Spaces:
Sleeping
Sleeping
gradio plot
Browse files
app.py
CHANGED
@@ -10,89 +10,110 @@ import gradio as gr
|
|
10 |
|
11 |
|
12 |
def compute_spatial_similarity(conv1, conv2):
|
13 |
-
|
14 |
-
|
15 |
-
|
16 |
-
|
17 |
-
|
18 |
-
|
19 |
-
|
20 |
-
|
21 |
-
|
22 |
-
|
23 |
-
|
24 |
-
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
-
|
31 |
-
|
32 |
-
|
|
|
33 |
|
34 |
# Get Layer 4
|
35 |
|
36 |
-
display_transform = transforms.Compose(
|
|
|
|
|
37 |
|
38 |
imagenet_transform = transforms.Compose(
|
39 |
-
|
40 |
-
|
41 |
-
|
42 |
-
|
|
|
|
|
|
|
43 |
|
44 |
|
45 |
class Wrapper(torch.nn.Module):
|
46 |
-
|
47 |
-
|
48 |
-
|
49 |
-
|
50 |
-
|
51 |
-
|
52 |
-
|
53 |
-
|
54 |
-
|
55 |
-
|
56 |
-
|
57 |
-
|
58 |
-
|
59 |
-
|
60 |
-
|
|
|
|
|
|
|
61 |
def get_layer4(input_image):
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
|
66 |
|
67 |
-
|
68 |
-
|
69 |
-
|
70 |
-
|
|
|
|
|
71 |
|
72 |
-
return reference_layer4.data.to('cpu').numpy()
|
73 |
|
74 |
# Visualization
|
75 |
def visualize_similarities(image1, image2):
|
76 |
-
|
77 |
-
|
78 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
79 |
|
80 |
-
|
81 |
-
|
82 |
-
|
83 |
-
|
|
|
84 |
|
85 |
-
|
86 |
-
|
87 |
-
|
|
|
88 |
|
89 |
-
fig.colorbar(im1, ax=axes[0])
|
90 |
-
fig.colorbar(im2, ax=axes[1])
|
91 |
-
plt.tight_layout()
|
92 |
-
return fig
|
93 |
|
94 |
# GRADIO APP
|
95 |
-
iface = gr.Interface(
|
96 |
-
|
97 |
-
|
98 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
10 |
|
11 |
|
12 |
def compute_spatial_similarity(conv1, conv2):
|
13 |
+
"""
|
14 |
+
Takes in the last convolutional layer from two images, computes the pooled output
|
15 |
+
feature, and then generates the spatial similarity map for both images.
|
16 |
+
"""
|
17 |
+
conv1 = conv1.reshape(-1, 7 * 7).T
|
18 |
+
conv2 = conv2.reshape(-1, 7 * 7).T
|
19 |
+
|
20 |
+
pool1 = np.mean(conv1, axis=0)
|
21 |
+
pool2 = np.mean(conv2, axis=0)
|
22 |
+
out_sz = (int(np.sqrt(conv1.shape[0])), int(np.sqrt(conv1.shape[0])))
|
23 |
+
conv1_normed = conv1 / np.linalg.norm(pool1) / conv1.shape[0]
|
24 |
+
conv2_normed = conv2 / np.linalg.norm(pool2) / conv2.shape[0]
|
25 |
+
im_similarity = np.zeros((conv1_normed.shape[0], conv1_normed.shape[0]))
|
26 |
+
|
27 |
+
for zz in range(conv1_normed.shape[0]):
|
28 |
+
repPx = mb.repmat(conv1_normed[zz, :], conv1_normed.shape[0], 1)
|
29 |
+
im_similarity[zz, :] = np.multiply(repPx, conv2_normed).sum(axis=1)
|
30 |
+
similarity1 = np.reshape(np.sum(im_similarity, axis=1), out_sz)
|
31 |
+
similarity2 = np.reshape(np.sum(im_similarity, axis=0), out_sz)
|
32 |
+
return similarity1, similarity2
|
33 |
+
|
34 |
|
35 |
# Get Layer 4
|
36 |
|
37 |
+
display_transform = transforms.Compose(
|
38 |
+
[transforms.Resize(256), transforms.CenterCrop((224, 224))]
|
39 |
+
)
|
40 |
|
41 |
imagenet_transform = transforms.Compose(
|
42 |
+
[
|
43 |
+
transforms.Resize(256),
|
44 |
+
transforms.CenterCrop((224, 224)),
|
45 |
+
transforms.ToTensor(),
|
46 |
+
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
|
47 |
+
]
|
48 |
+
)
|
49 |
|
50 |
|
51 |
class Wrapper(torch.nn.Module):
|
52 |
+
def __init__(self, model):
|
53 |
+
super(Wrapper, self).__init__()
|
54 |
+
self.model = model
|
55 |
+
self.layer4_ouputs = None
|
56 |
+
|
57 |
+
def fw_hook(module, input, output):
|
58 |
+
self.layer4_ouputs = output
|
59 |
+
|
60 |
+
self.model.layer4.register_forward_hook(fw_hook)
|
61 |
+
|
62 |
+
def forward(self, input):
|
63 |
+
_ = self.model(input)
|
64 |
+
return self.layer4_ouputs
|
65 |
+
|
66 |
+
def __repr__(self):
|
67 |
+
return "Wrapper"
|
68 |
+
|
69 |
+
|
70 |
def get_layer4(input_image):
|
71 |
+
l4_model = models.resnet50(pretrained=True)
|
72 |
+
# l4_model = l4_model.cuda()
|
73 |
+
l4_model.eval()
|
74 |
+
wrapped_model = Wrapper(l4_model)
|
75 |
|
76 |
+
with torch.no_grad():
|
77 |
+
data = imagenet_transform(input_image).unsqueeze(0)
|
78 |
+
# data = data.cuda()
|
79 |
+
reference_layer4 = wrapped_model(data)
|
80 |
+
|
81 |
+
return reference_layer4.data.to("cpu").numpy()
|
82 |
|
|
|
83 |
|
84 |
# Visualization
|
85 |
def visualize_similarities(image1, image2):
|
86 |
+
a = get_layer4(image1).squeeze()
|
87 |
+
b = get_layer4(image2).squeeze()
|
88 |
+
sim1, sim2 = compute_spatial_similarity(a, b)
|
89 |
+
|
90 |
+
fig, axes = plt.subplots(1, 2, figsize=(12, 5))
|
91 |
+
axes[0].imshow(display_transform(image1))
|
92 |
+
im1 = axes[0].imshow(
|
93 |
+
skimage.transform.resize(sim1, (224, 224)), alpha=0.6, cmap="jet"
|
94 |
+
)
|
95 |
+
# axes[0].colorbar()
|
96 |
|
97 |
+
axes[1].imshow(display_transform(image2))
|
98 |
+
im2 = axes[1].imshow(
|
99 |
+
skimage.transform.resize(sim2, (224, 224)), alpha=0.6, cmap="jet"
|
100 |
+
)
|
101 |
+
# axes[1].colorbar()
|
102 |
|
103 |
+
fig.colorbar(im1, ax=axes[0])
|
104 |
+
fig.colorbar(im2, ax=axes[1])
|
105 |
+
plt.tight_layout()
|
106 |
+
return fig
|
107 |
|
|
|
|
|
|
|
|
|
108 |
|
109 |
# GRADIO APP
|
110 |
+
iface = gr.Interface(
|
111 |
+
fn=visualize_similarities,
|
112 |
+
inputs=[
|
113 |
+
gr.inputs.Image(shape=(300, 300), type="pil"),
|
114 |
+
gr.inputs.Image(shape=(300, 300), type="pil"),
|
115 |
+
],
|
116 |
+
outputs=[gr.outputs.Plot(type="matplotlib")],
|
117 |
+
)
|
118 |
+
|
119 |
+
iface.launch()
|