Spaces:
Runtime error
Runtime error
Patrick Ramos
commited on
Commit
•
5345282
1
Parent(s):
461e0e8
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,130 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
from transformers import ViTFeatureExtractor, ViTModel
|
3 |
+
import torch
|
4 |
+
import matplotlib.pyplot as plt
|
5 |
+
from skops import hub_utils
|
6 |
+
from einops import reduce
|
7 |
+
import seaborn as sns
|
8 |
+
import pickle
|
9 |
+
|
10 |
+
labels = [
|
11 |
+
'tench',
|
12 |
+
'English springer',
|
13 |
+
'cassette player',
|
14 |
+
'chain saw',
|
15 |
+
'church',
|
16 |
+
'French horn',
|
17 |
+
'garbage truck',
|
18 |
+
'gas pump',
|
19 |
+
'golf ball',
|
20 |
+
'parachute'
|
21 |
+
]
|
22 |
+
|
23 |
+
# load DINO
|
24 |
+
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
25 |
+
feature_extractor = ViTFeatureExtractor.from_pretrained('facebook/dino-vitb16')
|
26 |
+
model = ViTModel.from_pretrained('facebook/dino-vitb16').eval().to(device)
|
27 |
+
|
28 |
+
# load logistic regression
|
29 |
+
!mkdir emb-gam-dino
|
30 |
+
hub_utils.download(repo_id='Ramos-Ramos/emb-gam-dino', dst='emb-gam-dino')
|
31 |
+
|
32 |
+
with open('emb-gam-dino/model.pkl', 'rb') as file:
|
33 |
+
logistic_regression = pickle.load(file)
|
34 |
+
|
35 |
+
def classify_and_heatmap(input_img):
|
36 |
+
# get patch embeddings
|
37 |
+
inputs = {k: v.to(device) for k, v in feature_extractor(input_img, return_tensors='pt').items()}
|
38 |
+
with torch.no_grad():
|
39 |
+
patch_embeddings = model(**inputs).last_hidden_state[0, 1:].cpu()
|
40 |
+
|
41 |
+
# get scores
|
42 |
+
scores = dict(zip(
|
43 |
+
labels,
|
44 |
+
logistic_regression.predict_proba(reduce(patch_embeddings, 'p d -> () d', 'sum'))[0]
|
45 |
+
))
|
46 |
+
|
47 |
+
# make plot
|
48 |
+
num_patches_side = model.config.image_size // model.config.patch_size
|
49 |
+
|
50 |
+
# set up figure
|
51 |
+
fig, axs = plt.subplots(2, 6, figsize=(12, 5))
|
52 |
+
gs = axs[0, 0].get_gridspec()
|
53 |
+
for ax in axs[:, 0]:
|
54 |
+
ax.remove()
|
55 |
+
ax_orig_img = fig.add_subplot(gs[:, 0])
|
56 |
+
|
57 |
+
# plot original image
|
58 |
+
img = feature_extractor.to_pil_image(
|
59 |
+
inputs['pixel_values'].squeeze(0) * torch.tensor(feature_extractor.image_std).view(-1, 1, 1) + torch.tensor(feature_extractor.image_mean).view(-1, 1, 1)
|
60 |
+
)
|
61 |
+
ax_orig_img.imshow(img)
|
62 |
+
ax_orig_img.axis('off')
|
63 |
+
|
64 |
+
# plot patch contributions
|
65 |
+
patch_contributions = logistic_regression.coef_ @ patch_embeddings.T.numpy()
|
66 |
+
vmin = patch_contributions.min()
|
67 |
+
vmax = patch_contributions.max()
|
68 |
+
|
69 |
+
# print(len(list(axs[:, 1:].flat)))
|
70 |
+
for i, ax in enumerate(axs[:, 1:].flat):
|
71 |
+
sns.heatmap(
|
72 |
+
patch_contributions[i].reshape(num_patches_side, num_patches_side),
|
73 |
+
ax=ax,
|
74 |
+
square=True,
|
75 |
+
vmin=vmin,
|
76 |
+
vmax=vmax,
|
77 |
+
)
|
78 |
+
ax.set_title(labels[i])
|
79 |
+
ax.set_xlabel(f'score={patch_contributions[i].sum():.2f}')
|
80 |
+
ax.set_xticks([])
|
81 |
+
ax.set_yticks([])
|
82 |
+
|
83 |
+
return scores, plt
|
84 |
+
|
85 |
+
description='''
|
86 |
+
This demo is a simple extension of [Emb-GAM (Singh & Gao, 2022)](https://arxiv.org/abs/2209.11799) to images. It does image classification on [Imagenette](https://github.com/fastai/imagenette) and visualizes the contrbutions of each image patch to each label.
|
87 |
+
'''
|
88 |
+
|
89 |
+
article='''
|
90 |
+
Under the hood, we use [DINO](https://arxiv.org/abs/2104.14294) to extract patch embeddings and a logistic regression model following the set up of the [offical Emb-GAM implementation](https://github.com/csinva/emb-gam).
|
91 |
+
|
92 |
+
Citation for stuff involved (not our papers):
|
93 |
+
```bibtex
|
94 |
+
@article{singh2022emb,
|
95 |
+
title={Emb-GAM: an Interpretable and Efficient Predictor using Pre-trained Language Models},
|
96 |
+
author={Singh, Chandan and Gao, Jianfeng},
|
97 |
+
journal={arXiv preprint arXiv:2209.11799},
|
98 |
+
year={2022}
|
99 |
+
}
|
100 |
+
|
101 |
+
@InProceedings{Caron_2021_ICCV,
|
102 |
+
author = {Caron, Mathilde and Touvron, Hugo and Misra, Ishan and J\'egou, Herv\'e and Mairal, Julien and Bojanowski, Piotr and Joulin, Armand},
|
103 |
+
title = {Emerging Properties in Self-Supervised Vision Transformers},
|
104 |
+
booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
|
105 |
+
month = {October},
|
106 |
+
year = {2021},
|
107 |
+
pages = {9650-9660}
|
108 |
+
}
|
109 |
+
|
110 |
+
@misc{imagenette,
|
111 |
+
author = {fast.ai},
|
112 |
+
title = {Imagenette},
|
113 |
+
url = {https://github.com/fastai/imagenette},
|
114 |
+
}
|
115 |
+
```
|
116 |
+
'''
|
117 |
+
|
118 |
+
demo = gr.Interface(
|
119 |
+
fn=classify_and_heatmap,
|
120 |
+
inputs=gr.Image(shape=(224, 224), type='pil', label='Input Image'),
|
121 |
+
outputs=[
|
122 |
+
gr.Label(label='Class'),
|
123 |
+
gr.Plot(label='Patch Contributions')
|
124 |
+
],
|
125 |
+
title='Emb-GAM DINO',
|
126 |
+
description=description,
|
127 |
+
article=article,
|
128 |
+
examples=['./examples/english_springer.png', './examples/golf_ball.png']
|
129 |
+
)
|
130 |
+
demo.launch(debug=True)
|