init demo

app.py

@@ -0,0 +1,174 @@
+import os
+from io import BytesIO
+from pathlib import Path
+from random import shuffle
+
+import cv2
+import gradio as gr
+import matplotlib.pyplot as plt
+import numpy as np
+import torch
+from mini_resnet import CustomResNet
+from PIL import Image
+from pytorch_grad_cam import GradCAM
+from pytorch_grad_cam.utils.image import show_cam_on_image
+from torchvision import transforms as T
+
+mean = (0.49139968, 0.48215841, 0.44653091)
+std = (0.24703223, 0.24348513, 0.26158784)
+transforms = T.Compose([T.ToTensor(), T.Normalize(mean=mean, std=std)])
+classes = ("plane", "car", "bird", "cat", "deer", "dog", "frog", "horse", "ship", "truck")
+softmax = torch.nn.Softmax(dim=0)
+
+model = CustomResNet()
+model.load_state_dict(torch.load("weights/weights.pt", map_location=torch.device("cpu")))
+model.eval()
+
+misclf_path = "images/miss_classified"
+mis_classified_imgs = list(Path(misclf_path).glob("*"))
+
+
+def get_traget_layer(block: str, layer: int):
+    layer_num = 0 if layer == 0 else -1
+    if block == "block1":
+        return model.layer1[layer_num]
+    if block == "block2":
+        return model.layer2[layer_num]
+    if block == "block3":
+        return model.layer3[layer_num]
+
+
+default_cam = GradCAM(model=model, target_layers=[get_traget_layer("block3", -1)])
+
+
+def make_image(p: Path | str, pred: str, label: str):
+    im = cv2.imread(str(p))
+    im = cv2.resize(im, (64, 64))
+
+    plt.imshow(im)
+    plt.title(f"{pred} / {label}")
+    plt.axis("off")
+
+    buffer = BytesIO()
+    plt.savefig(buffer, format="png")
+    buffer.seek(0)
+
+    img_array = np.frombuffer(buffer.getvalue(), dtype=np.uint8)
+    buffer.close()
+
+    # Decode the image array using OpenCV
+    im = cv2.imdecode(img_array, cv2.IMREAD_COLOR)
+    return im
+
+
+@torch.inference_mode()
+def predict_img(img: np.ndarray, top_k: int = 10):
+    preds = model(img)
+    preds = softmax(preds.flatten())
+    preds = {classes[i]: float(preds[i]) for i in range(10)}
+    preds = {
+        k: v for k, v in sorted(preds.items(), key=lambda item: item[1], reverse=True)[:top_k]
+    }
+
+    return preds
+
+
+def display_cam(cam: GradCAM, org_img: np.ndarray, img: torch.Tensor, transparency: float):
+    grayscale_cam = cam(input_tensor=img, targets=None)
+    grayscale_cam = grayscale_cam[0, :]
+    visualization = show_cam_on_image(
+        org_img / 255, grayscale_cam, use_rgb=True, image_weight=transparency
+    )
+    return visualization
+
+
+def inference(
+    org_img: np.ndarray,
+    top_k: int,
+    show_cam: str,
+    num_cam_imgs: int,
+    cam_block: str,
+    target_layer_num: int,
+    transparency: float,
+    show_misclf: str,
+    num_misclf: int,
+):
+    input_img = transforms(org_img)
+    input_img = input_img.unsqueeze(0)
+
+    preds = predict_img(input_img, top_k)
+    org_img = display_cam(default_cam, org_img, input_img, transparency)
+
+    shuffle(mis_classified_imgs)
+    cam_outputs = []
+    if show_cam:
+        img_list = []
+
+        target_layers = [get_traget_layer(cam_block, target_layer_num)]
+        cam = GradCAM(model=model, target_layers=target_layers, use_cuda=False)
+        for p in mis_classified_imgs[:num_cam_imgs]:
+            im = cv2.imread(str(p))
+            inp_im = transforms(im)
+            inp_im = inp_im.unsqueeze(0)
+
+            grayscale_cam = cam(input_tensor=inp_im, targets=None)
+
+            grayscale_cam = grayscale_cam[0, :]
+            visualization = show_cam_on_image(
+                im / 255, grayscale_cam, use_rgb=True, image_weight=transparency
+            )
+            cam_outputs.append(visualization)
+
+        del cam, img_list
+
+    misclf_images_output = []
+    if show_misclf:
+        img_list = []
+        gt = []
+        for p in mis_classified_imgs[:num_misclf]:
+            img_list.append(transforms(Image.open(p).convert("RGB")))
+            gt.append(p.name.split("_")[0])
+
+        misclf_out = softmax(model(torch.stack(img_list))).argmax(dim=1).tolist()
+        del img_list
+        for imp, pred, label in zip(mis_classified_imgs[:num_misclf], misclf_out, gt):
+            pred = classes[pred]
+            misclf_images_output.append(make_image(imp, pred, label))
+
+    return org_img, preds, cam_outputs, misclf_images_output
+
+
+title = "CIFAR10 trained on Custom Model inspired by ResNet with GradCAM"
+description = "A simple Gradio interface to infer on ResNet model, and get GradCAM results. You can see the code <a href='https://github.com/anantgupta129/TorcHood'>here</a> & <a href='https://colab.research.google.com/github/anantgupta129/ERA-V1/blob/main/session12/notebooks/s12_train.ipynb'>training notebook</a>"
+
+examples = [["images/examples/cat.jpg", 3, True, 5, "block3", 1, 0.5, True, 5], ["images/examples/dog.jpg", 5, True, 5, "block3", 1, 0.5, True, 5]]
+demo = gr.Interface(
+    inference,
+    inputs=[
+        gr.Image(shape=(32, 32), label="Input Image"),
+        gr.Slider(1, 10, value=3, step=1, label="Top K predictions"),
+        gr.Checkbox(label="Show Grad Cam"),
+        gr.Slider(1, 20, value=5, step=1, label="Number of images"),
+        gr.Radio(label="Which Block?", choices=["block1", "block2", "block3"]),
+        gr.Slider(0, 1, value=1, step=1, label="Which Layer?"),
+        gr.Slider(0, 1, value=0.5, label="Opacity of GradCAM"),
+        gr.Checkbox(label="Show Misclassified Images"),
+        gr.Slider(1, 20, value=5, step=5, label="Number of Misclassification Images"),
+    ],
+    outputs=[
+        gr.Image(shape=(32, 32), label="Output", width=128, height=128),
+        "label",
+        gr.Gallery(label="GradCAM Output"),
+        gr.Gallery(
+            label="Misclassified Images Pred/G.T.",
+            columns=[2],
+            rows=[2],
+            object_fit="contain",
+            height="auto",
+        ),
+    ],
+    title=title,
+    description=description,
+    examples=examples,
+)
+demo.launch()

mini_resnet.py

@@ -0,0 +1,89 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+# from common import BaseNet
+
+
+class ResBlock(nn.Module):
+    def __init__(self, in_planes: int, out_planes: int, stride: int = 1, drop: float = 0) -> None:
+        super().__init__()
+        self.dropout = nn.Dropout2d(drop)
+
+        self.conv1 = nn.Conv2d(
+            in_planes,
+            out_planes,
+            kernel_size=3,
+            stride=stride,
+            padding=1,
+            bias=False,
+        )
+        self.bn1 = nn.BatchNorm2d(out_planes)
+
+        self.conv2 = nn.Conv2d(
+            out_planes,
+            out_planes,
+            kernel_size=3,
+            stride=stride,
+            padding=1,
+            bias=False,
+        )
+        self.bn2 = nn.BatchNorm2d(out_planes)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        out = F.relu(self.bn1(self.conv1(x)))
+        out = self.dropout(out)
+        out = self.bn2(self.conv2(out))
+        out += x
+        out = F.relu(out)
+        out = self.dropout(out)
+
+        return out
+
+
+class CustomResNet(nn.Module):
+    def __init__(self, drop: float = 0, num_classes: int = 10) -> None:
+        super().__init__()
+
+        # perp layer
+        self.perlayer = nn.Sequential(
+            nn.Conv2d(3, 64, 3, padding=1, bias=False),
+            nn.BatchNorm2d(64),
+            nn.ReLU(),
+            nn.Dropout2d(drop),
+        )
+        self.layer1 = nn.Sequential(
+            nn.Conv2d(64, 128, 3, padding=1, bias=False),
+            nn.MaxPool2d(2, 2),
+            nn.BatchNorm2d(128),
+            nn.ReLU(),
+            nn.Dropout2d(drop),
+            ResBlock(128, 128, drop=drop),
+        )
+        self.layer2 = nn.Sequential(
+            nn.Conv2d(128, 256, 3, padding=1, bias=False),
+            nn.MaxPool2d(2, 2),
+            nn.BatchNorm2d(256),
+            nn.ReLU(),
+            nn.Dropout2d(drop),
+        )
+        self.layer3 = nn.Sequential(
+            nn.Conv2d(256, 512, 3, padding=1, bias=False),
+            nn.MaxPool2d(2, 2),
+            nn.BatchNorm2d(512),
+            nn.ReLU(),
+            nn.Dropout2d(drop),
+            ResBlock(512, 512, drop=drop),
+        )
+        self.pool = nn.MaxPool2d(4)
+        self.out = nn.Conv2d(512, num_classes, 1, bias=False)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.perlayer(x)
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.pool(x)
+        x = self.out(x)
+
+        return x.view(-1, 10)

weights/weights.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:537219db09391c540e4a68ace7708857fb1dce2f5b0e0d325eba082bafc2643e
+size 26325330