Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,108 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import numpy as np
|
3 |
+
import gradio as gr
|
4 |
+
import plotly.express as px
|
5 |
+
import plotly.graph_objects as go
|
6 |
+
from sklearn.decomposition import PCA
|
7 |
+
from torchvision import transforms as T
|
8 |
+
from sklearn.preprocessing import MinMaxScaler
|
9 |
+
|
10 |
+
|
11 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
12 |
+
|
13 |
+
dino = torch.hub.load('facebookresearch/dinov2', 'dinov2_vitb14')
|
14 |
+
dino.eval()
|
15 |
+
dino.to(device)
|
16 |
+
|
17 |
+
pca = PCA(n_components=3)
|
18 |
+
scaler = MinMaxScaler(clip=True)
|
19 |
+
|
20 |
+
def plot_img(img_array: np.array) -> go.Figure:
|
21 |
+
fig = px.imshow(img_array)
|
22 |
+
fig.update_layout(
|
23 |
+
xaxis=dict(showticklabels=False),
|
24 |
+
yaxis=dict(showticklabels=False)
|
25 |
+
)
|
26 |
+
|
27 |
+
return fig
|
28 |
+
|
29 |
+
|
30 |
+
def app_fn(
|
31 |
+
img: np.ndarray,
|
32 |
+
threshold: float,
|
33 |
+
object_larger_than_bg: bool
|
34 |
+
) -> go.Figure:
|
35 |
+
IMAGENET_DEFAULT_MEAN = (0.485, 0.456, 0.406)
|
36 |
+
IMAGENET_DEFAULT_STD = (0.229, 0.224, 0.225)
|
37 |
+
|
38 |
+
patch_h = 40
|
39 |
+
patch_w = 40
|
40 |
+
|
41 |
+
transform = T.Compose([
|
42 |
+
T.Resize((14 * patch_h, 14 * patch_w)),
|
43 |
+
T.Normalize(mean=IMAGENET_DEFAULT_MEAN, std=IMAGENET_DEFAULT_STD),
|
44 |
+
])
|
45 |
+
|
46 |
+
img = torch.from_numpy(img).type(torch.float).permute(2, 0, 1) / 255
|
47 |
+
img_tensor = transform(img).unsqueeze(0).to(device)
|
48 |
+
|
49 |
+
with torch.no_grad():
|
50 |
+
out = dino.forward_features(img_tensor)
|
51 |
+
|
52 |
+
features = out["x_prenorm"][:, 1:, :]
|
53 |
+
features = features.squeeze(0)
|
54 |
+
features = features.cpu().numpy()
|
55 |
+
|
56 |
+
pca_features = pca.fit_transform(features)
|
57 |
+
pca_features = scaler.fit_transform(pca_features)
|
58 |
+
|
59 |
+
if object_larger_than_bg:
|
60 |
+
pca_features_bg = pca_features[:, 0] > threshold
|
61 |
+
else:
|
62 |
+
pca_features_bg = pca_features[:, 0] < threshold
|
63 |
+
|
64 |
+
pca_features_fg = ~pca_features_bg
|
65 |
+
|
66 |
+
pca_features_fg_seg = pca.fit_transform(features[pca_features_fg])
|
67 |
+
|
68 |
+
pca_features_fg_seg = scaler.fit_transform(pca_features_fg_seg)
|
69 |
+
|
70 |
+
pca_features_rgb = np.zeros((patch_h * patch_w, 3))
|
71 |
+
pca_features_rgb[pca_features_bg] = 0
|
72 |
+
pca_features_rgb[pca_features_fg] = pca_features_fg_seg
|
73 |
+
pca_features_rgb = pca_features_rgb.reshape(patch_h, patch_w, 3)
|
74 |
+
|
75 |
+
|
76 |
+
fig_pca = plot_img(pca_features_rgb)
|
77 |
+
|
78 |
+
return fig_pca
|
79 |
+
|
80 |
+
if __name__=="__main__":
|
81 |
+
title = "DINOv2"
|
82 |
+
with gr.Blocks(title=title) as demo:
|
83 |
+
gr.Markdown(f"# {title}")
|
84 |
+
gr.Markdown(
|
85 |
+
"""
|
86 |
+
"""
|
87 |
+
)
|
88 |
+
with gr.Row():
|
89 |
+
threshold = gr.Slider(minimum=0, maximum=1, value=0.6, step=0.05, label="Threshold")
|
90 |
+
object_larger_than_bg = gr.Checkbox(label="Object Larger than Background", value=False)
|
91 |
+
btn = gr.Button(label="Visualize")
|
92 |
+
with gr.Row():
|
93 |
+
img = gr.Image()
|
94 |
+
fig_pca = gr.Plot(label="PCA Features")
|
95 |
+
|
96 |
+
btn.click(fn=app_fn, inputs=[img, threshold, object_larger_than_bg], outputs=[fig_pca])
|
97 |
+
examples = gr.Examples(
|
98 |
+
examples=[
|
99 |
+
["assets/neca-the-cat.jpeg", 0.6, True],
|
100 |
+
["assets/dog.png", 0.7, False]
|
101 |
+
],
|
102 |
+
inputs=[img, threshold, object_larger_than_bg],
|
103 |
+
outputs=[fig_pca],
|
104 |
+
fn=app_fn,
|
105 |
+
cache_examples=True
|
106 |
+
)
|
107 |
+
|
108 |
+
demo.queue(max_size=5).launch()
|