F-G Fernandez
commited on
Commit
·
fc36b00
1
Parent(s):
a4f6936
fix: Fixed resizing and switched to ONNX
Browse files- app.py +53 -21
- requirements.txt +5 -2
app.py
CHANGED
@@ -1,39 +1,71 @@
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import argparse
|
|
|
2 |
|
3 |
import gradio as gr
|
4 |
-
import
|
|
|
|
|
5 |
from PIL import Image
|
6 |
-
from torchvision.transforms import Compose, ConvertImageDtype, Normalize, PILToTensor, Resize
|
7 |
-
from torchvision.transforms.functional import InterpolationMode
|
8 |
|
9 |
-
from holocron import models
|
10 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
11 |
|
12 |
-
|
|
|
|
|
|
|
13 |
|
14 |
-
|
15 |
-
|
16 |
-
|
17 |
-
|
18 |
-
|
19 |
-
])
|
20 |
|
21 |
-
|
22 |
-
img = Image.fromarray(img.astype('uint8'), 'RGB')
|
23 |
-
img = preprocessor(img)
|
24 |
-
with torch.inference_mode():
|
25 |
-
prediction = torch.nn.functional.softmax(model(img.unsqueeze(0))[0], dim=0)
|
26 |
-
return {class_name: float(conf) for class_name, conf in zip(model.default_cfg['classes'], prediction)}
|
27 |
|
28 |
-
|
29 |
outputs = gr.outputs.Label(num_top_classes=3)
|
30 |
|
31 |
gr.Interface(
|
32 |
fn=predict,
|
33 |
-
inputs=[
|
34 |
outputs=outputs,
|
35 |
title="Holocron: image classification demo",
|
36 |
-
article=(
|
37 |
-
|
|
|
|
|
|
|
38 |
live=True,
|
39 |
).launch()
|
|
|
1 |
+
# Copyright (C) 2022, François-Guillaume Fernandez.
|
2 |
+
|
3 |
+
# This program is licensed under the Apache License 2.0.
|
4 |
+
# See LICENSE or go to <https://www.apache.org/licenses/LICENSE-2.0> for full license details.
|
5 |
+
|
6 |
import argparse
|
7 |
+
import json
|
8 |
|
9 |
import gradio as gr
|
10 |
+
import numpy as np
|
11 |
+
import onnxruntime
|
12 |
+
from huggingface_hub import hf_hub_download
|
13 |
from PIL import Image
|
|
|
|
|
14 |
|
|
|
15 |
|
16 |
+
REPO = "frgfm/rexnet1_0x"
|
17 |
+
|
18 |
+
# Download model config & checkpoint
|
19 |
+
with open(hf_hub_download(args.repo, filename="config.json"), "rb") as f:
|
20 |
+
cfg = json.load(f)
|
21 |
+
|
22 |
+
ort_session = onnxruntime.InferenceSession(hf_hub_download(args.repo, filename="model.onnx"))
|
23 |
+
|
24 |
+
def preprocess_image(pil_img: Image.Image) -> np.ndarray:
|
25 |
+
"""Preprocess an image for inference
|
26 |
+
|
27 |
+
Args:
|
28 |
+
pil_img: a valid pillow image
|
29 |
+
|
30 |
+
Returns:
|
31 |
+
the resized and normalized image of shape (1, C, H, W)
|
32 |
+
"""
|
33 |
+
|
34 |
+
# Resizing (PIL takes (W, H) order for resizing)
|
35 |
+
img = pil_img.resize(cfg["input_shape"][-2:][::-1], Image.BILINEAR)
|
36 |
+
# (H, W, C) --> (C, H, W)
|
37 |
+
img = np.asarray(img).transpose((2, 0, 1)).astype(np.float32) / 255
|
38 |
+
# Normalization
|
39 |
+
img -= np.array(cfg["mean"])[:, None, None]
|
40 |
+
img /= np.array(cfg["std"])[:, None, None]
|
41 |
+
|
42 |
+
return img[None, ...]
|
43 |
|
44 |
+
def predict(image):
|
45 |
+
# Preprocessing
|
46 |
+
np_img = preprocess_image(image)
|
47 |
+
ort_input = {ort_session.get_inputs()[0].name: np_img}
|
48 |
|
49 |
+
# Inference
|
50 |
+
ort_out = ort_session.run(None, ort_input)
|
51 |
+
# Post-processing
|
52 |
+
out_exp = np.exp(ort_out[0][0])
|
53 |
+
probs = out_exp / out_exp.sum()
|
|
|
54 |
|
55 |
+
return {class_name: float(conf) for class_name, conf in zip(cfg["classes"], probs)}
|
|
|
|
|
|
|
|
|
|
|
56 |
|
57 |
+
img = gr.inputs.Image(type="pil")
|
58 |
outputs = gr.outputs.Label(num_top_classes=3)
|
59 |
|
60 |
gr.Interface(
|
61 |
fn=predict,
|
62 |
+
inputs=[img],
|
63 |
outputs=outputs,
|
64 |
title="Holocron: image classification demo",
|
65 |
+
article=(
|
66 |
+
"<p style='text-align: center'><a href='https://github.com/frgfm/Holocron'>"
|
67 |
+
"Github Repo</a> | "
|
68 |
+
"<a href='https://frgfm.github.io/Holocron/'>Documentation</a></p>"
|
69 |
+
),
|
70 |
live=True,
|
71 |
).launch()
|
requirements.txt
CHANGED
@@ -1,2 +1,5 @@
|
|
1 |
-
|
2 |
-
|
|
|
|
|
|
|
|
1 |
+
gradio>=3.0.2,<4.0.0
|
2 |
+
Pillow>=8.4.0
|
3 |
+
onnxruntime>=1.10.0,<2.0.0
|
4 |
+
huggingface-hub>=0.4.0,<1.0.0
|
5 |
+
numpy>=1.19.5,<2.0.0
|