amp.GradScaler(enabled=amp)
+    stopper, stop = EarlyStopping(patience=opt.patience), False
+    compute_loss = ComputeLoss(model)  # init loss class
+    callbacks.run('on_train_start')
+    LOGGER.info(f'Image sizes {imgsz} train, {imgsz} val\n'
+                f'Using {train_loader.num_workers * WORLD_SIZE} dataloader workers\n'
+                f"Logging results to {colorstr('bold', save_dir)}\n"
+                f'Starting training for {epochs} epochs...')
+    for epoch in range(start_epoch, epochs):  # epoch ------------------------------------------------------------------
+        callbacks.run('on_train_epoch_start')
+        model.train()
+
+        # Update image weights (optional, single-GPU only)
+        if opt.image_weights:
+            cw = model.class_weights.cpu().numpy() * (1 - maps) ** 2 / nc  # class weights
+            iw = labels_to_image_weights(dataset.labels, nc=nc, class_weights=cw)  # image weights
+            dataset.indices = random.choices(range(dataset.n), weights=iw, k=dataset.n)  # rand weighted idx
+        if epoch == (epochs - opt.close_mosaic):
+            LOGGER.info("Closing dataloader mosaic")
+            dataset.mosaic = False
+
+        # Update mosaic border (optional)
+        # b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs)
+        # dataset.mosaic_border = [b - imgsz, -b]  # height, width borders
+
+        mloss = torch.zeros(3, device=device)  # mean losses
+        if RANK != -1:
+            train_loader.sampler.set_epoch(epoch)
+        pbar = enumerate(train_loader)
+        LOGGER.info(('\n' + '%11s' * 7) % ('Epoch', 'GPU_mem', 'box_loss', 'cls_loss', 'dfl_loss', 'Instances', 'Size'))
+        if RANK in {-1, 0}:
+            pbar = tqdm(pbar, total=nb, bar_format=TQDM_BAR_FORMAT)  # progress bar
+        optimizer.zero_grad()
+        for i, (imgs, targets, paths, _) in pbar:  # batch -------------------------------------------------------------
+            callbacks.run('on_train_batch_start')
+            ni = i + nb * epoch  # number integrated batches (since train start)
+            imgs = imgs.to(device, non_blocking=True).float() / 255  # uint8 to float32, 0-255 to 0.0-1.0
+
+            # Warmup
+            if ni <= nw:
+                xi = [0, nw]  # x interp
+                # compute_loss.gr = np.interp(ni, xi, [0.0, 1.0])  # iou loss ratio (obj_loss = 1.0 or iou)
+                accumulate = max(1, np.interp(ni, xi, [1, nbs / batch_size]).round())
+                for j, x in enumerate(optimizer.param_groups):
+                    # bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
+                    x['lr'] = np.interp(ni, xi, [hyp['warmup_bias_lr'] if j == 0 else 0.0, x['initial_lr'] * lf(epoch)])
+                    if 'momentum' in x:
+                        x['momentum'] = np.interp(ni, xi, [hyp['warmup_momentum'], hyp['momentum']])
+
+            # Multi-scale
+            if opt.multi_scale:
+                sz = random.randrange(imgsz * 0.5, imgsz * 1.5 + gs) // gs * gs  # size
+                sf = sz / max(imgs.shape[2:])  # scale factor
+                if sf != 1:
+                    ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]]  # new shape (stretched to gs-multiple)
+                    imgs = nn.functional.interpolate(imgs, size=ns, mode='bilinear', align_corners=False)
+
+            # Forward
+            with torch.cuda.amp.autocast(amp):
+                pred = model(imgs)  # forward
+                loss, loss_items = compute_loss(pred, targets.to(device))  # loss scaled by batch_size
+                if RANK != -1:
+                    loss *= WORLD_SIZE  # gradient averaged between devices in DDP mode
+                if opt.quad:
+                    loss *= 4.
+
+            # Backward
+            scaler.scale(loss).backward()
+
+            # Optimize - https://pytorch.org/docs/master/notes/amp_examples.html
+            if ni - last_opt_step >= accumulate:
+                scaler.unscale_(optimizer)  # unscale gradients
+                torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=10.0)  # clip gradients
+                scaler.step(optimizer)  # optimizer.step
+                scaler.update()
+                optimizer.zero_grad()
+                if ema:
+                    ema.update(model)
+                last_opt_step = ni
+
+            # Log
+            if RANK in {-1, 0}:
+                mloss = (mloss * i + loss_items) / (i + 1)  # update mean losses
+                mem = f'{torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0:.3g}G'  # (GB)
+                pbar.set_description(('%11s' * 2 + '%11.4g' * 5) %
+                                     (f'{epoch}/{epochs - 1}', mem, *mloss, targets.shape[0], imgs.shape[-1]))
+                callbacks.run('on_train_batch_end', model, ni, imgs, targets, paths, list(mloss))
+                if callbacks.stop_training:
+                    return
+            # end batch ------------------------------------------------------------------------------------------------
+
+        # Scheduler
+        lr = [x['lr'] for x in optimizer.param_groups]  # for loggers
+        scheduler.step()
+
+        if RANK in {-1, 0}:
+            # mAP
+            callbacks.run('on_train_epoch_end', epoch=epoch)
+            ema.update_attr(model, include=['yaml', 'nc', 'hyp', 'names', 'stride', 'class_weights'])
+            final_epoch = (epoch + 1 == epochs) or stopper.possible_stop
+            if not noval or final_epoch:  # Calculate mAP
+                results, maps, _ = validate.run(data_dict,
+                                                batch_size=batch_size // WORLD_SIZE * 2,
+                                                imgsz=imgsz,
+                                                half=amp,
+                                                model=ema.ema,
+                                                single_cls=single_cls,
+                                                dataloader=val_loader,
+                                                save_dir=save_dir,
+                                                plots=False,
+                                                callbacks=callbacks,
+                                                compute_loss=compute_loss)
+
+            # Update best mAP
+            fi = fitness(np.array(results).reshape(1, -1))  # weighted combination of [P, R, [email protected], [email protected]]
+            stop = stopper(epoch=epoch, fitness=fi)  # early stop check
+            if fi > best_fitness:
+                best_fitness = fi
+            log_vals = list(mloss) + list(results) + lr
+            callbacks.run('on_fit_epoch_end', log_vals, epoch, best_fitness, fi)
+
+            # Save model
+            if (not nosave) or (final_epoch and not evolve):  # if save
+                ckpt = {
+                    'epoch': epoch,
+                    'best_fitness': best_fitness,
+                    'model': deepcopy(de_parallel(model)).half(),
+                    'ema': deepcopy(ema.ema).half(),
+                    'updates': ema.updates,
+                    'optimizer': optimizer.state_dict(),
+                    'opt': vars(opt),
+                    'git': GIT_INFO,  # {remote, branch, commit} if a git repo
+                    'date': datetime.now().isoformat()}
+
+                # Save last, best and delete
+                torch.save(ckpt, last)
+                if best_fitness == fi:
+                    torch.save(ckpt, best)
+                if opt.save_period > 0 and epoch % opt.save_period == 0:
+                    torch.save(ckpt, w / f'epoch{epoch}.pt')
+                del ckpt
+                callbacks.run('on_model_save', last, epoch, final_epoch, best_fitness, fi)
+
+        # EarlyStopping
+        if RANK != -1:  # if DDP training
+            broadcast_list = [stop if RANK == 0 else None]
+            dist.broadcast_object_list(broadcast_list, 0)  # broadcast 'stop' to all ranks
+            if RANK != 0:
+                stop = broadcast_list[0]
+        if stop:
+            break  # must break all DDP ranks
+
+        # end epoch ----------------------------------------------------------------------------------------------------
+    # end training -----------------------------------------------------------------------------------------------------
+    if RANK in {-1, 0}:
+        LOGGER.info(f'\n{epoch - start_epoch + 1} epochs completed in {(time.time() - t0) / 3600:.3f} hours.')
+        for f in last, best:
+            if f.exists():
+                if f is last:
+                    strip_optimizer(f, last_striped)  # strip optimizers
+                else:
+                    strip_optimizer(f, best_striped)  # strip optimizers
+                if f is best:
+                    LOGGER.info(f'\nValidating {f}...')
+                    results, _, _ = validate.run(
+                        data_dict,
+                        batch_size=batch_size // WORLD_SIZE * 2,
+                        imgsz=imgsz,
+                        model=attempt_load(f, device).half(),
+                        single_cls=single_cls,
+                        dataloader=val_loader,
+                        save_dir=save_dir,
+                        save_json=is_coco,
+                        verbose=True,
+                        plots=plots,
+                        callbacks=callbacks,
+                        compute_loss=compute_loss)  # val best model with plots
+                    if is_coco:
+                        callbacks.run('on_fit_epoch_end', list(mloss) + list(results) + lr, epoch, best_fitness, fi)
+
+        callbacks.run('on_train_end', last, best, epoch, results)
+
+    torch.cuda.empty_cache()
+    return results
+
+
+def parse_opt(known=False):
+    parser = argparse.ArgumentParser()
+    # parser.add_argument('--weights', type=str, default=ROOT / 'yolo.pt', help='initial weights path')
+    # parser.add_argument('--cfg', type=str, default='', help='model.yaml path')
+    parser.add_argument('--weights', type=str, default='', help='initial weights path')
+    parser.add_argument('--cfg', type=str, default='yolo.yaml', help='model.yaml path')
+    parser.add_argument('--data', type=str, default=ROOT / 'data/coco128.yaml', help='dataset.yaml path')
+    parser.add_argument('--hyp', type=str, default=ROOT / 'data/hyps/hyp.scratch-low.yaml', help='hyperparameters path')
+    parser.add_argument('--epochs', type=int, default=100, help='total training epochs')
+    parser.add_argument('--batch-size', type=int, default=16, help='total batch size for all GPUs, -1 for autobatch')
+    parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=640, help='train, val image size (pixels)')
+    parser.add_argument('--rect', action='store_true', help='rectangular training')
+    parser.add_argument('--resume', nargs='?', const=True, default=False, help='resume most recent training')
+    parser.add_argument('--nosave', action='store_true', help='only save final checkpoint')
+    parser.add_argument('--noval', action='store_true', help='only validate final epoch')
+    parser.add_argument('--noautoanchor', action='store_true', help='disable AutoAnchor')
+    parser.add_argument('--noplots', action='store_true', help='save no plot files')
+    parser.add_argument('--evolve', type=int, nargs='?', const=300, help='evolve hyperparameters for x generations')
+    parser.add_argument('--bucket', type=str, default='', help='gsutil bucket')
+    parser.add_argument('--cache', type=str, nargs='?', const='ram', help='image --cache ram/disk')
+    parser.add_argument('--image-weights', action='store_true', help='use weighted image selection for training')
+    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
+    parser.add_argument('--multi-scale', action='store_true', help='vary img-size +/- 50%%')
+    parser.add_argument('--single-cls', action='store_true', help='train multi-class data as single-class')
+    parser.add_argument('--optimizer', type=str, choices=['SGD', 'Adam', 'AdamW', 'LION'], default='SGD', help='optimizer')
+    parser.add_argument('--sync-bn', action='store_true', help='use SyncBatchNorm, only available in DDP mode')
+    parser.add_argument('--workers', type=int, default=8, help='max dataloader workers (per RANK in DDP mode)')
+    parser.add_argument('--project', default=ROOT / 'runs/train', help='save to project/name')
+    parser.add_argument('--name', default='exp', help='save to project/name')
+    parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')
+    parser.add_argument('--quad', action='store_true', help='quad dataloader')
+    parser.add_argument('--cos-lr', action='store_true', help='cosine LR scheduler')
+    parser.add_argument('--flat-cos-lr', action='store_true', help='flat cosine LR scheduler')
+    parser.add_argument('--fixed-lr', action='store_true', help='fixed LR scheduler')
+    parser.add_argument('--label-smoothing', type=float, default=0.0, help='Label smoothing epsilon')
+    parser.add_argument('--patience', type=int, default=100, help='EarlyStopping patience (epochs without improvement)')
+    parser.add_argument('--freeze', nargs='+', type=int, default=[0], help='Freeze layers: backbone=10, first3=0 1 2')
+    parser.add_argument('--save-period', type=int, default=-1, help='Save checkpoint every x epochs (disabled if < 1)')
+    parser.add_argument('--seed', type=int, default=0, help='Global training seed')
+    parser.add_argument('--local_rank', type=int, default=-1, help='Automatic DDP Multi-GPU argument, do not modify')
+    parser.add_argument('--min-items', type=int, default=0, help='Experimental')
+    parser.add_argument('--close-mosaic', type=int, default=0, help='Experimental')
+
+    # Logger arguments
+    parser.add_argument('--entity', default=None, help='Entity')
+    parser.add_argument('--upload_dataset', nargs='?', const=True, default=False, help='Upload data, "val" option')
+    parser.add_argument('--bbox_interval', type=int, default=-1, help='Set bounding-box image logging interval')
+    parser.add_argument('--artifact_alias', type=str, default='latest', help='Version of dataset artifact to use')
+
+    return parser.parse_known_args()[0] if known else parser.parse_args()
+
+
+def main(opt, callbacks=Callbacks()):
+    # Checks
+    if RANK in {-1, 0}:
+        print_args(vars(opt))
+
+    # Resume (from specified or most recent last.pt)
+    if opt.resume and not check_comet_resume(opt) and not opt.evolve:
+        last = Path(check_file(opt.resume) if isinstance(opt.resume, str) else get_latest_run())
+        opt_yaml = last.parent.parent / 'opt.yaml'  # train options yaml
+        opt_data = opt.data  # original dataset
+        if opt_yaml.is_file():
+            with open(opt_yaml, errors='ignore') as f:
+                d = yaml.safe_load(f)
+        else:
+            d = torch.load(last, map_location='cpu')['opt']
+        opt = argparse.Namespace(**d)  # replace
+        opt.cfg, opt.weights, opt.resume = '', str(last), True  # reinstate
+        if is_url(opt_data):
+            opt.data = check_file(opt_data)  # avoid HUB resume auth timeout
+    else:
+        opt.data, opt.cfg, opt.hyp, opt.weights, opt.project = \
+            check_file(opt.data), check_yaml(opt.cfg), check_yaml(opt.hyp), str(opt.weights), str(opt.project)  # checks
+        assert len(opt.cfg) or len(opt.weights), 'either --cfg or --weights must be specified'
+        if opt.evolve:
+            if opt.project == str(ROOT / 'runs/train'):  # if default project name, rename to runs/evolve
+                opt.project = str(ROOT / 'runs/evolve')
+            opt.exist_ok, opt.resume = opt.resume, False  # pass resume to exist_ok and disable resume
+        if opt.name == 'cfg':
+            opt.name = Path(opt.cfg).stem  # use model.yaml as name
+        opt.save_dir = str(increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok))
+
+    # DDP mode
+    device = select_device(opt.device, batch_size=opt.batch_size)
+    if LOCAL_RANK != -1:
+        msg = 'is not compatible with YOLO Multi-GPU DDP training'
+        assert not opt.image_weights, f'--image-weights {msg}'
+        assert not opt.evolve, f'--evolve {msg}'
+        assert opt.batch_size != -1, f'AutoBatch with --batch-size -1 {msg}, please pass a valid --batch-size'
+        assert opt.batch_size % WORLD_SIZE == 0, f'--batch-size {opt.batch_size} must be multiple of WORLD_SIZE'
+        assert torch.cuda.device_count() > LOCAL_RANK, 'insufficient CUDA devices for DDP command'
+        torch.cuda.set_device(LOCAL_RANK)
+        device = torch.device('cuda', LOCAL_RANK)
+        dist.init_process_group(backend="nccl" if dist.is_nccl_available() else "gloo")
+
+    # Train
+    if not opt.evolve:
+        train(opt.hyp, opt, device, callbacks)
+
+    # Evolve hyperparameters (optional)
+    else:
+        # Hyperparameter evolution metadata (mutation scale 0-1, lower_limit, upper_limit)
+        meta = {
+            'lr0': (1, 1e-5, 1e-1),  # initial learning rate (SGD=1E-2, Adam=1E-3)
+            'lrf': (1, 0.01, 1.0),  # final OneCycleLR learning rate (lr0 * lrf)
+            'momentum': (0.3, 0.6, 0.98),  # SGD momentum/Adam beta1
+            'weight_decay': (1, 0.0, 0.001),  # optimizer weight decay
+            'warmup_epochs': (1, 0.0, 5.0),  # warmup epochs (fractions ok)
+            'warmup_momentum': (1, 0.0, 0.95),  # warmup initial momentum
+            'warmup_bias_lr': (1, 0.0, 0.2),  # warmup initial bias lr
+            'box': (1, 0.02, 0.2),  # box loss gain
+            'cls': (1, 0.2, 4.0),  # cls loss gain
+            'cls_pw': (1, 0.5, 2.0),  # cls BCELoss positive_weight
+            'obj': (1, 0.2, 4.0),  # obj loss gain (scale with pixels)
+            'obj_pw': (1, 0.5, 2.0),  # obj BCELoss positive_weight
+            'iou_t': (0, 0.1, 0.7),  # IoU training threshold
+            'anchor_t': (1, 2.0, 8.0),  # anchor-multiple threshold
+            'anchors': (2, 2.0, 10.0),  # anchors per output grid (0 to ignore)
+            'fl_gamma': (0, 0.0, 2.0),  # focal loss gamma (efficientDet default gamma=1.5)
+            'hsv_h': (1, 0.0, 0.1),  # image HSV-Hue augmentation (fraction)
+            'hsv_s': (1, 0.0, 0.9),  # image HSV-Saturation augmentation (fraction)
+            'hsv_v': (1, 0.0, 0.9),  # image HSV-Value augmentation (fraction)
+            'degrees': (1, 0.0, 45.0),  # image rotation (+/- deg)
+            'translate': (1, 0.0, 0.9),  # image translation (+/- fraction)
+            'scale': (1, 0.0, 0.9),  # image scale (+/- gain)
+            'shear': (1, 0.0, 10.0),  # image shear (+/- deg)
+            'perspective': (0, 0.0, 0.001),  # image perspective (+/- fraction), range 0-0.001
+            'flipud': (1, 0.0, 1.0),  # image flip up-down (probability)
+            'fliplr': (0, 0.0, 1.0),  # image flip left-right (probability)
+            'mosaic': (1, 0.0, 1.0),  # image mixup (probability)
+            'mixup': (1, 0.0, 1.0),  # image mixup (probability)
+            'copy_paste': (1, 0.0, 1.0)}  # segment copy-paste (probability)
+
+        with open(opt.hyp, errors='ignore') as f:
+            hyp = yaml.safe_load(f)  # load hyps dict
+            if 'anchors' not in hyp:  # anchors commented in hyp.yaml
+                hyp['anchors'] = 3
+        if opt.noautoanchor:
+            del hyp['anchors'], meta['anchors']
+        opt.noval, opt.nosave, save_dir = True, True, Path(opt.save_dir)  # only val/save final epoch
+        # ei = [isinstance(x, (int, float)) for x in hyp.values()]  # evolvable indices
+        evolve_yaml, evolve_csv = save_dir / 'hyp_evolve.yaml', save_dir / 'evolve.csv'
+        if opt.bucket:
+            os.system(f'gsutil cp gs://{opt.bucket}/evolve.csv {evolve_csv}')  # download evolve.csv if exists
+
+        for _ in range(opt.evolve):  # generations to evolve
+            if evolve_csv.exists():  # if evolve.csv exists: select best hyps and mutate
+                # Select parent(s)
+                parent = 'single'  # parent selection method: 'single' or 'weighted'
+                x = np.loadtxt(evolve_csv, ndmin=2, delimiter=',', skiprows=1)
+                n = min(5, len(x))  # number of previous results to consider
+                x = x[np.argsort(-fitness(x))][:n]  # top n mutations
+                w = fitness(x) - fitness(x).min() + 1E-6  # weights (sum > 0)
+                if parent == 'single' or len(x) == 1:
+                    # x = x[random.randint(0, n - 1)]  # random selection
+                    x = x[random.choices(range(n), weights=w)[0]]  # weighted selection
+                elif parent == 'weighted':
+                    x = (x * w.reshape(n, 1)).sum(0) / w.sum()  # weighted combination
+
+                # Mutate
+                mp, s = 0.8, 0.2  # mutation probability, sigma
+                npr = np.random
+                npr.seed(int(time.time()))
+                g = np.array([meta[k][0] for k in hyp.keys()])  # gains 0-1
+                ng = len(meta)
+                v = np.ones(ng)
+                while all(v == 1):  # mutate until a change occurs (prevent duplicates)
+                    v = (g * (npr.random(ng) < mp) * npr.randn(ng) * npr.random() * s + 1).clip(0.3, 3.0)
+                for i, k in enumerate(hyp.keys()):  # plt.hist(v.ravel(), 300)
+                    hyp[k] = float(x[i + 7] * v[i])  # mutate
+
+            # Constrain to limits
+            for k, v in meta.items():
+                hyp[k] = max(hyp[k], v[1])  # lower limit
+                hyp[k] = min(hyp[k], v[2])  # upper limit
+                hyp[k] = round(hyp[k], 5)  # significant digits
+
+            # Train mutation
+            results = train(hyp.copy(), opt, device, callbacks)
+            callbacks = Callbacks()
+            # Write mutation results
+            keys = ('metrics/precision', 'metrics/recall', 'metrics/mAP_0.5', 'metrics/mAP_0.5:0.95', 'val/box_loss',
+                    'val/obj_loss', 'val/cls_loss')
+            print_mutation(keys, results, hyp.copy(), save_dir, opt.bucket)
+
+        # Plot results
+        plot_evolve(evolve_csv)
+        LOGGER.info(f'Hyperparameter evolution finished {opt.evolve} generations\n'
+                    f"Results saved to {colorstr('bold', save_dir)}\n"
+                    f'Usage example: $ python train.py --hyp {evolve_yaml}')
+
+
+def run(**kwargs):
+    # Usage: import train; train.run(data='coco128.yaml', imgsz=320, weights='yolo.pt')
+    opt = parse_opt(True)
+    for k, v in kwargs.items():
+        setattr(opt, k, v)
+    main(opt)
+    return opt
+
+
+if __name__ == "__main__":
+    opt = parse_opt()
+    main(opt)

train_dual.py

@@ -0,0 +1,644 @@
+import argparse
+import math
+import os
+import random
+import sys
+import time
+from copy import deepcopy
+from datetime import datetime
+from pathlib import Path
+
+import numpy as np
+import torch
+import torch.distributed as dist
+import torch.nn as nn
+import yaml
+from torch.optim import lr_scheduler
+from tqdm import tqdm
+
+FILE = Path(__file__).resolve()
+ROOT = FILE.parents[0]  # YOLO root directory
+if str(ROOT) not in sys.path:
+    sys.path.append(str(ROOT))  # add ROOT to PATH
+ROOT = Path(os.path.relpath(ROOT, Path.cwd()))  # relative
+
+import val_dual as validate  # for end-of-epoch mAP
+from models.experimental import attempt_load
+from models.yolo import Model
+from utils.autoanchor import check_anchors
+from utils.autobatch import check_train_batch_size
+from utils.callbacks import Callbacks
+from utils.dataloaders import create_dataloader
+from utils.downloads import attempt_download, is_url
+from utils.general import (LOGGER, TQDM_BAR_FORMAT, check_amp, check_dataset, check_file, check_git_info,
+                           check_git_status, check_img_size, check_requirements, check_suffix, check_yaml, colorstr,
+                           get_latest_run, increment_path, init_seeds, intersect_dicts, labels_to_class_weights,
+                           labels_to_image_weights, methods, one_cycle, print_args, print_mutation, strip_optimizer,
+                           yaml_save, one_flat_cycle)
+from utils.loggers import Loggers
+from utils.loggers.comet.comet_utils import check_comet_resume
+from utils.loss_tal_dual import ComputeLoss
+#from utils.loss_tal_dual import ComputeLossLH as ComputeLoss
+#from utils.loss_tal_dual import ComputeLossLHCF as ComputeLoss
+from utils.metrics import fitness
+from utils.plots import plot_evolve
+from utils.torch_utils import (EarlyStopping, ModelEMA, de_parallel, select_device, smart_DDP, smart_optimizer,
+                               smart_resume, torch_distributed_zero_first)
+
+LOCAL_RANK = int(os.getenv('LOCAL_RANK', -1))  # https://pytorch.org/docs/stable/elastic/run.html
+RANK = int(os.getenv('RANK', -1))
+WORLD_SIZE = int(os.getenv('WORLD_SIZE', 1))
+GIT_INFO = None#check_git_info()
+
+
+def train(hyp, opt, device, callbacks):  # hyp is path/to/hyp.yaml or hyp dictionary
+    save_dir, epochs, batch_size, weights, single_cls, evolve, data, cfg, resume, noval, nosave, workers, freeze = \
+        Path(opt.save_dir), opt.epochs, opt.batch_size, opt.weights, opt.single_cls, opt.evolve, opt.data, opt.cfg, \
+        opt.resume, opt.noval, opt.nosave, opt.workers, opt.freeze
+    callbacks.run('on_pretrain_routine_start')
+
+    # Directories
+    w = save_dir / 'weights'  # weights dir
+    (w.parent if evolve else w).mkdir(parents=True, exist_ok=True)  # make dir
+    last, best = w / 'last.pt', w / 'best.pt'
+
+    # Hyperparameters
+    if isinstance(hyp, str):
+        with open(hyp, errors='ignore') as f:
+            hyp = yaml.safe_load(f)  # load hyps dict
+    LOGGER.info(colorstr('hyperparameters: ') + ', '.join(f'{k}={v}' for k, v in hyp.items()))
+    hyp['anchor_t'] = 5.0
+    opt.hyp = hyp.copy()  # for saving hyps to checkpoints
+
+    # Save run settings
+    if not evolve:
+        yaml_save(save_dir / 'hyp.yaml', hyp)
+        yaml_save(save_dir / 'opt.yaml', vars(opt))
+
+    # Loggers
+    data_dict = None
+    if RANK in {-1, 0}:
+        loggers = Loggers(save_dir, weights, opt, hyp, LOGGER)  # loggers instance
+
+        # Register actions
+        for k in methods(loggers):
+            callbacks.register_action(k, callback=getattr(loggers, k))
+
+        # Process custom dataset artifact link
+        data_dict = loggers.remote_dataset
+        if resume:  # If resuming runs from remote artifact
+            weights, epochs, hyp, batch_size = opt.weights, opt.epochs, opt.hyp, opt.batch_size
+
+    # Config
+    plots = not evolve and not opt.noplots  # create plots
+    cuda = device.type != 'cpu'
+    init_seeds(opt.seed + 1 + RANK, deterministic=True)
+    with torch_distributed_zero_first(LOCAL_RANK):
+        data_dict = data_dict or check_dataset(data)  # check if None
+    train_path, val_path = data_dict['train'], data_dict['val']
+    nc = 1 if single_cls else int(data_dict['nc'])  # number of classes
+    names = {0: 'item'} if single_cls and len(data_dict['names']) != 1 else data_dict['names']  # class names
+    #is_coco = isinstance(val_path, str) and val_path.endswith('coco/val2017.txt')  # COCO dataset
+    is_coco = isinstance(val_path, str) and val_path.endswith('val2017.txt')  # COCO dataset
+
+    # Model
+    check_suffix(weights, '.pt')  # check weights
+    pretrained = weights.endswith('.pt')
+    if pretrained:
+        with torch_distributed_zero_first(LOCAL_RANK):
+            weights = attempt_download(weights)  # download if not found locally
+        ckpt = torch.load(weights, map_location='cpu')  # load checkpoint to CPU to avoid CUDA memory leak
+        model = Model(cfg or ckpt['model'].yaml, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device)  # create
+        exclude = ['anchor'] if (cfg or hyp.get('anchors')) and not resume else []  # exclude keys
+        csd = ckpt['model'].float().state_dict()  # checkpoint state_dict as FP32
+        csd = intersect_dicts(csd, model.state_dict(), exclude=exclude)  # intersect
+        model.load_state_dict(csd, strict=False)  # load
+        LOGGER.info(f'Transferred {len(csd)}/{len(model.state_dict())} items from {weights}')  # report
+    else:
+        model = Model(cfg, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device)  # create
+    amp = check_amp(model)  # check AMP
+
+    # Freeze
+    freeze = [f'model.{x}.' for x in (freeze if len(freeze) > 1 else range(freeze[0]))]  # layers to freeze
+    for k, v in model.named_parameters():
+        # v.requires_grad = True  # train all layers TODO: uncomment this line as in master
+        # v.register_hook(lambda x: torch.nan_to_num(x))  # NaN to 0 (commented for erratic training results)
+        if any(x in k for x in freeze):
+            LOGGER.info(f'freezing {k}')
+            v.requires_grad = False
+
+    # Image size
+    gs = max(int(model.stride.max()), 32)  # grid size (max stride)
+    imgsz = check_img_size(opt.imgsz, gs, floor=gs * 2)  # verify imgsz is gs-multiple
+
+    # Batch size
+    if RANK == -1 and batch_size == -1:  # single-GPU only, estimate best batch size
+        batch_size = check_train_batch_size(model, imgsz, amp)
+        loggers.on_params_update({"batch_size": batch_size})
+
+    # Optimizer
+    nbs = 64  # nominal batch size
+    accumulate = max(round(nbs / batch_size), 1)  # accumulate loss before optimizing
+    hyp['weight_decay'] *= batch_size * accumulate / nbs  # scale weight_decay
+    optimizer = smart_optimizer(model, opt.optimizer, hyp['lr0'], hyp['momentum'], hyp['weight_decay'])
+
+    # Scheduler
+    if opt.cos_lr:
+        lf = one_cycle(1, hyp['lrf'], epochs)  # cosine 1->hyp['lrf']
+    elif opt.flat_cos_lr:
+        lf = one_flat_cycle(1, hyp['lrf'], epochs)  # flat cosine 1->hyp['lrf']        
+    elif opt.fixed_lr:
+        lf = lambda x: 1.0
+    else:
+        lf = lambda x: (1 - x / epochs) * (1.0 - hyp['lrf']) + hyp['lrf']  # linear
+
+    # def lf(x):  # saw
+    #     return (1 - (x % 30) / 30) * (1 - x / epochs) * (1.0 - hyp['lrf']) + hyp['lrf']
+    #
+    # def lf(x):  # triangle start at min
+    #     return 2 * abs(x / 30 - math.floor(x / 30 + 1 / 2)) * (1 - x / epochs) * (1.0 - hyp['lrf']) + hyp['lrf']
+    #
+    # def lf(x):  # triangle start at max
+    #     return 2 * abs(x / 32 + .5 - math.floor(x / 32 + 1)) * (1 - x / epochs) * (1.0 - hyp['lrf']) + hyp['lrf']
+
+    scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
+    # from utils.plots import plot_lr_scheduler; plot_lr_scheduler(optimizer, scheduler, epochs)
+
+    # EMA
+    ema = ModelEMA(model) if RANK in {-1, 0} else None
+
+    # Resume
+    best_fitness, start_epoch = 0.0, 0
+    if pretrained:
+        if resume:
+            best_fitness, start_epoch, epochs = smart_resume(ckpt, optimizer, ema, weights, epochs, resume)
+        del ckpt, csd
+
+    # DP mode
+    if cuda and RANK == -1 and torch.cuda.device_count() > 1:
+        LOGGER.warning('WARNING ⚠️ DP not recommended, use torch.distributed.run for best DDP Multi-GPU results.')
+        model = torch.nn.DataParallel(model)
+
+    # SyncBatchNorm
+    if opt.sync_bn and cuda and RANK != -1:
+        model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device)
+        LOGGER.info('Using SyncBatchNorm()')
+
+    # Trainloader
+    train_loader, dataset = create_dataloader(train_path,
+                                              imgsz,
+                                              batch_size // WORLD_SIZE,
+                                              gs,
+                                              single_cls,
+                                              hyp=hyp,
+                                              augment=True,
+                                              cache=None if opt.cache == 'val' else opt.cache,
+                                              rect=opt.rect,
+                                              rank=LOCAL_RANK,
+                                              workers=workers,
+                                              image_weights=opt.image_weights,
+                                              close_mosaic=opt.close_mosaic != 0,
+                                              quad=opt.quad,
+                                              prefix=colorstr('train: '),
+                                              shuffle=True,
+                                              min_items=opt.min_items)
+    labels = np.concatenate(dataset.labels, 0)
+    mlc = int(labels[:, 0].max())  # max label class
+    assert mlc < nc, f'Label class {mlc} exceeds nc={nc} in {data}. Possible class labels are 0-{nc - 1}'
+
+    # Process 0
+    if RANK in {-1, 0}:
+        val_loader = create_dataloader(val_path,
+                                       imgsz,
+                                       batch_size // WORLD_SIZE * 2,
+                                       gs,
+                                       single_cls,
+                                       hyp=hyp,
+                                       cache=None if noval else opt.cache,
+                                       rect=True,
+                                       rank=-1,
+                                       workers=workers * 2,
+                                       pad=0.5,
+                                       prefix=colorstr('val: '))[0]
+
+        if not resume:
+            # if not opt.noautoanchor:
+            #     check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz)  # run AutoAnchor
+            model.half().float()  # pre-reduce anchor precision
+
+        callbacks.run('on_pretrain_routine_end', labels, names)
+
+    # DDP mode
+    if cuda and RANK != -1:
+        model = smart_DDP(model)
+
+    # Model attributes
+    nl = de_parallel(model).model[-1].nl  # number of detection layers (to scale hyps)
+    #hyp['box'] *= 3 / nl  # scale to layers
+    #hyp['cls'] *= nc / 80 * 3 / nl  # scale to classes and layers
+    #hyp['obj'] *= (imgsz / 640) ** 2 * 3 / nl  # scale to image size and layers
+    hyp['label_smoothing'] = opt.label_smoothing
+    model.nc = nc  # attach number of classes to model
+    model.hyp = hyp  # attach hyperparameters to model
+    model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) * nc  # attach class weights
+    model.names = names
+
+    # Start training
+    t0 = time.time()
+    nb = len(train_loader)  # number of batches
+    nw = max(round(hyp['warmup_epochs'] * nb), 100)  # number of warmup iterations, max(3 epochs, 100 iterations)
+    # nw = min(nw, (epochs - start_epoch) / 2 * nb)  # limit warmup to < 1/2 of training
+    last_opt_step = -1
+    maps = np.zeros(nc)  # mAP per class
+    results = (0, 0, 0, 0, 0, 0, 0)  # P, R, [email protected], [email protected], val_loss(box, obj, cls)
+    scheduler.last_epoch = start_epoch - 1  # do not move
+    scaler = torch.cuda.amp.GradScaler(enabled=amp)
+    stopper, stop = EarlyStopping(patience=opt.patience), False
+    compute_loss = ComputeLoss(model)  # init loss class
+    callbacks.run('on_train_start')
+    LOGGER.info(f'Image sizes {imgsz} train, {imgsz} val\n'
+                f'Using {train_loader.num_workers * WORLD_SIZE} dataloader workers\n'
+                f"Logging results to {colorstr('bold', save_dir)}\n"
+                f'Starting training for {epochs} epochs...')
+    for epoch in range(start_epoch, epochs):  # epoch ------------------------------------------------------------------
+        callbacks.run('on_train_epoch_start')
+        model.train()
+
+        # Update image weights (optional, single-GPU only)
+        if opt.image_weights:
+            cw = model.class_weights.cpu().numpy() * (1 - maps) ** 2 / nc  # class weights
+            iw = labels_to_image_weights(dataset.labels, nc=nc, class_weights=cw)  # image weights
+            dataset.indices = random.choices(range(dataset.n), weights=iw, k=dataset.n)  # rand weighted idx
+        if epoch == (epochs - opt.close_mosaic):
+            LOGGER.info("Closing dataloader mosaic")
+            dataset.mosaic = False
+
+        # Update mosaic border (optional)
+        # b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs)
+        # dataset.mosaic_border = [b - imgsz, -b]  # height, width borders
+
+        mloss = torch.zeros(3, device=device)  # mean losses
+        if RANK != -1:
+            train_loader.sampler.set_epoch(epoch)
+        pbar = enumerate(train_loader)
+        LOGGER.info(('\n' + '%11s' * 7) % ('Epoch', 'GPU_mem', 'box_loss', 'cls_loss', 'dfl_loss', 'Instances', 'Size'))
+        if RANK in {-1, 0}:
+            pbar = tqdm(pbar, total=nb, bar_format=TQDM_BAR_FORMAT)  # progress bar
+        optimizer.zero_grad()
+        for i, (imgs, targets, paths, _) in pbar:  # batch -------------------------------------------------------------
+            callbacks.run('on_train_batch_start')
+            ni = i + nb * epoch  # number integrated batches (since train start)
+            imgs = imgs.to(device, non_blocking=True).float() / 255  # uint8 to float32, 0-255 to 0.0-1.0
+
+            # Warmup
+            if ni <= nw:
+                xi = [0, nw]  # x interp
+                # compute_loss.gr = np.interp(ni, xi, [0.0, 1.0])  # iou loss ratio (obj_loss = 1.0 or iou)
+                accumulate = max(1, np.interp(ni, xi, [1, nbs / batch_size]).round())
+                for j, x in enumerate(optimizer.param_groups):
+                    # bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
+                    x['lr'] = np.interp(ni, xi, [hyp['warmup_bias_lr'] if j == 0 else 0.0, x['initial_lr'] * lf(epoch)])
+                    if 'momentum' in x:
+                        x['momentum'] = np.interp(ni, xi, [hyp['warmup_momentum'], hyp['momentum']])
+
+            # Multi-scale
+            if opt.multi_scale:
+                sz = random.randrange(imgsz * 0.5, imgsz * 1.5 + gs) // gs * gs  # size
+                sf = sz / max(imgs.shape[2:])  # scale factor
+                if sf != 1:
+                    ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]]  # new shape (stretched to gs-multiple)
+                    imgs = nn.functional.interpolate(imgs, size=ns, mode='bilinear', align_corners=False)
+
+            # Forward
+            with torch.cuda.amp.autocast(amp):
+                pred = model(imgs)  # forward
+                loss, loss_items = compute_loss(pred, targets.to(device))  # loss scaled by batch_size
+                if RANK != -1:
+                    loss *= WORLD_SIZE  # gradient averaged between devices in DDP mode
+                if opt.quad:
+                    loss *= 4.
+
+            # Backward
+            scaler.scale(loss).backward()
+
+            # Optimize - https://pytorch.org/docs/master/notes/amp_examples.html
+            if ni - last_opt_step >= accumulate:
+                scaler.unscale_(optimizer)  # unscale gradients
+                torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=10.0)  # clip gradients
+                scaler.step(optimizer)  # optimizer.step
+                scaler.update()
+                optimizer.zero_grad()
+                if ema:
+                    ema.update(model)
+                last_opt_step = ni
+
+            # Log
+            if RANK in {-1, 0}:
+                mloss = (mloss * i + loss_items) / (i + 1)  # update mean losses
+                mem = f'{torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0:.3g}G'  # (GB)
+                pbar.set_description(('%11s' * 2 + '%11.4g' * 5) %
+                                     (f'{epoch}/{epochs - 1}', mem, *mloss, targets.shape[0], imgs.shape[-1]))
+                callbacks.run('on_train_batch_end', model, ni, imgs, targets, paths, list(mloss))
+                if callbacks.stop_training:
+                    return
+            # end batch ------------------------------------------------------------------------------------------------
+
+        # Scheduler
+        lr = [x['lr'] for x in optimizer.param_groups]  # for loggers
+        scheduler.step()
+
+        if RANK in {-1, 0}:
+            # mAP
+            callbacks.run('on_train_epoch_end', epoch=epoch)
+            ema.update_attr(model, include=['yaml', 'nc', 'hyp', 'names', 'stride', 'class_weights'])
+            final_epoch = (epoch + 1 == epochs) or stopper.possible_stop
+            if not noval or final_epoch:  # Calculate mAP
+                results, maps, _ = validate.run(data_dict,
+                                                batch_size=batch_size // WORLD_SIZE * 2,
+                                                imgsz=imgsz,
+                                                half=amp,
+                                                model=ema.ema,
+                                                single_cls=single_cls,
+                                                dataloader=val_loader,
+                                                save_dir=save_dir,
+                                                plots=False,
+                                                callbacks=callbacks,
+                                                compute_loss=compute_loss)
+
+            # Update best mAP
+            fi = fitness(np.array(results).reshape(1, -1))  # weighted combination of [P, R, [email protected], [email protected]]
+            stop = stopper(epoch=epoch, fitness=fi)  # early stop check
+            if fi > best_fitness:
+                best_fitness = fi
+            log_vals = list(mloss) + list(results) + lr
+            callbacks.run('on_fit_epoch_end', log_vals, epoch, best_fitness, fi)
+
+            # Save model
+            if (not nosave) or (final_epoch and not evolve):  # if save
+                ckpt = {
+                    'epoch': epoch,
+                    'best_fitness': best_fitness,
+                    'model': deepcopy(de_parallel(model)).half(),
+                    'ema': deepcopy(ema.ema).half(),
+                    'updates': ema.updates,
+                    'optimizer': optimizer.state_dict(),
+                    'opt': vars(opt),
+                    'git': GIT_INFO,  # {remote, branch, commit} if a git repo
+                    'date': datetime.now().isoformat()}
+
+                # Save last, best and delete
+                torch.save(ckpt, last)
+                if best_fitness == fi:
+                    torch.save(ckpt, best)
+                if opt.save_period > 0 and epoch % opt.save_period == 0:
+                    torch.save(ckpt, w / f'epoch{epoch}.pt')
+                del ckpt
+                callbacks.run('on_model_save', last, epoch, final_epoch, best_fitness, fi)
+
+        # EarlyStopping
+        if RANK != -1:  # if DDP training
+            broadcast_list = [stop if RANK == 0 else None]
+            dist.broadcast_object_list(broadcast_list, 0)  # broadcast 'stop' to all ranks
+            if RANK != 0:
+                stop = broadcast_list[0]
+        if stop:
+            break  # must break all DDP ranks
+
+        # end epoch ----------------------------------------------------------------------------------------------------
+    # end training -----------------------------------------------------------------------------------------------------
+    if RANK in {-1, 0}:
+        LOGGER.info(f'\n{epoch - start_epoch + 1} epochs completed in {(time.time() - t0) / 3600:.3f} hours.')
+        for f in last, best:
+            if f.exists():
+                strip_optimizer(f)  # strip optimizers
+                if f is best:
+                    LOGGER.info(f'\nValidating {f}...')
+                    results, _, _ = validate.run(
+                        data_dict,
+                        batch_size=batch_size // WORLD_SIZE * 2,
+                        imgsz=imgsz,
+                        model=attempt_load(f, device).half(),
+                        single_cls=single_cls,
+                        dataloader=val_loader,
+                        save_dir=save_dir,
+                        save_json=is_coco,
+                        verbose=True,
+                        plots=plots,
+                        callbacks=callbacks,
+                        compute_loss=compute_loss)  # val best model with plots
+                    if is_coco:
+                        callbacks.run('on_fit_epoch_end', list(mloss) + list(results) + lr, epoch, best_fitness, fi)
+
+        callbacks.run('on_train_end', last, best, epoch, results)
+
+    torch.cuda.empty_cache()
+    return results
+
+
+def parse_opt(known=False):
+    parser = argparse.ArgumentParser()
+    # parser.add_argument('--weights', type=str, default=ROOT / 'yolo.pt', help='initial weights path')
+    # parser.add_argument('--cfg', type=str, default='', help='model.yaml path')
+    parser.add_argument('--weights', type=str, default='', help='initial weights path')
+    parser.add_argument('--cfg', type=str, default='yolo.yaml', help='model.yaml path')
+    parser.add_argument('--data', type=str, default=ROOT / 'data/coco.yaml', help='dataset.yaml path')
+    parser.add_argument('--hyp', type=str, default=ROOT / 'data/hyps/hyp.scratch-high.yaml', help='hyperparameters path')
+    parser.add_argument('--epochs', type=int, default=100, help='total training epochs')
+    parser.add_argument('--batch-size', type=int, default=16, help='total batch size for all GPUs, -1 for autobatch')
+    parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=640, help='train, val image size (pixels)')
+    parser.add_argument('--rect', action='store_true', help='rectangular training')
+    parser.add_argument('--resume', nargs='?', const=True, default=False, help='resume most recent training')
+    parser.add_argument('--nosave', action='store_true', help='only save final checkpoint')
+    parser.add_argument('--noval', action='store_true', help='only validate final epoch')
+    parser.add_argument('--noautoanchor', action='store_true', help='disable AutoAnchor')
+    parser.add_argument('--noplots', action='store_true', help='save no plot files')
+    parser.add_argument('--evolve', type=int, nargs='?', const=300, help='evolve hyperparameters for x generations')
+    parser.add_argument('--bucket', type=str, default='', help='gsutil bucket')
+    parser.add_argument('--cache', type=str, nargs='?', const='ram', help='image --cache ram/disk')
+    parser.add_argument('--image-weights', action='store_true', help='use weighted image selection for training')
+    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
+    parser.add_argument('--multi-scale', action='store_true', help='vary img-size +/- 50%%')
+    parser.add_argument('--single-cls', action='store_true', help='train multi-class data as single-class')
+    parser.add_argument('--optimizer', type=str, choices=['SGD', 'Adam', 'AdamW', 'LION'], default='SGD', help='optimizer')
+    parser.add_argument('--sync-bn', action='store_true', help='use SyncBatchNorm, only available in DDP mode')
+    parser.add_argument('--workers', type=int, default=8, help='max dataloader workers (per RANK in DDP mode)')
+    parser.add_argument('--project', default=ROOT / 'runs/train', help='save to project/name')
+    parser.add_argument('--name', default='exp', help='save to project/name')
+    parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')
+    parser.add_argument('--quad', action='store_true', help='quad dataloader')
+    parser.add_argument('--cos-lr', action='store_true', help='cosine LR scheduler')
+    parser.add_argument('--flat-cos-lr', action='store_true', help='flat cosine LR scheduler')
+    parser.add_argument('--fixed-lr', action='store_true', help='fixed LR scheduler')
+    parser.add_argument('--label-smoothing', type=float, default=0.0, help='Label smoothing epsilon')
+    parser.add_argument('--patience', type=int, default=100, help='EarlyStopping patience (epochs without improvement)')
+    parser.add_argument('--freeze', nargs='+', type=int, default=[0], help='Freeze layers: backbone=10, first3=0 1 2')
+    parser.add_argument('--save-period', type=int, default=-1, help='Save checkpoint every x epochs (disabled if < 1)')
+    parser.add_argument('--seed', type=int, default=0, help='Global training seed')
+    parser.add_argument('--local_rank', type=int, default=-1, help='Automatic DDP Multi-GPU argument, do not modify')
+    parser.add_argument('--min-items', type=int, default=0, help='Experimental')
+    parser.add_argument('--close-mosaic', type=int, default=0, help='Experimental')
+
+    # Logger arguments
+    parser.add_argument('--entity', default=None, help='Entity')
+    parser.add_argument('--upload_dataset', nargs='?', const=True, default=False, help='Upload data, "val" option')
+    parser.add_argument('--bbox_interval', type=int, default=-1, help='Set bounding-box image logging interval')
+    parser.add_argument('--artifact_alias', type=str, default='latest', help='Version of dataset artifact to use')
+
+    return parser.parse_known_args()[0] if known else parser.parse_args()
+
+
+def main(opt, callbacks=Callbacks()):
+    # Checks
+    if RANK in {-1, 0}:
+        print_args(vars(opt))
+        #check_git_status()
+        #check_requirements()
+
+    # Resume (from specified or most recent last.pt)
+    if opt.resume and not check_comet_resume(opt) and not opt.evolve:
+        last = Path(check_file(opt.resume) if isinstance(opt.resume, str) else get_latest_run())
+        opt_yaml = last.parent.parent / 'opt.yaml'  # train options yaml
+        opt_data = opt.data  # original dataset
+        if opt_yaml.is_file():
+            with open(opt_yaml, errors='ignore') as f:
+                d = yaml.safe_load(f)
+        else:
+            d = torch.load(last, map_location='cpu')['opt']
+        opt = argparse.Namespace(**d)  # replace
+        opt.cfg, opt.weights, opt.resume = '', str(last), True  # reinstate
+        if is_url(opt_data):
+            opt.data = check_file(opt_data)  # avoid HUB resume auth timeout
+    else:
+        opt.data, opt.cfg, opt.hyp, opt.weights, opt.project = \
+            check_file(opt.data), check_yaml(opt.cfg), check_yaml(opt.hyp), str(opt.weights), str(opt.project)  # checks
+        assert len(opt.cfg) or len(opt.weights), 'either --cfg or --weights must be specified'
+        if opt.evolve:
+            if opt.project == str(ROOT / 'runs/train'):  # if default project name, rename to runs/evolve
+                opt.project = str(ROOT / 'runs/evolve')
+            opt.exist_ok, opt.resume = opt.resume, False  # pass resume to exist_ok and disable resume
+        if opt.name == 'cfg':
+            opt.name = Path(opt.cfg).stem  # use model.yaml as name
+        opt.save_dir = str(increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok))
+
+    # DDP mode
+    device = select_device(opt.device, batch_size=opt.batch_size)
+    if LOCAL_RANK != -1:
+        msg = 'is not compatible with YOLO Multi-GPU DDP training'
+        assert not opt.image_weights, f'--image-weights {msg}'
+        assert not opt.evolve, f'--evolve {msg}'
+        assert opt.batch_size != -1, f'AutoBatch with --batch-size -1 {msg}, please pass a valid --batch-size'
+        assert opt.batch_size % WORLD_SIZE == 0, f'--batch-size {opt.batch_size} must be multiple of WORLD_SIZE'
+        assert torch.cuda.device_count() > LOCAL_RANK, 'insufficient CUDA devices for DDP command'
+        torch.cuda.set_device(LOCAL_RANK)
+        device = torch.device('cuda', LOCAL_RANK)
+        dist.init_process_group(backend="nccl" if dist.is_nccl_available() else "gloo")
+
+    # Train
+    if not opt.evolve:
+        train(opt.hyp, opt, device, callbacks)
+
+    # Evolve hyperparameters (optional)
+    else:
+        # Hyperparameter evolution metadata (mutation scale 0-1, lower_limit, upper_limit)
+        meta = {
+            'lr0': (1, 1e-5, 1e-1),  # initial learning rate (SGD=1E-2, Adam=1E-3)
+            'lrf': (1, 0.01, 1.0),  # final OneCycleLR learning rate (lr0 * lrf)
+            'momentum': (0.3, 0.6, 0.98),  # SGD momentum/Adam beta1
+            'weight_decay': (1, 0.0, 0.001),  # optimizer weight decay
+            'warmup_epochs': (1, 0.0, 5.0),  # warmup epochs (fractions ok)
+            'warmup_momentum': (1, 0.0, 0.95),  # warmup initial momentum
+            'warmup_bias_lr': (1, 0.0, 0.2),  # warmup initial bias lr
+            'box': (1, 0.02, 0.2),  # box loss gain
+            'cls': (1, 0.2, 4.0),  # cls loss gain
+            'cls_pw': (1, 0.5, 2.0),  # cls BCELoss positive_weight
+            'obj': (1, 0.2, 4.0),  # obj loss gain (scale with pixels)
+            'obj_pw': (1, 0.5, 2.0),  # obj BCELoss positive_weight
+            'iou_t': (0, 0.1, 0.7),  # IoU training threshold
+            'anchor_t': (1, 2.0, 8.0),  # anchor-multiple threshold
+            'anchors': (2, 2.0, 10.0),  # anchors per output grid (0 to ignore)
+            'fl_gamma': (0, 0.0, 2.0),  # focal loss gamma (efficientDet default gamma=1.5)
+            'hsv_h': (1, 0.0, 0.1),  # image HSV-Hue augmentation (fraction)
+            'hsv_s': (1, 0.0, 0.9),  # image HSV-Saturation augmentation (fraction)
+            'hsv_v': (1, 0.0, 0.9),  # image HSV-Value augmentation (fraction)
+            'degrees': (1, 0.0, 45.0),  # image rotation (+/- deg)
+            'translate': (1, 0.0, 0.9),  # image translation (+/- fraction)
+            'scale': (1, 0.0, 0.9),  # image scale (+/- gain)
+            'shear': (1, 0.0, 10.0),  # image shear (+/- deg)
+            'perspective': (0, 0.0, 0.001),  # image perspective (+/- fraction), range 0-0.001
+            'flipud': (1, 0.0, 1.0),  # image flip up-down (probability)
+            'fliplr': (0, 0.0, 1.0),  # image flip left-right (probability)
+            'mosaic': (1, 0.0, 1.0),  # image mixup (probability)
+            'mixup': (1, 0.0, 1.0),  # image mixup (probability)
+            'copy_paste': (1, 0.0, 1.0)}  # segment copy-paste (probability)
+
+        with open(opt.hyp, errors='ignore') as f:
+            hyp = yaml.safe_load(f)  # load hyps dict
+            if 'anchors' not in hyp:  # anchors commented in hyp.yaml
+                hyp['anchors'] = 3
+        if opt.noautoanchor:
+            del hyp['anchors'], meta['anchors']
+        opt.noval, opt.nosave, save_dir = True, True, Path(opt.save_dir)  # only val/save final epoch
+        # ei = [isinstance(x, (int, float)) for x in hyp.values()]  # evolvable indices
+        evolve_yaml, evolve_csv = save_dir / 'hyp_evolve.yaml', save_dir / 'evolve.csv'
+        if opt.bucket:
+            os.system(f'gsutil cp gs://{opt.bucket}/evolve.csv {evolve_csv}')  # download evolve.csv if exists
+
+        for _ in range(opt.evolve):  # generations to evolve
+            if evolve_csv.exists():  # if evolve.csv exists: select best hyps and mutate
+                # Select parent(s)
+                parent = 'single'  # parent selection method: 'single' or 'weighted'
+                x = np.loadtxt(evolve_csv, ndmin=2, delimiter=',', skiprows=1)
+                n = min(5, len(x))  # number of previous results to consider
+                x = x[np.argsort(-fitness(x))][:n]  # top n mutations
+                w = fitness(x) - fitness(x).min() + 1E-6  # weights (sum > 0)
+                if parent == 'single' or len(x) == 1:
+                    # x = x[random.randint(0, n - 1)]  # random selection
+                    x = x[random.choices(range(n), weights=w)[0]]  # weighted selection
+                elif parent == 'weighted':
+                    x = (x * w.reshape(n, 1)).sum(0) / w.sum()  # weighted combination
+
+                # Mutate
+                mp, s = 0.8, 0.2  # mutation probability, sigma
+                npr = np.random
+                npr.seed(int(time.time()))
+                g = np.array([meta[k][0] for k in hyp.keys()])  # gains 0-1
+                ng = len(meta)
+                v = np.ones(ng)
+                while all(v == 1):  # mutate until a change occurs (prevent duplicates)
+                    v = (g * (npr.random(ng) < mp) * npr.randn(ng) * npr.random() * s + 1).clip(0.3, 3.0)
+                for i, k in enumerate(hyp.keys()):  # plt.hist(v.ravel(), 300)
+                    hyp[k] = float(x[i + 7] * v[i])  # mutate
+
+            # Constrain to limits
+            for k, v in meta.items():
+                hyp[k] = max(hyp[k], v[1])  # lower limit
+                hyp[k] = min(hyp[k], v[2])  # upper limit
+                hyp[k] = round(hyp[k], 5)  # significant digits
+
+            # Train mutation
+            results = train(hyp.copy(), opt, device, callbacks)
+            callbacks = Callbacks()
+            # Write mutation results
+            keys = ('metrics/precision', 'metrics/recall', 'metrics/mAP_0.5', 'metrics/mAP_0.5:0.95', 'val/box_loss',
+                    'val/obj_loss', 'val/cls_loss')
+            print_mutation(keys, results, hyp.copy(), save_dir, opt.bucket)
+
+        # Plot results
+        plot_evolve(evolve_csv)
+        LOGGER.info(f'Hyperparameter evolution finished {opt.evolve} generations\n'
+                    f"Results saved to {colorstr('bold', save_dir)}\n"
+                    f'Usage example: $ python train.py --hyp {evolve_yaml}')
+
+
+def run(**kwargs):
+    # Usage: import train; train.run(data='coco128.yaml', imgsz=320, weights='yolo.pt')
+    opt = parse_opt(True)
+    for k, v in kwargs.items():
+        setattr(opt, k, v)
+    main(opt)
+    return opt
+
+
+if __name__ == "__main__":
+    opt = parse_opt()
+    main(opt)

train_triple.py

@@ -0,0 +1,636 @@
+import argparse
+import math
+import os
+import random
+import sys
+import time
+from copy import deepcopy
+from datetime import datetime
+from pathlib import Path
+
+import numpy as np
+import torch
+import torch.distributed as dist
+import torch.nn as nn
+import yaml
+from torch.optim import lr_scheduler
+from tqdm import tqdm
+
+FILE = Path(__file__).resolve()
+ROOT = FILE.parents[0]  # YOLO root directory
+if str(ROOT) not in sys.path:
+    sys.path.append(str(ROOT))  # add ROOT to PATH
+ROOT = Path(os.path.relpath(ROOT, Path.cwd()))  # relative
+
+import val_triple as validate  # for end-of-epoch mAP
+from models.experimental import attempt_load
+from models.yolo import Model
+from utils.autoanchor import check_anchors
+from utils.autobatch import check_train_batch_size
+from utils.callbacks import Callbacks
+from utils.dataloaders import create_dataloader
+from utils.downloads import attempt_download, is_url
+from utils.general import (LOGGER, TQDM_BAR_FORMAT, check_amp, check_dataset, check_file, check_git_info,
+                           check_git_status, check_img_size, check_requirements, check_suffix, check_yaml, colorstr,
+                           get_latest_run, increment_path, init_seeds, intersect_dicts, labels_to_class_weights,
+                           labels_to_image_weights, methods, one_cycle, print_args, print_mutation, strip_optimizer,
+                           yaml_save)
+from utils.loggers import Loggers
+from utils.loggers.comet.comet_utils import check_comet_resume
+from utils.loss_tal_triple import ComputeLoss
+from utils.metrics import fitness
+from utils.plots import plot_evolve
+from utils.torch_utils import (EarlyStopping, ModelEMA, de_parallel, select_device, smart_DDP, smart_optimizer,
+                               smart_resume, torch_distributed_zero_first)
+
+LOCAL_RANK = int(os.getenv('LOCAL_RANK', -1))  # https://pytorch.org/docs/stable/elastic/run.html
+RANK = int(os.getenv('RANK', -1))
+WORLD_SIZE = int(os.getenv('WORLD_SIZE', 1))
+GIT_INFO = None#check_git_info()
+
+
+def train(hyp, opt, device, callbacks):  # hyp is path/to/hyp.yaml or hyp dictionary
+    save_dir, epochs, batch_size, weights, single_cls, evolve, data, cfg, resume, noval, nosave, workers, freeze = \
+        Path(opt.save_dir), opt.epochs, opt.batch_size, opt.weights, opt.single_cls, opt.evolve, opt.data, opt.cfg, \
+        opt.resume, opt.noval, opt.nosave, opt.workers, opt.freeze
+    callbacks.run('on_pretrain_routine_start')
+
+    # Directories
+    w = save_dir / 'weights'  # weights dir
+    (w.parent if evolve else w).mkdir(parents=True, exist_ok=True)  # make dir
+    last, best = w / 'last.pt', w / 'best.pt'
+
+    # Hyperparameters
+    if isinstance(hyp, str):
+        with open(hyp, errors='ignore') as f:
+            hyp = yaml.safe_load(f)  # load hyps dict
+    LOGGER.info(colorstr('hyperparameters: ') + ', '.join(f'{k}={v}' for k, v in hyp.items()))
+    hyp['anchor_t'] = 5.0
+    opt.hyp = hyp.copy()  # for saving hyps to checkpoints
+
+    # Save run settings
+    if not evolve:
+        yaml_save(save_dir / 'hyp.yaml', hyp)
+        yaml_save(save_dir / 'opt.yaml', vars(opt))
+
+    # Loggers
+    data_dict = None
+    if RANK in {-1, 0}:
+        loggers = Loggers(save_dir, weights, opt, hyp, LOGGER)  # loggers instance
+
+        # Register actions
+        for k in methods(loggers):
+            callbacks.register_action(k, callback=getattr(loggers, k))
+
+        # Process custom dataset artifact link
+        data_dict = loggers.remote_dataset
+        if resume:  # If resuming runs from remote artifact
+            weights, epochs, hyp, batch_size = opt.weights, opt.epochs, opt.hyp, opt.batch_size
+
+    # Config
+    plots = not evolve and not opt.noplots  # create plots
+    cuda = device.type != 'cpu'
+    init_seeds(opt.seed + 1 + RANK, deterministic=True)
+    with torch_distributed_zero_first(LOCAL_RANK):
+        data_dict = data_dict or check_dataset(data)  # check if None
+    train_path, val_path = data_dict['train'], data_dict['val']
+    nc = 1 if single_cls else int(data_dict['nc'])  # number of classes
+    names = {0: 'item'} if single_cls and len(data_dict['names']) != 1 else data_dict['names']  # class names
+    #is_coco = isinstance(val_path, str) and val_path.endswith('coco/val2017.txt')  # COCO dataset
+    is_coco = isinstance(val_path, str) and val_path.endswith('val2017.txt')  # COCO dataset
+
+    # Model
+    check_suffix(weights, '.pt')  # check weights
+    pretrained = weights.endswith('.pt')
+    if pretrained:
+        with torch_distributed_zero_first(LOCAL_RANK):
+            weights = attempt_download(weights)  # download if not found locally
+        ckpt = torch.load(weights, map_location='cpu')  # load checkpoint to CPU to avoid CUDA memory leak
+        model = Model(cfg or ckpt['model'].yaml, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device)  # create
+        exclude = ['anchor'] if (cfg or hyp.get('anchors')) and not resume else []  # exclude keys
+        csd = ckpt['model'].float().state_dict()  # checkpoint state_dict as FP32
+        csd = intersect_dicts(csd, model.state_dict(), exclude=exclude)  # intersect
+        model.load_state_dict(csd, strict=False)  # load
+        LOGGER.info(f'Transferred {len(csd)}/{len(model.state_dict())} items from {weights}')  # report
+    else:
+        model = Model(cfg, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device)  # create
+    amp = check_amp(model)  # check AMP
+
+    # Freeze
+    freeze = [f'model.{x}.' for x in (freeze if len(freeze) > 1 else range(freeze[0]))]  # layers to freeze
+    for k, v in model.named_parameters():
+        # v.requires_grad = True  # train all layers TODO: uncomment this line as in master
+        # v.register_hook(lambda x: torch.nan_to_num(x))  # NaN to 0 (commented for erratic training results)
+        if any(x in k for x in freeze):
+            LOGGER.info(f'freezing {k}')
+            v.requires_grad = False
+
+    # Image size
+    gs = max(int(model.stride.max()), 32)  # grid size (max stride)
+    imgsz = check_img_size(opt.imgsz, gs, floor=gs * 2)  # verify imgsz is gs-multiple
+
+    # Batch size
+    if RANK == -1 and batch_size == -1:  # single-GPU only, estimate best batch size
+        batch_size = check_train_batch_size(model, imgsz, amp)
+        loggers.on_params_update({"batch_size": batch_size})
+
+    # Optimizer
+    nbs = 64  # nominal batch size
+    accumulate = max(round(nbs / batch_size), 1)  # accumulate loss before optimizing
+    hyp['weight_decay'] *= batch_size * accumulate / nbs  # scale weight_decay
+    optimizer = smart_optimizer(model, opt.optimizer, hyp['lr0'], hyp['momentum'], hyp['weight_decay'])
+
+    # Scheduler
+    if opt.cos_lr:
+        lf = one_cycle(1, hyp['lrf'], epochs)  # cosine 1->hyp['lrf']
+    else:
+        lf = lambda x: (1 - x / epochs) * (1.0 - hyp['lrf']) + hyp['lrf']  # linear
+
+    # def lf(x):  # saw
+    #     return (1 - (x % 30) / 30) * (1 - x / epochs) * (1.0 - hyp['lrf']) + hyp['lrf']
+    #
+    # def lf(x):  # triangle start at min
+    #     return 2 * abs(x / 30 - math.floor(x / 30 + 1 / 2)) * (1 - x / epochs) * (1.0 - hyp['lrf']) + hyp['lrf']
+    #
+    # def lf(x):  # triangle start at max
+    #     return 2 * abs(x / 32 + .5 - math.floor(x / 32 + 1)) * (1 - x / epochs) * (1.0 - hyp['lrf']) + hyp['lrf']
+
+    scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
+    # from utils.plots import plot_lr_scheduler; plot_lr_scheduler(optimizer, scheduler, epochs)
+
+    # EMA
+    ema = ModelEMA(model) if RANK in {-1, 0} else None
+
+    # Resume
+    best_fitness, start_epoch = 0.0, 0
+    if pretrained:
+        if resume:
+            best_fitness, start_epoch, epochs = smart_resume(ckpt, optimizer, ema, weights, epochs, resume)
+        del ckpt, csd
+
+    # DP mode
+    if cuda and RANK == -1 and torch.cuda.device_count() > 1:
+        LOGGER.warning('WARNING ⚠️ DP not recommended, use torch.distributed.run for best DDP Multi-GPU results.')
+        model = torch.nn.DataParallel(model)
+
+    # SyncBatchNorm
+    if opt.sync_bn and cuda and RANK != -1:
+        model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device)
+        LOGGER.info('Using SyncBatchNorm()')
+
+    # Trainloader
+    train_loader, dataset = create_dataloader(train_path,
+                                              imgsz,
+                                              batch_size // WORLD_SIZE,
+                                              gs,
+                                              single_cls,
+                                              hyp=hyp,
+                                              augment=True,
+                                              cache=None if opt.cache == 'val' else opt.cache,
+                                              rect=opt.rect,
+                                              rank=LOCAL_RANK,
+                                              workers=workers,
+                                              image_weights=opt.image_weights,
+                                              close_mosaic=opt.close_mosaic != 0,
+                                              quad=opt.quad,
+                                              prefix=colorstr('train: '),
+                                              shuffle=True,
+                                              min_items=opt.min_items)
+    labels = np.concatenate(dataset.labels, 0)
+    mlc = int(labels[:, 0].max())  # max label class
+    assert mlc < nc, f'Label class {mlc} exceeds nc={nc} in {data}. Possible class labels are 0-{nc - 1}'
+
+    # Process 0
+    if RANK in {-1, 0}:
+        val_loader = create_dataloader(val_path,
+                                       imgsz,
+                                       batch_size // WORLD_SIZE * 2,
+                                       gs,
+                                       single_cls,
+                                       hyp=hyp,
+                                       cache=None if noval else opt.cache,
+                                       rect=True,
+                                       rank=-1,
+                                       workers=workers * 2,
+                                       pad=0.5,
+                                       prefix=colorstr('val: '))[0]
+
+        if not resume:
+            # if not opt.noautoanchor:
+            #     check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz)  # run AutoAnchor
+            model.half().float()  # pre-reduce anchor precision
+
+        callbacks.run('on_pretrain_routine_end', labels, names)
+
+    # DDP mode
+    if cuda and RANK != -1:
+        model = smart_DDP(model)
+
+    # Model attributes
+    nl = de_parallel(model).model[-1].nl  # number of detection layers (to scale hyps)
+    #hyp['box'] *= 3 / nl  # scale to layers
+    #hyp['cls'] *= nc / 80 * 3 / nl  # scale to classes and layers
+    #hyp['obj'] *= (imgsz / 640) ** 2 * 3 / nl  # scale to image size and layers
+    hyp['label_smoothing'] = opt.label_smoothing
+    model.nc = nc  # attach number of classes to model
+    model.hyp = hyp  # attach hyperparameters to model
+    model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) * nc  # attach class weights
+    model.names = names
+
+    # Start training
+    t0 = time.time()
+    nb = len(train_loader)  # number of batches
+    nw = max(round(hyp['warmup_epochs'] * nb), 100)  # number of warmup iterations, max(3 epochs, 100 iterations)
+    # nw = min(nw, (epochs - start_epoch) / 2 * nb)  # limit warmup to < 1/2 of training
+    last_opt_step = -1
+    maps = np.zeros(nc)  # mAP per class
+    results = (0, 0, 0, 0, 0, 0, 0)  # P, R, [email protected], [email protected], val_loss(box, obj, cls)
+    scheduler.last_epoch = start_epoch - 1  # do not move
+    scaler = torch.cuda.amp.GradScaler(enabled=amp)
+    stopper, stop = EarlyStopping(patience=opt.patience), False
+    compute_loss = ComputeLoss(model)  # init loss class
+    callbacks.run('on_train_start')
+    LOGGER.info(f'Image sizes {imgsz} train, {imgsz} val\n'
+                f'Using {train_loader.num_workers * WORLD_SIZE} dataloader workers\n'
+                f"Logging results to {colorstr('bold', save_dir)}\n"
+                f'Starting training for {epochs} epochs...')
+    for epoch in range(start_epoch, epochs):  # epoch ------------------------------------------------------------------
+        callbacks.run('on_train_epoch_start')
+        model.train()
+
+        # Update image weights (optional, single-GPU only)
+        if opt.image_weights:
+            cw = model.class_weights.cpu().numpy() * (1 - maps) ** 2 / nc  # class weights
+            iw = labels_to_image_weights(dataset.labels, nc=nc, class_weights=cw)  # image weights
+            dataset.indices = random.choices(range(dataset.n), weights=iw, k=dataset.n)  # rand weighted idx
+        if epoch == (epochs - opt.close_mosaic):
+            LOGGER.info("Closing dataloader mosaic")
+            dataset.mosaic = False
+
+        # Update mosaic border (optional)
+        # b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs)
+        # dataset.mosaic_border = [b - imgsz, -b]  # height, width borders
+
+        mloss = torch.zeros(3, device=device)  # mean losses
+        if RANK != -1:
+            train_loader.sampler.set_epoch(epoch)
+        pbar = enumerate(train_loader)
+        LOGGER.info(('\n' + '%11s' * 7) % ('Epoch', 'GPU_mem', 'box_loss', 'cls_loss', 'dfl_loss', 'Instances', 'Size'))
+        if RANK in {-1, 0}:
+            pbar = tqdm(pbar, total=nb, bar_format=TQDM_BAR_FORMAT)  # progress bar
+        optimizer.zero_grad()
+        for i, (imgs, targets, paths, _) in pbar:  # batch -------------------------------------------------------------
+            callbacks.run('on_train_batch_start')
+            ni = i + nb * epoch  # number integrated batches (since train start)
+            imgs = imgs.to(device, non_blocking=True).float() / 255  # uint8 to float32, 0-255 to 0.0-1.0
+
+            # Warmup
+            if ni <= nw:
+                xi = [0, nw]  # x interp
+                # compute_loss.gr = np.interp(ni, xi, [0.0, 1.0])  # iou loss ratio (obj_loss = 1.0 or iou)
+                accumulate = max(1, np.interp(ni, xi, [1, nbs / batch_size]).round())
+                for j, x in enumerate(optimizer.param_groups):
+                    # bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
+                    x['lr'] = np.interp(ni, xi, [hyp['warmup_bias_lr'] if j == 0 else 0.0, x['initial_lr'] * lf(epoch)])
+                    if 'momentum' in x:
+                        x['momentum'] = np.interp(ni, xi, [hyp['warmup_momentum'], hyp['momentum']])
+
+            # Multi-scale
+            if opt.multi_scale:
+                sz = random.randrange(imgsz * 0.5, imgsz * 1.5 + gs) // gs * gs  # size
+                sf = sz / max(imgs.shape[2:])  # scale factor
+                if sf != 1:
+                    ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]]  # new shape (stretched to gs-multiple)
+                    imgs = nn.functional.interpolate(imgs, size=ns, mode='bilinear', align_corners=False)
+
+            # Forward
+            with torch.cuda.amp.autocast(amp):
+                pred = model(imgs)  # forward
+                loss, loss_items = compute_loss(pred, targets.to(device))  # loss scaled by batch_size
+                if RANK != -1:
+                    loss *= WORLD_SIZE  # gradient averaged between devices in DDP mode
+                if opt.quad:
+                    loss *= 4.
+
+            # Backward
+            scaler.scale(loss).backward()
+
+            # Optimize - https://pytorch.org/docs/master/notes/amp_examples.html
+            if ni - last_opt_step >= accumulate:
+                scaler.unscale_(optimizer)  # unscale gradients
+                torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=10.0)  # clip gradients
+                scaler.step(optimizer)  # optimizer.step
+                scaler.update()
+                optimizer.zero_grad()
+                if ema:
+                    ema.update(model)
+                last_opt_step = ni
+
+            # Log
+            if RANK in {-1, 0}:
+                mloss = (mloss * i + loss_items) / (i + 1)  # update mean losses
+                mem = f'{torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0:.3g}G'  # (GB)
+                pbar.set_description(('%11s' * 2 + '%11.4g' * 5) %
+                                     (f'{epoch}/{epochs - 1}', mem, *mloss, targets.shape[0], imgs.shape[-1]))
+                callbacks.run('on_train_batch_end', model, ni, imgs, targets, paths, list(mloss))
+                if callbacks.stop_training:
+                    return
+            # end batch ------------------------------------------------------------------------------------------------
+
+        # Scheduler
+        lr = [x['lr'] for x in optimizer.param_groups]  # for loggers
+        scheduler.step()
+
+        if RANK in {-1, 0}:
+            # mAP
+            callbacks.run('on_train_epoch_end', epoch=epoch)
+            ema.update_attr(model, include=['yaml', 'nc', 'hyp', 'names', 'stride', 'class_weights'])
+            final_epoch = (epoch + 1 == epochs) or stopper.possible_stop
+            if not noval or final_epoch:  # Calculate mAP
+                results, maps, _ = validate.run(data_dict,
+                                                batch_size=batch_size // WORLD_SIZE * 2,
+                                                imgsz=imgsz,
+                                                half=amp,
+                                                model=ema.ema,
+                                                single_cls=single_cls,
+                                                dataloader=val_loader,
+                                                save_dir=save_dir,
+                                                plots=False,
+                                                callbacks=callbacks,
+                                                compute_loss=compute_loss)
+
+            # Update best mAP
+            fi = fitness(np.array(results).reshape(1, -1))  # weighted combination of [P, R, [email protected], [email protected]]
+            stop = stopper(epoch=epoch, fitness=fi)  # early stop check
+            if fi > best_fitness:
+                best_fitness = fi
+            log_vals = list(mloss) + list(results) + lr
+            callbacks.run('on_fit_epoch_end', log_vals, epoch, best_fitness, fi)
+
+            # Save model
+            if (not nosave) or (final_epoch and not evolve):  # if save
+                ckpt = {
+                    'epoch': epoch,
+                    'best_fitness': best_fitness,
+                    'model': deepcopy(de_parallel(model)).half(),
+                    'ema': deepcopy(ema.ema).half(),
+                    'updates': ema.updates,
+                    'optimizer': optimizer.state_dict(),
+                    'opt': vars(opt),
+                    'git': GIT_INFO,  # {remote, branch, commit} if a git repo
+                    'date': datetime.now().isoformat()}
+
+                # Save last, best and delete
+                torch.save(ckpt, last)
+                if best_fitness == fi:
+                    torch.save(ckpt, best)
+                if opt.save_period > 0 and epoch % opt.save_period == 0:
+                    torch.save(ckpt, w / f'epoch{epoch}.pt')
+                del ckpt
+                callbacks.run('on_model_save', last, epoch, final_epoch, best_fitness, fi)
+
+        # EarlyStopping
+        if RANK != -1:  # if DDP training
+            broadcast_list = [stop if RANK == 0 else None]
+            dist.broadcast_object_list(broadcast_list, 0)  # broadcast 'stop' to all ranks
+            if RANK != 0:
+                stop = broadcast_list[0]
+        if stop:
+            break  # must break all DDP ranks
+
+        # end epoch ----------------------------------------------------------------------------------------------------
+    # end training -----------------------------------------------------------------------------------------------------
+    if RANK in {-1, 0}:
+        LOGGER.info(f'\n{epoch - start_epoch + 1} epochs completed in {(time.time() - t0) / 3600:.3f} hours.')
+        for f in last, best:
+            if f.exists():
+                strip_optimizer(f)  # strip optimizers
+                if f is best:
+                    LOGGER.info(f'\nValidating {f}...')
+                    results, _, _ = validate.run(
+                        data_dict,
+                        batch_size=batch_size // WORLD_SIZE * 2,
+                        imgsz=imgsz,
+                        model=attempt_load(f, device).half(),
+                        single_cls=single_cls,
+                        dataloader=val_loader,
+                        save_dir=save_dir,
+                        save_json=is_coco,
+                        verbose=True,
+                        plots=plots,
+                        callbacks=callbacks,
+                        compute_loss=compute_loss)  # val best model with plots
+                    if is_coco:
+                        callbacks.run('on_fit_epoch_end', list(mloss) + list(results) + lr, epoch, best_fitness, fi)
+
+        callbacks.run('on_train_end', last, best, epoch, results)
+
+    torch.cuda.empty_cache()
+    return results
+
+
+def parse_opt(known=False):
+    parser = argparse.ArgumentParser()
+    # parser.add_argument('--weights', type=str, default=ROOT / 'yolo.pt', help='initial weights path')
+    # parser.add_argument('--cfg', type=str, default='', help='model.yaml path')
+    parser.add_argument('--weights', type=str, default='', help='initial weights path')
+    parser.add_argument('--cfg', type=str, default='yolo.yaml', help='model.yaml path')
+    parser.add_argument('--data', type=str, default=ROOT / 'data/coco.yaml', help='dataset.yaml path')
+    parser.add_argument('--hyp', type=str, default=ROOT / 'data/hyps/hyp.scratch-high.yaml', help='hyperparameters path')
+    parser.add_argument('--epochs', type=int, default=100, help='total training epochs')
+    parser.add_argument('--batch-size', type=int, default=16, help='total batch size for all GPUs, -1 for autobatch')
+    parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=640, help='train, val image size (pixels)')
+    parser.add_argument('--rect', action='store_true', help='rectangular training')
+    parser.add_argument('--resume', nargs='?', const=True, default=False, help='resume most recent training')
+    parser.add_argument('--nosave', action='store_true', help='only save final checkpoint')
+    parser.add_argument('--noval', action='store_true', help='only validate final epoch')
+    parser.add_argument('--noautoanchor', action='store_true', help='disable AutoAnchor')
+    parser.add_argument('--noplots', action='store_true', help='save no plot files')
+    parser.add_argument('--evolve', type=int, nargs='?', const=300, help='evolve hyperparameters for x generations')
+    parser.add_argument('--bucket', type=str, default='', help='gsutil bucket')
+    parser.add_argument('--cache', type=str, nargs='?', const='ram', help='image --cache ram/disk')
+    parser.add_argument('--image-weights', action='store_true', help='use weighted image selection for training')
+    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
+    parser.add_argument('--multi-scale', action='store_true', help='vary img-size +/- 50%%')
+    parser.add_argument('--single-cls', action='store_true', help='train multi-class data as single-class')
+    parser.add_argument('--optimizer', type=str, choices=['SGD', 'Adam', 'AdamW', 'LION'], default='SGD', help='optimizer')
+    parser.add_argument('--sync-bn', action='store_true', help='use SyncBatchNorm, only available in DDP mode')
+    parser.add_argument('--workers', type=int, default=8, help='max dataloader workers (per RANK in DDP mode)')
+    parser.add_argument('--project', default=ROOT / 'runs/train', help='save to project/name')
+    parser.add_argument('--name', default='exp', help='save to project/name')
+    parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')
+    parser.add_argument('--quad', action='store_true', help='quad dataloader')
+    parser.add_argument('--cos-lr', action='store_true', help='cosine LR scheduler')
+    parser.add_argument('--label-smoothing', type=float, default=0.0, help='Label smoothing epsilon')
+    parser.add_argument('--patience', type=int, default=100, help='EarlyStopping patience (epochs without improvement)')
+    parser.add_argument('--freeze', nargs='+', type=int, default=[0], help='Freeze layers: backbone=10, first3=0 1 2')
+    parser.add_argument('--save-period', type=int, default=-1, help='Save checkpoint every x epochs (disabled if < 1)')
+    parser.add_argument('--seed', type=int, default=0, help='Global training seed')
+    parser.add_argument('--local_rank', type=int, default=-1, help='Automatic DDP Multi-GPU argument, do not modify')
+    parser.add_argument('--min-items', type=int, default=0, help='Experimental')
+    parser.add_argument('--close-mosaic', type=int, default=0, help='Experimental')
+
+    # Logger arguments
+    parser.add_argument('--entity', default=None, help='Entity')
+    parser.add_argument('--upload_dataset', nargs='?', const=True, default=False, help='Upload data, "val" option')
+    parser.add_argument('--bbox_interval', type=int, default=-1, help='Set bounding-box image logging interval')
+    parser.add_argument('--artifact_alias', type=str, default='latest', help='Version of dataset artifact to use')
+
+    return parser.parse_known_args()[0] if known else parser.parse_args()
+
+
+def main(opt, callbacks=Callbacks()):
+    # Checks
+    if RANK in {-1, 0}:
+        print_args(vars(opt))
+        #check_git_status()
+        #check_requirements()
+
+    # Resume (from specified or most recent last.pt)
+    if opt.resume and not check_comet_resume(opt) and not opt.evolve:
+        last = Path(check_file(opt.resume) if isinstance(opt.resume, str) else get_latest_run())
+        opt_yaml = last.parent.parent / 'opt.yaml'  # train options yaml
+        opt_data = opt.data  # original dataset
+        if opt_yaml.is_file():
+            with open(opt_yaml, errors='ignore') as f:
+                d = yaml.safe_load(f)
+        else:
+            d = torch.load(last, map_location='cpu')['opt']
+        opt = argparse.Namespace(**d)  # replace
+        opt.cfg, opt.weights, opt.resume = '', str(last), True  # reinstate
+        if is_url(opt_data):
+            opt.data = check_file(opt_data)  # avoid HUB resume auth timeout
+    else:
+        opt.data, opt.cfg, opt.hyp, opt.weights, opt.project = \
+            check_file(opt.data), check_yaml(opt.cfg), check_yaml(opt.hyp), str(opt.weights), str(opt.project)  # checks
+        assert len(opt.cfg) or len(opt.weights), 'either --cfg or --weights must be specified'
+        if opt.evolve:
+            if opt.project == str(ROOT / 'runs/train'):  # if default project name, rename to runs/evolve
+                opt.project = str(ROOT / 'runs/evolve')
+            opt.exist_ok, opt.resume = opt.resume, False  # pass resume to exist_ok and disable resume
+        if opt.name == 'cfg':
+            opt.name = Path(opt.cfg).stem  # use model.yaml as name
+        opt.save_dir = str(increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok))
+
+    # DDP mode
+    device = select_device(opt.device, batch_size=opt.batch_size)
+    if LOCAL_RANK != -1:
+        msg = 'is not compatible with YOLO Multi-GPU DDP training'
+        assert not opt.image_weights, f'--image-weights {msg}'
+        assert not opt.evolve, f'--evolve {msg}'
+        assert opt.batch_size != -1, f'AutoBatch with --batch-size -1 {msg}, please pass a valid --batch-size'
+        assert opt.batch_size % WORLD_SIZE == 0, f'--batch-size {opt.batch_size} must be multiple of WORLD_SIZE'
+        assert torch.cuda.device_count() > LOCAL_RANK, 'insufficient CUDA devices for DDP command'
+        torch.cuda.set_device(LOCAL_RANK)
+        device = torch.device('cuda', LOCAL_RANK)
+        dist.init_process_group(backend="nccl" if dist.is_nccl_available() else "gloo")
+
+    # Train
+    if not opt.evolve:
+        train(opt.hyp, opt, device, callbacks)
+
+    # Evolve hyperparameters (optional)
+    else:
+        # Hyperparameter evolution metadata (mutation scale 0-1, lower_limit, upper_limit)
+        meta = {
+            'lr0': (1, 1e-5, 1e-1),  # initial learning rate (SGD=1E-2, Adam=1E-3)
+            'lrf': (1, 0.01, 1.0),  # final OneCycleLR learning rate (lr0 * lrf)
+            'momentum': (0.3, 0.6, 0.98),  # SGD momentum/Adam beta1
+            'weight_decay': (1, 0.0, 0.001),  # optimizer weight decay
+            'warmup_epochs': (1, 0.0, 5.0),  # warmup epochs (fractions ok)
+            'warmup_momentum': (1, 0.0, 0.95),  # warmup initial momentum
+            'warmup_bias_lr': (1, 0.0, 0.2),  # warmup initial bias lr
+            'box': (1, 0.02, 0.2),  # box loss gain
+            'cls': (1, 0.2, 4.0),  # cls loss gain
+            'cls_pw': (1, 0.5, 2.0),  # cls BCELoss positive_weight
+            'obj': (1, 0.2, 4.0),  # obj loss gain (scale with pixels)
+            'obj_pw': (1, 0.5, 2.0),  # obj BCELoss positive_weight
+            'iou_t': (0, 0.1, 0.7),  # IoU training threshold
+            'anchor_t': (1, 2.0, 8.0),  # anchor-multiple threshold
+            'anchors': (2, 2.0, 10.0),  # anchors per output grid (0 to ignore)
+            'fl_gamma': (0, 0.0, 2.0),  # focal loss gamma (efficientDet default gamma=1.5)
+            'hsv_h': (1, 0.0, 0.1),  # image HSV-Hue augmentation (fraction)
+            'hsv_s': (1, 0.0, 0.9),  # image HSV-Saturation augmentation (fraction)
+            'hsv_v': (1, 0.0, 0.9),  # image HSV-Value augmentation (fraction)
+            'degrees': (1, 0.0, 45.0),  # image rotation (+/- deg)
+            'translate': (1, 0.0, 0.9),  # image translation (+/- fraction)
+            'scale': (1, 0.0, 0.9),  # image scale (+/- gain)
+            'shear': (1, 0.0, 10.0),  # image shear (+/- deg)
+            'perspective': (0, 0.0, 0.001),  # image perspective (+/- fraction), range 0-0.001
+            'flipud': (1, 0.0, 1.0),  # image flip up-down (probability)
+            'fliplr': (0, 0.0, 1.0),  # image flip left-right (probability)
+            'mosaic': (1, 0.0, 1.0),  # image mixup (probability)
+            'mixup': (1, 0.0, 1.0),  # image mixup (probability)
+            'copy_paste': (1, 0.0, 1.0)}  # segment copy-paste (probability)
+
+        with open(opt.hyp, errors='ignore') as f:
+            hyp = yaml.safe_load(f)  # load hyps dict
+            if 'anchors' not in hyp:  # anchors commented in hyp.yaml
+                hyp['anchors'] = 3
+        if opt.noautoanchor:
+            del hyp['anchors'], meta['anchors']
+        opt.noval, opt.nosave, save_dir = True, True, Path(opt.save_dir)  # only val/save final epoch
+        # ei = [isinstance(x, (int, float)) for x in hyp.values()]  # evolvable indices
+        evolve_yaml, evolve_csv = save_dir / 'hyp_evolve.yaml', save_dir / 'evolve.csv'
+        if opt.bucket:
+            os.system(f'gsutil cp gs://{opt.bucket}/evolve.csv {evolve_csv}')  # download evolve.csv if exists
+
+        for _ in range(opt.evolve):  # generations to evolve
+            if evolve_csv.exists():  # if evolve.csv exists: select best hyps and mutate
+                # Select parent(s)
+                parent = 'single'  # parent selection method: 'single' or 'weighted'
+                x = np.loadtxt(evolve_csv, ndmin=2, delimiter=',', skiprows=1)
+                n = min(5, len(x))  # number of previous results to consider
+                x = x[np.argsort(-fitness(x))][:n]  # top n mutations
+                w = fitness(x) - fitness(x).min() + 1E-6  # weights (sum > 0)
+                if parent == 'single' or len(x) == 1:
+                    # x = x[random.randint(0, n - 1)]  # random selection
+                    x = x[random.choices(range(n), weights=w)[0]]  # weighted selection
+                elif parent == 'weighted':
+                    x = (x * w.reshape(n, 1)).sum(0) / w.sum()  # weighted combination
+
+                # Mutate
+                mp, s = 0.8, 0.2  # mutation probability, sigma
+                npr = np.random
+                npr.seed(int(time.time()))
+                g = np.array([meta[k][0] for k in hyp.keys()])  # gains 0-1
+                ng = len(meta)
+                v = np.ones(ng)
+                while all(v == 1):  # mutate until a change occurs (prevent duplicates)
+                    v = (g * (npr.random(ng) < mp) * npr.randn(ng) * npr.random() * s + 1).clip(0.3, 3.0)
+                for i, k in enumerate(hyp.keys()):  # plt.hist(v.ravel(), 300)
+                    hyp[k] = float(x[i + 7] * v[i])  # mutate
+
+            # Constrain to limits
+            for k, v in meta.items():
+                hyp[k] = max(hyp[k], v[1])  # lower limit
+                hyp[k] = min(hyp[k], v[2])  # upper limit
+                hyp[k] = round(hyp[k], 5)  # significant digits
+
+            # Train mutation
+            results = train(hyp.copy(), opt, device, callbacks)
+            callbacks = Callbacks()
+            # Write mutation results
+            keys = ('metrics/precision', 'metrics/recall', 'metrics/mAP_0.5', 'metrics/mAP_0.5:0.95', 'val/box_loss',
+                    'val/obj_loss', 'val/cls_loss')
+            print_mutation(keys, results, hyp.copy(), save_dir, opt.bucket)
+
+        # Plot results
+        plot_evolve(evolve_csv)
+        LOGGER.info(f'Hyperparameter evolution finished {opt.evolve} generations\n'
+                    f"Results saved to {colorstr('bold', save_dir)}\n"
+                    f'Usage example: $ python train.py --hyp {evolve_yaml}')
+
+
+def run(**kwargs):
+    # Usage: import train; train.run(data='coco128.yaml', imgsz=320, weights='yolo.pt')
+    opt = parse_opt(True)
+    for k, v in kwargs.items():
+        setattr(opt, k, v)
+    main(opt)
+    return opt
+
+
+if __name__ == "__main__":
+    opt = parse_opt()
+    main(opt)

utils/__init__.py

@@ -0,0 +1,75 @@
+import contextlib
+import platform
+import threading
+
+
+def emojis(str=''):
+    # Return platform-dependent emoji-safe version of string
+    return str.encode().decode('ascii', 'ignore') if platform.system() == 'Windows' else str
+
+
+class TryExcept(contextlib.ContextDecorator):
+    # YOLOv5 TryExcept class. Usage: @TryExcept() decorator or 'with TryExcept():' context manager
+    def __init__(self, msg=''):
+        self.msg = msg
+
+    def __enter__(self):
+        pass
+
+    def __exit__(self, exc_type, value, traceback):
+        if value:
+            print(emojis(f"{self.msg}{': ' if self.msg else ''}{value}"))
+        return True
+
+
+def threaded(func):
+    # Multi-threads a target function and returns thread. Usage: @threaded decorator
+    def wrapper(*args, **kwargs):
+        thread = threading.Thread(target=func, args=args, kwargs=kwargs, daemon=True)
+        thread.start()
+        return thread
+
+    return wrapper
+
+
+def join_threads(verbose=False):
+    # Join all daemon threads, i.e. atexit.register(lambda: join_threads())
+    main_thread = threading.current_thread()
+    for t in threading.enumerate():
+        if t is not main_thread:
+            if verbose:
+                print(f'Joining thread {t.name}')
+            t.join()
+
+
+def notebook_init(verbose=True):
+    # Check system software and hardware
+    print('Checking setup...')
+
+    import os
+    import shutil
+
+    from utils.general import check_font, check_requirements, is_colab
+    from utils.torch_utils import select_device  # imports
+
+    check_font()
+
+    import psutil
+    from IPython import display  # to display images and clear console output
+
+    if is_colab():
+        shutil.rmtree('/content/sample_data', ignore_errors=True)  # remove colab /sample_data directory
+
+    # System info
+    if verbose:
+        gb = 1 << 30  # bytes to GiB (1024 ** 3)
+        ram = psutil.virtual_memory().total
+        total, used, free = shutil.disk_usage("/")
+        display.clear_output()
+        s = f'({os.cpu_count()} CPUs, {ram / gb:.1f} GB RAM, {(total - free) / gb:.1f}/{total / gb:.1f} GB disk)'
+    else:
+        s = ''
+
+    select_device(newline=False)
+    print(emojis(f'Setup complete ✅ {s}'))
+    return display

utils/activations.py

@@ -0,0 +1,98 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class SiLU(nn.Module):
+    # SiLU activation https://arxiv.org/pdf/1606.08415.pdf
+    @staticmethod
+    def forward(x):
+        return x * torch.sigmoid(x)
+
+
+class Hardswish(nn.Module):
+    # Hard-SiLU activation
+    @staticmethod
+    def forward(x):
+        # return x * F.hardsigmoid(x)  # for TorchScript and CoreML
+        return x * F.hardtanh(x + 3, 0.0, 6.0) / 6.0  # for TorchScript, CoreML and ONNX
+
+
+class Mish(nn.Module):
+    # Mish activation https://github.com/digantamisra98/Mish
+    @staticmethod
+    def forward(x):
+        return x * F.softplus(x).tanh()
+
+
+class MemoryEfficientMish(nn.Module):
+    # Mish activation memory-efficient
+    class F(torch.autograd.Function):
+
+        @staticmethod
+        def forward(ctx, x):
+            ctx.save_for_backward(x)
+            return x.mul(torch.tanh(F.softplus(x)))  # x * tanh(ln(1 + exp(x)))
+
+        @staticmethod
+        def backward(ctx, grad_output):
+            x = ctx.saved_tensors[0]
+            sx = torch.sigmoid(x)
+            fx = F.softplus(x).tanh()
+            return grad_output * (fx + x * sx * (1 - fx * fx))
+
+    def forward(self, x):
+        return self.F.apply(x)
+
+
+class FReLU(nn.Module):
+    # FReLU activation https://arxiv.org/abs/2007.11824
+    def __init__(self, c1, k=3):  # ch_in, kernel
+        super().__init__()
+        self.conv = nn.Conv2d(c1, c1, k, 1, 1, groups=c1, bias=False)
+        self.bn = nn.BatchNorm2d(c1)
+
+    def forward(self, x):
+        return torch.max(x, self.bn(self.conv(x)))
+
+
+class AconC(nn.Module):
+    r""" ACON activation (activate or not)
+    AconC: (p1*x-p2*x) * sigmoid(beta*(p1*x-p2*x)) + p2*x, beta is a learnable parameter
+    according to "Activate or Not: Learning Customized Activation" <https://arxiv.org/pdf/2009.04759.pdf>.
+    """
+
+    def __init__(self, c1):
+        super().__init__()
+        self.p1 = nn.Parameter(torch.randn(1, c1, 1, 1))
+        self.p2 = nn.Parameter(torch.randn(1, c1, 1, 1))
+        self.beta = nn.Parameter(torch.ones(1, c1, 1, 1))
+
+    def forward(self, x):
+        dpx = (self.p1 - self.p2) * x
+        return dpx * torch.sigmoid(self.beta * dpx) + self.p2 * x
+
+
+class MetaAconC(nn.Module):
+    r""" ACON activation (activate or not)
+    MetaAconC: (p1*x-p2*x) * sigmoid(beta*(p1*x-p2*x)) + p2*x, beta is generated by a small network
+    according to "Activate or Not: Learning Customized Activation" <https://arxiv.org/pdf/2009.04759.pdf>.
+    """
+
+    def __init__(self, c1, k=1, s=1, r=16):  # ch_in, kernel, stride, r
+        super().__init__()
+        c2 = max(r, c1 // r)
+        self.p1 = nn.Parameter(torch.randn(1, c1, 1, 1))
+        self.p2 = nn.Parameter(torch.randn(1, c1, 1, 1))
+        self.fc1 = nn.Conv2d(c1, c2, k, s, bias=True)
+        self.fc2 = nn.Conv2d(c2, c1, k, s, bias=True)
+        # self.bn1 = nn.BatchNorm2d(c2)
+        # self.bn2 = nn.BatchNorm2d(c1)
+
+    def forward(self, x):
+        y = x.mean(dim=2, keepdims=True).mean(dim=3, keepdims=True)
+        # batch-size 1 bug/instabilities https://github.com/ultralytics/yolov5/issues/2891
+        # beta = torch.sigmoid(self.bn2(self.fc2(self.bn1(self.fc1(y)))))  # bug/unstable
+        beta = torch.sigmoid(self.fc2(self.fc1(y)))  # bug patch BN layers removed
+        dpx = (self.p1 - self.p2) * x
+        return dpx * torch.sigmoid(beta * dpx) + self.p2 * x

utils/augmentations.py

@@ -0,0 +1,395 @@
+import math
+import random
+
+import cv2
+import numpy as np
+import torch
+import torchvision.transforms as T
+import torchvision.transforms.functional as TF
+
+from utils.general import LOGGER, check_version, colorstr, resample_segments, segment2box, xywhn2xyxy
+from utils.metrics import bbox_ioa
+
+IMAGENET_MEAN = 0.485, 0.456, 0.406  # RGB mean
+IMAGENET_STD = 0.229, 0.224, 0.225  # RGB standard deviation
+
+
+class Albumentations:
+    # YOLOv5 Albumentations class (optional, only used if package is installed)
+    def __init__(self, size=640):
+        self.transform = None
+        prefix = colorstr('albumentations: ')
+        try:
+            import albumentations as A
+            check_version(A.__version__, '1.0.3', hard=True)  # version requirement
+
+            T = [
+                A.RandomResizedCrop(height=size, width=size, scale=(0.8, 1.0), ratio=(0.9, 1.11), p=0.0),
+                A.Blur(p=0.01),
+                A.MedianBlur(p=0.01),
+                A.ToGray(p=0.01),
+                A.CLAHE(p=0.01),
+                A.RandomBrightnessContrast(p=0.0),
+                A.RandomGamma(p=0.0),
+                A.ImageCompression(quality_lower=75, p=0.0)]  # transforms
+            self.transform = A.Compose(T, bbox_params=A.BboxParams(format='yolo', label_fields=['class_labels']))
+
+            LOGGER.info(prefix + ', '.join(f'{x}'.replace('always_apply=False, ', '') for x in T if x.p))
+        except ImportError:  # package not installed, skip
+            pass
+        except Exception as e:
+            LOGGER.info(f'{prefix}{e}')
+
+    def __call__(self, im, labels, p=1.0):
+        if self.transform and random.random() < p:
+            new = self.transform(image=im, bboxes=labels[:, 1:], class_labels=labels[:, 0])  # transformed
+            im, labels = new['image'], np.array([[c, *b] for c, b in zip(new['class_labels'], new['bboxes'])])
+        return im, labels
+
+
+def normalize(x, mean=IMAGENET_MEAN, std=IMAGENET_STD, inplace=False):
+    # Denormalize RGB images x per ImageNet stats in BCHW format, i.e. = (x - mean) / std
+    return TF.normalize(x, mean, std, inplace=inplace)
+
+
+def denormalize(x, mean=IMAGENET_MEAN, std=IMAGENET_STD):
+    # Denormalize RGB images x per ImageNet stats in BCHW format, i.e. = x * std + mean
+    for i in range(3):
+        x[:, i] = x[:, i] * std[i] + mean[i]
+    return x
+
+
+def augment_hsv(im, hgain=0.5, sgain=0.5, vgain=0.5):
+    # HSV color-space augmentation
+    if hgain or sgain or vgain:
+        r = np.random.uniform(-1, 1, 3) * [hgain, sgain, vgain] + 1  # random gains
+        hue, sat, val = cv2.split(cv2.cvtColor(im, cv2.COLOR_BGR2HSV))
+        dtype = im.dtype  # uint8
+
+        x = np.arange(0, 256, dtype=r.dtype)
+        lut_hue = ((x * r[0]) % 180).astype(dtype)
+        lut_sat = np.clip(x * r[1], 0, 255).astype(dtype)
+        lut_val = np.clip(x * r[2], 0, 255).astype(dtype)
+
+        im_hsv = cv2.merge((cv2.LUT(hue, lut_hue), cv2.LUT(sat, lut_sat), cv2.LUT(val, lut_val)))
+        cv2.cvtColor(im_hsv, cv2.COLOR_HSV2BGR, dst=im)  # no return needed
+
+
+def hist_equalize(im, clahe=True, bgr=False):
+    # Equalize histogram on BGR image 'im' with im.shape(n,m,3) and range 0-255
+    yuv = cv2.cvtColor(im, cv2.COLOR_BGR2YUV if bgr else cv2.COLOR_RGB2YUV)
+    if clahe:
+        c = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8, 8))
+        yuv[:, :, 0] = c.apply(yuv[:, :, 0])
+    else:
+        yuv[:, :, 0] = cv2.equalizeHist(yuv[:, :, 0])  # equalize Y channel histogram
+    return cv2.cvtColor(yuv, cv2.COLOR_YUV2BGR if bgr else cv2.COLOR_YUV2RGB)  # convert YUV image to RGB
+
+
+def replicate(im, labels):
+    # Replicate labels
+    h, w = im.shape[:2]
+    boxes = labels[:, 1:].astype(int)
+    x1, y1, x2, y2 = boxes.T
+    s = ((x2 - x1) + (y2 - y1)) / 2  # side length (pixels)
+    for i in s.argsort()[:round(s.size * 0.5)]:  # smallest indices
+        x1b, y1b, x2b, y2b = boxes[i]
+        bh, bw = y2b - y1b, x2b - x1b
+        yc, xc = int(random.uniform(0, h - bh)), int(random.uniform(0, w - bw))  # offset x, y
+        x1a, y1a, x2a, y2a = [xc, yc, xc + bw, yc + bh]
+        im[y1a:y2a, x1a:x2a] = im[y1b:y2b, x1b:x2b]  # im4[ymin:ymax, xmin:xmax]
+        labels = np.append(labels, [[labels[i, 0], x1a, y1a, x2a, y2a]], axis=0)
+
+    return im, labels
+
+
+def letterbox(im, new_shape=(640, 640), color=(114, 114, 114), auto=True, scaleFill=False, scaleup=True, stride=32):
+    # Resize and pad image while meeting stride-multiple constraints
+    shape = im.shape[:2]  # current shape [height, width]
+    if isinstance(new_shape, int):
+        new_shape = (new_shape, new_shape)
+
+    # Scale ratio (new / old)
+    r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])
+    if not scaleup:  # only scale down, do not scale up (for better val mAP)
+        r = min(r, 1.0)
+
+    # Compute padding
+    ratio = r, r  # width, height ratios
+    new_unpad = int(round(shape[1] * r)), int(round(shape[0] * r))
+    dw, dh = new_shape[1] - new_unpad[0], new_shape[0] - new_unpad[1]  # wh padding
+    if auto:  # minimum rectangle
+        dw, dh = np.mod(dw, stride), np.mod(dh, stride)  # wh padding
+    elif scaleFill:  # stretch
+        dw, dh = 0.0, 0.0
+        new_unpad = (new_shape[1], new_shape[0])
+        ratio = new_shape[1] / shape[1], new_shape[0] / shape[0]  # width, height ratios
+
+    dw /= 2  # divide padding into 2 sides
+    dh /= 2
+
+    if shape[::-1] != new_unpad:  # resize
+        im = cv2.resize(im, new_unpad, interpolation=cv2.INTER_LINEAR)
+    top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1))
+    left, right = int(round(dw - 0.1)), int(round(dw + 0.1))
+    im = cv2.copyMakeBorder(im, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color)  # add border
+    return im, ratio, (dw, dh)
+
+
+def random_perspective(im,
+                       targets=(),
+                       segments=(),
+                       degrees=10,
+                       translate=.1,
+                       scale=.1,
+                       shear=10,
+                       perspective=0.0,
+                       border=(0, 0)):
+    # torchvision.transforms.RandomAffine(degrees=(-10, 10), translate=(0.1, 0.1), scale=(0.9, 1.1), shear=(-10, 10))
+    # targets = [cls, xyxy]
+
+    height = im.shape[0] + border[0] * 2  # shape(h,w,c)
+    width = im.shape[1] + border[1] * 2
+
+    # Center
+    C = np.eye(3)
+    C[0, 2] = -im.shape[1] / 2  # x translation (pixels)
+    C[1, 2] = -im.shape[0] / 2  # y translation (pixels)
+
+    # Perspective
+    P = np.eye(3)
+    P[2, 0] = random.uniform(-perspective, perspective)  # x perspective (about y)
+    P[2, 1] = random.uniform(-perspective, perspective)  # y perspective (about x)
+
+    # Rotation and Scale
+    R = np.eye(3)
+    a = random.uniform(-degrees, degrees)
+    # a += random.choice([-180, -90, 0, 90])  # add 90deg rotations to small rotations
+    s = random.uniform(1 - scale, 1 + scale)
+    # s = 2 ** random.uniform(-scale, scale)
+    R[:2] = cv2.getRotationMatrix2D(angle=a, center=(0, 0), scale=s)
+
+    # Shear
+    S = np.eye(3)
+    S[0, 1] = math.tan(random.uniform(-shear, shear) * math.pi / 180)  # x shear (deg)
+    S[1, 0] = math.tan(random.uniform(-shear, shear) * math.pi / 180)  # y shear (deg)
+
+    # Translation
+    T = np.eye(3)
+    T[0, 2] = random.uniform(0.5 - translate, 0.5 + translate) * width  # x translation (pixels)
+    T[1, 2] = random.uniform(0.5 - translate, 0.5 + translate) * height  # y translation (pixels)
+
+    # Combined rotation matrix
+    M = T @ S @ R @ P @ C  # order of operations (right to left) is IMPORTANT
+    if (border[0] != 0) or (border[1] != 0) or (M != np.eye(3)).any():  # image changed
+        if perspective:
+            im = cv2.warpPerspective(im, M, dsize=(width, height), borderValue=(114, 114, 114))
+        else:  # affine
+            im = cv2.warpAffine(im, M[:2], dsize=(width, height), borderValue=(114, 114, 114))
+
+    # Visualize
+    # import matplotlib.pyplot as plt
+    # ax = plt.subplots(1, 2, figsize=(12, 6))[1].ravel()
+    # ax[0].imshow(im[:, :, ::-1])  # base
+    # ax[1].imshow(im2[:, :, ::-1])  # warped
+
+    # Transform label coordinates
+    n = len(targets)
+    if n:
+        use_segments = any(x.any() for x in segments)
+        new = np.zeros((n, 4))
+        if use_segments:  # warp segments
+            segments = resample_segments(segments)  # upsample
+            for i, segment in enumerate(segments):
+                xy = np.ones((len(segment), 3))
+                xy[:, :2] = segment
+                xy = xy @ M.T  # transform
+                xy = xy[:, :2] / xy[:, 2:3] if perspective else xy[:, :2]  # perspective rescale or affine
+
+                # clip
+                new[i] = segment2box(xy, width, height)
+
+        else:  # warp boxes
+            xy = np.ones((n * 4, 3))
+            xy[:, :2] = targets[:, [1, 2, 3, 4, 1, 4, 3, 2]].reshape(n * 4, 2)  # x1y1, x2y2, x1y2, x2y1
+            xy = xy @ M.T  # transform
+            xy = (xy[:, :2] / xy[:, 2:3] if perspective else xy[:, :2]).reshape(n, 8)  # perspective rescale or affine
+
+            # create new boxes
+            x = xy[:, [0, 2, 4, 6]]
+            y = xy[:, [1, 3, 5, 7]]
+            new = np.concatenate((x.min(1), y.min(1), x.max(1), y.max(1))).reshape(4, n).T
+
+            # clip
+            new[:, [0, 2]] = new[:, [0, 2]].clip(0, width)
+            new[:, [1, 3]] = new[:, [1, 3]].clip(0, height)
+
+        # filter candidates
+        i = box_candidates(box1=targets[:, 1:5].T * s, box2=new.T, area_thr=0.01 if use_segments else 0.10)
+        targets = targets[i]
+        targets[:, 1:5] = new[i]
+
+    return im, targets
+
+
+def copy_paste(im, labels, segments, p=0.5):
+    # Implement Copy-Paste augmentation https://arxiv.org/abs/2012.07177, labels as nx5 np.array(cls, xyxy)
+    n = len(segments)
+    if p and n:
+        h, w, c = im.shape  # height, width, channels
+        im_new = np.zeros(im.shape, np.uint8)
+
+        # calculate ioa first then select indexes randomly
+        boxes = np.stack([w - labels[:, 3], labels[:, 2], w - labels[:, 1], labels[:, 4]], axis=-1)  # (n, 4)
+        ioa = bbox_ioa(boxes, labels[:, 1:5])  # intersection over area
+        indexes = np.nonzero((ioa < 0.30).all(1))[0]  # (N, )
+        n = len(indexes)
+        for j in random.sample(list(indexes), k=round(p * n)):
+            l, box, s = labels[j], boxes[j], segments[j]
+            labels = np.concatenate((labels, [[l[0], *box]]), 0)
+            segments.append(np.concatenate((w - s[:, 0:1], s[:, 1:2]), 1))
+            cv2.drawContours(im_new, [segments[j].astype(np.int32)], -1, (1, 1, 1), cv2.FILLED)
+
+        result = cv2.flip(im, 1)  # augment segments (flip left-right)
+        i = cv2.flip(im_new, 1).astype(bool)
+        im[i] = result[i]  # cv2.imwrite('debug.jpg', im)  # debug
+
+    return im, labels, segments
+
+
+def cutout(im, labels, p=0.5):
+    # Applies image cutout augmentation https://arxiv.org/abs/1708.04552
+    if random.random() < p:
+        h, w = im.shape[:2]
+        scales = [0.5] * 1 + [0.25] * 2 + [0.125] * 4 + [0.0625] * 8 + [0.03125] * 16  # image size fraction
+        for s in scales:
+            mask_h = random.randint(1, int(h * s))  # create random masks
+            mask_w = random.randint(1, int(w * s))
+
+            # box
+            xmin = max(0, random.randint(0, w) - mask_w // 2)
+            ymin = max(0, random.randint(0, h) - mask_h // 2)
+            xmax = min(w, xmin + mask_w)
+            ymax = min(h, ymin + mask_h)
+
+            # apply random color mask
+            im[ymin:ymax, xmin:xmax] = [random.randint(64, 191) for _ in range(3)]
+
+            # return unobscured labels
+            if len(labels) and s > 0.03:
+                box = np.array([[xmin, ymin, xmax, ymax]], dtype=np.float32)
+                ioa = bbox_ioa(box, xywhn2xyxy(labels[:, 1:5], w, h))[0]  # intersection over area
+                labels = labels[ioa < 0.60]  # remove >60% obscured labels
+
+    return labels
+
+
+def mixup(im, labels, im2, labels2):
+    # Applies MixUp augmentation https://arxiv.org/pdf/1710.09412.pdf
+    r = np.random.beta(32.0, 32.0)  # mixup ratio, alpha=beta=32.0
+    im = (im * r + im2 * (1 - r)).astype(np.uint8)
+    labels = np.concatenate((labels, labels2), 0)
+    return im, labels
+
+
+def box_candidates(box1, box2, wh_thr=2, ar_thr=100, area_thr=0.1, eps=1e-16):  # box1(4,n), box2(4,n)
+    # Compute candidate boxes: box1 before augment, box2 after augment, wh_thr (pixels), aspect_ratio_thr, area_ratio
+    w1, h1 = box1[2] - box1[0], box1[3] - box1[1]
+    w2, h2 = box2[2] - box2[0], box2[3] - box2[1]
+    ar = np.maximum(w2 / (h2 + eps), h2 / (w2 + eps))  # aspect ratio
+    return (w2 > wh_thr) & (h2 > wh_thr) & (w2 * h2 / (w1 * h1 + eps) > area_thr) & (ar < ar_thr)  # candidates
+
+
+def classify_albumentations(
+        augment=True,
+        size=224,
+        scale=(0.08, 1.0),
+        ratio=(0.75, 1.0 / 0.75),  # 0.75, 1.33
+        hflip=0.5,
+        vflip=0.0,
+        jitter=0.4,
+        mean=IMAGENET_MEAN,
+        std=IMAGENET_STD,
+        auto_aug=False):
+    # YOLOv5 classification Albumentations (optional, only used if package is installed)
+    prefix = colorstr('albumentations: ')
+    try:
+        import albumentations as A
+        from albumentations.pytorch import ToTensorV2
+        check_version(A.__version__, '1.0.3', hard=True)  # version requirement
+        if augment:  # Resize and crop
+            T = [A.RandomResizedCrop(height=size, width=size, scale=scale, ratio=ratio)]
+            if auto_aug:
+                # TODO: implement AugMix, AutoAug & RandAug in albumentation
+                LOGGER.info(f'{prefix}auto augmentations are currently not supported')
+            else:
+                if hflip > 0:
+                    T += [A.HorizontalFlip(p=hflip)]
+                if vflip > 0:
+                    T += [A.VerticalFlip(p=vflip)]
+                if jitter > 0:
+                    color_jitter = (float(jitter),) * 3  # repeat value for brightness, contrast, satuaration, 0 hue
+                    T += [A.ColorJitter(*color_jitter, 0)]
+        else:  # Use fixed crop for eval set (reproducibility)
+            T = [A.SmallestMaxSize(max_size=size), A.CenterCrop(height=size, width=size)]
+        T += [A.Normalize(mean=mean, std=std), ToTensorV2()]  # Normalize and convert to Tensor
+        LOGGER.info(prefix + ', '.join(f'{x}'.replace('always_apply=False, ', '') for x in T if x.p))
+        return A.Compose(T)
+
+    except ImportError:  # package not installed, skip
+        LOGGER.warning(f'{prefix}⚠️ not found, install with `pip install albumentations` (recommended)')
+    except Exception as e:
+        LOGGER.info(f'{prefix}{e}')
+
+
+def classify_transforms(size=224):
+    # Transforms to apply if albumentations not installed
+    assert isinstance(size, int), f'ERROR: classify_transforms size {size} must be integer, not (list, tuple)'
+    # T.Compose([T.ToTensor(), T.Resize(size), T.CenterCrop(size), T.Normalize(IMAGENET_MEAN, IMAGENET_STD)])
+    return T.Compose([CenterCrop(size), ToTensor(), T.Normalize(IMAGENET_MEAN, IMAGENET_STD)])
+
+
+class LetterBox:
+    # YOLOv5 LetterBox class for image preprocessing, i.e. T.Compose([LetterBox(size), ToTensor()])
+    def __init__(self, size=(640, 640), auto=False, stride=32):
+        super().__init__()
+        self.h, self.w = (size, size) if isinstance(size, int) else size
+        self.auto = auto  # pass max size integer, automatically solve for short side using stride
+        self.stride = stride  # used with auto
+
+    def __call__(self, im):  # im = np.array HWC
+        imh, imw = im.shape[:2]
+        r = min(self.h / imh, self.w / imw)  # ratio of new/old
+        h, w = round(imh * r), round(imw * r)  # resized image
+        hs, ws = (math.ceil(x / self.stride) * self.stride for x in (h, w)) if self.auto else self.h, self.w
+        top, left = round((hs - h) / 2 - 0.1), round((ws - w) / 2 - 0.1)
+        im_out = np.full((self.h, self.w, 3), 114, dtype=im.dtype)
+        im_out[top:top + h, left:left + w] = cv2.resize(im, (w, h), interpolation=cv2.INTER_LINEAR)
+        return im_out
+
+
+class CenterCrop:
+    # YOLOv5 CenterCrop class for image preprocessing, i.e. T.Compose([CenterCrop(size), ToTensor()])
+    def __init__(self, size=640):
+        super().__init__()
+        self.h, self.w = (size, size) if isinstance(size, int) else size
+
+    def __call__(self, im):  # im = np.array HWC
+        imh, imw = im.shape[:2]
+        m = min(imh, imw)  # min dimension
+        top, left = (imh - m) // 2, (imw - m) // 2
+        return cv2.resize(im[top:top + m, left:left + m], (self.w, self.h), interpolation=cv2.INTER_LINEAR)
+
+
+class ToTensor:
+    # YOLOv5 ToTensor class for image preprocessing, i.e. T.Compose([LetterBox(size), ToTensor()])
+    def __init__(self, half=False):
+        super().__init__()
+        self.half = half
+
+    def __call__(self, im):  # im = np.array HWC in BGR order
+        im = np.ascontiguousarray(im.transpose((2, 0, 1))[::-1])  # HWC to CHW -> BGR to RGB -> contiguous
+        im = torch.from_numpy(im)  # to torch
+        im = im.half() if self.half else im.float()  # uint8 to fp16/32
+        im /= 255.0  # 0-255 to 0.0-1.0
+        return im

utils/autoanchor.py

@@ -0,0 +1,164 @@
+import random
+
+import numpy as np
+import torch
+import yaml
+from tqdm import tqdm
+
+from utils import TryExcept
+from utils.general import LOGGER, TQDM_BAR_FORMAT, colorstr
+
+PREFIX = colorstr('AutoAnchor: ')
+
+
+def check_anchor_order(m):
+    # Check anchor order against stride order for YOLOv5 Detect() module m, and correct if necessary
+    a = m.anchors.prod(-1).mean(-1).view(-1)  # mean anchor area per output layer
+    da = a[-1] - a[0]  # delta a
+    ds = m.stride[-1] - m.stride[0]  # delta s
+    if da and (da.sign() != ds.sign()):  # same order
+        LOGGER.info(f'{PREFIX}Reversing anchor order')
+        m.anchors[:] = m.anchors.flip(0)
+
+
+@TryExcept(f'{PREFIX}ERROR')
+def check_anchors(dataset, model, thr=4.0, imgsz=640):
+    # Check anchor fit to data, recompute if necessary
+    m = model.module.model[-1] if hasattr(model, 'module') else model.model[-1]  # Detect()
+    shapes = imgsz * dataset.shapes / dataset.shapes.max(1, keepdims=True)
+    scale = np.random.uniform(0.9, 1.1, size=(shapes.shape[0], 1))  # augment scale
+    wh = torch.tensor(np.concatenate([l[:, 3:5] * s for s, l in zip(shapes * scale, dataset.labels)])).float()  # wh
+
+    def metric(k):  # compute metric
+        r = wh[:, None] / k[None]
+        x = torch.min(r, 1 / r).min(2)[0]  # ratio metric
+        best = x.max(1)[0]  # best_x
+        aat = (x > 1 / thr).float().sum(1).mean()  # anchors above threshold
+        bpr = (best > 1 / thr).float().mean()  # best possible recall
+        return bpr, aat
+
+    stride = m.stride.to(m.anchors.device).view(-1, 1, 1)  # model strides
+    anchors = m.anchors.clone() * stride  # current anchors
+    bpr, aat = metric(anchors.cpu().view(-1, 2))
+    s = f'\n{PREFIX}{aat:.2f} anchors/target, {bpr:.3f} Best Possible Recall (BPR). '
+    if bpr > 0.98:  # threshold to recompute
+        LOGGER.info(f'{s}Current anchors are a good fit to dataset ✅')
+    else:
+        LOGGER.info(f'{s}Anchors are a poor fit to dataset ⚠️, attempting to improve...')
+        na = m.anchors.numel() // 2  # number of anchors
+        anchors = kmean_anchors(dataset, n=na, img_size=imgsz, thr=thr, gen=1000, verbose=False)
+        new_bpr = metric(anchors)[0]
+        if new_bpr > bpr:  # replace anchors
+            anchors = torch.tensor(anchors, device=m.anchors.device).type_as(m.anchors)
+            m.anchors[:] = anchors.clone().view_as(m.anchors)
+            check_anchor_order(m)  # must be in pixel-space (not grid-space)
+            m.anchors /= stride
+            s = f'{PREFIX}Done ✅ (optional: update model *.yaml to use these anchors in the future)'
+        else:
+            s = f'{PREFIX}Done ⚠️ (original anchors better than new anchors, proceeding with original anchors)'
+        LOGGER.info(s)
+
+
+def kmean_anchors(dataset='./data/coco128.yaml', n=9, img_size=640, thr=4.0, gen=1000, verbose=True):
+    """ Creates kmeans-evolved anchors from training dataset
+
+        Arguments:
+            dataset: path to data.yaml, or a loaded dataset
+            n: number of anchors
+            img_size: image size used for training
+            thr: anchor-label wh ratio threshold hyperparameter hyp['anchor_t'] used for training, default=4.0
+            gen: generations to evolve anchors using genetic algorithm
+            verbose: print all results
+
+        Return:
+            k: kmeans evolved anchors
+
+        Usage:
+            from utils.autoanchor import *; _ = kmean_anchors()
+    """
+    from scipy.cluster.vq import kmeans
+
+    npr = np.random
+    thr = 1 / thr
+
+    def metric(k, wh):  # compute metrics
+        r = wh[:, None] / k[None]
+        x = torch.min(r, 1 / r).min(2)[0]  # ratio metric
+        # x = wh_iou(wh, torch.tensor(k))  # iou metric
+        return x, x.max(1)[0]  # x, best_x
+
+    def anchor_fitness(k):  # mutation fitness
+        _, best = metric(torch.tensor(k, dtype=torch.float32), wh)
+        return (best * (best > thr).float()).mean()  # fitness
+
+    def print_results(k, verbose=True):
+        k = k[np.argsort(k.prod(1))]  # sort small to large
+        x, best = metric(k, wh0)
+        bpr, aat = (best > thr).float().mean(), (x > thr).float().mean() * n  # best possible recall, anch > thr
+        s = f'{PREFIX}thr={thr:.2f}: {bpr:.4f} best possible recall, {aat:.2f} anchors past thr\n' \
+            f'{PREFIX}n={n}, img_size={img_size}, metric_all={x.mean():.3f}/{best.mean():.3f}-mean/best, ' \
+            f'past_thr={x[x > thr].mean():.3f}-mean: '
+        for x in k:
+            s += '%i,%i, ' % (round(x[0]), round(x[1]))
+        if verbose:
+            LOGGER.info(s[:-2])
+        return k
+
+    if isinstance(dataset, str):  # *.yaml file
+        with open(dataset, errors='ignore') as f:
+            data_dict = yaml.safe_load(f)  # model dict
+        from utils.dataloaders import LoadImagesAndLabels
+        dataset = LoadImagesAndLabels(data_dict['train'], augment=True, rect=True)
+
+    # Get label wh
+    shapes = img_size * dataset.shapes / dataset.shapes.max(1, keepdims=True)
+    wh0 = np.concatenate([l[:, 3:5] * s for s, l in zip(shapes, dataset.labels)])  # wh
+
+    # Filter
+    i = (wh0 < 3.0).any(1).sum()
+    if i:
+        LOGGER.info(f'{PREFIX}WARNING ⚠️ Extremely small objects found: {i} of {len(wh0)} labels are <3 pixels in size')
+    wh = wh0[(wh0 >= 2.0).any(1)].astype(np.float32)  # filter > 2 pixels
+    # wh = wh * (npr.rand(wh.shape[0], 1) * 0.9 + 0.1)  # multiply by random scale 0-1
+
+    # Kmeans init
+    try:
+        LOGGER.info(f'{PREFIX}Running kmeans for {n} anchors on {len(wh)} points...')
+        assert n <= len(wh)  # apply overdetermined constraint
+        s = wh.std(0)  # sigmas for whitening
+        k = kmeans(wh / s, n, iter=30)[0] * s  # points
+        assert n == len(k)  # kmeans may return fewer points than requested if wh is insufficient or too similar
+    except Exception:
+        LOGGER.warning(f'{PREFIX}WARNING ⚠️ switching strategies from kmeans to random init')
+        k = np.sort(npr.rand(n * 2)).reshape(n, 2) * img_size  # random init
+    wh, wh0 = (torch.tensor(x, dtype=torch.float32) for x in (wh, wh0))
+    k = print_results(k, verbose=False)
+
+    # Plot
+    # k, d = [None] * 20, [None] * 20
+    # for i in tqdm(range(1, 21)):
+    #     k[i-1], d[i-1] = kmeans(wh / s, i)  # points, mean distance
+    # fig, ax = plt.subplots(1, 2, figsize=(14, 7), tight_layout=True)
+    # ax = ax.ravel()
+    # ax[0].plot(np.arange(1, 21), np.array(d) ** 2, marker='.')
+    # fig, ax = plt.subplots(1, 2, figsize=(14, 7))  # plot wh
+    # ax[0].hist(wh[wh[:, 0]<100, 0],400)
+    # ax[1].hist(wh[wh[:, 1]<100, 1],400)
+    # fig.savefig('wh.png', dpi=200)
+
+    # Evolve
+    f, sh, mp, s = anchor_fitness(k), k.shape, 0.9, 0.1  # fitness, generations, mutation prob, sigma
+    pbar = tqdm(range(gen), bar_format=TQDM_BAR_FORMAT)  # progress bar
+    for _ in pbar:
+        v = np.ones(sh)
+        while (v == 1).all():  # mutate until a change occurs (prevent duplicates)
+            v = ((npr.random(sh) < mp) * random.random() * npr.randn(*sh) * s + 1).clip(0.3, 3.0)
+        kg = (k.copy() * v).clip(min=2.0)
+        fg = anchor_fitness(kg)
+        if fg > f:
+            f, k = fg, kg.copy()
+            pbar.desc = f'{PREFIX}Evolving anchors with Genetic Algorithm: fitness = {f:.4f}'
+            if verbose:
+                print_results(k, verbose)
+
+    return print_results(k).astype(np.float32)

utils/autobatch.py

@@ -0,0 +1,67 @@
+from copy import deepcopy
+
+import numpy as np
+import torch
+
+from utils.general import LOGGER, colorstr
+from utils.torch_utils import profile
+
+
+def check_train_batch_size(model, imgsz=640, amp=True):
+    # Check YOLOv5 training batch size
+    with torch.cuda.amp.autocast(amp):
+        return autobatch(deepcopy(model).train(), imgsz)  # compute optimal batch size
+
+
+def autobatch(model, imgsz=640, fraction=0.8, batch_size=16):
+    # Automatically estimate best YOLOv5 batch size to use `fraction` of available CUDA memory
+    # Usage:
+    #     import torch
+    #     from utils.autobatch import autobatch
+    #     model = torch.hub.load('ultralytics/yolov5', 'yolov5s', autoshape=False)
+    #     print(autobatch(model))
+
+    # Check device
+    prefix = colorstr('AutoBatch: ')
+    LOGGER.info(f'{prefix}Computing optimal batch size for --imgsz {imgsz}')
+    device = next(model.parameters()).device  # get model device
+    if device.type == 'cpu':
+        LOGGER.info(f'{prefix}CUDA not detected, using default CPU batch-size {batch_size}')
+        return batch_size
+    if torch.backends.cudnn.benchmark:
+        LOGGER.info(f'{prefix} ⚠️ Requires torch.backends.cudnn.benchmark=False, using default batch-size {batch_size}')
+        return batch_size
+
+    # Inspect CUDA memory
+    gb = 1 << 30  # bytes to GiB (1024 ** 3)
+    d = str(device).upper()  # 'CUDA:0'
+    properties = torch.cuda.get_device_properties(device)  # device properties
+    t = properties.total_memory / gb  # GiB total
+    r = torch.cuda.memory_reserved(device) / gb  # GiB reserved
+    a = torch.cuda.memory_allocated(device) / gb  # GiB allocated
+    f = t - (r + a)  # GiB free
+    LOGGER.info(f'{prefix}{d} ({properties.name}) {t:.2f}G total, {r:.2f}G reserved, {a:.2f}G allocated, {f:.2f}G free')
+
+    # Profile batch sizes
+    batch_sizes = [1, 2, 4, 8, 16]
+    try:
+        img = [torch.empty(b, 3, imgsz, imgsz) for b in batch_sizes]
+        results = profile(img, model, n=3, device=device)
+    except Exception as e:
+        LOGGER.warning(f'{prefix}{e}')
+
+    # Fit a solution
+    y = [x[2] for x in results if x]  # memory [2]
+    p = np.polyfit(batch_sizes[:len(y)], y, deg=1)  # first degree polynomial fit
+    b = int((f * fraction - p[1]) / p[0])  # y intercept (optimal batch size)
+    if None in results:  # some sizes failed
+        i = results.index(None)  # first fail index
+        if b >= batch_sizes[i]:  # y intercept above failure point
+            b = batch_sizes[max(i - 1, 0)]  # select prior safe point
+    if b < 1 or b > 1024:  # b outside of safe range
+        b = batch_size
+        LOGGER.warning(f'{prefix}WARNING ⚠️ CUDA anomaly detected, recommend restart environment and retry command.')
+
+    fraction = (np.polyval(p, b) + r + a) / t  # actual fraction predicted
+    LOGGER.info(f'{prefix}Using batch-size {b} for {d} {t * fraction:.2f}G/{t:.2f}G ({fraction * 100:.0f}%) ✅')
+    return b