Spaces:
Sleeping
Sleeping
Update saliency_gradio.py
Browse files- saliency_gradio.py +34 -39
saliency_gradio.py
CHANGED
@@ -10,13 +10,10 @@ hf_dir = snapshot_download(repo_id="alexanderkroner/MSI-Net")
|
|
10 |
|
11 |
def get_target_shape(original_shape):
|
12 |
original_aspect_ratio = original_shape[0] / original_shape[1]
|
13 |
-
|
14 |
square_mode = abs(original_aspect_ratio - 1.0)
|
15 |
landscape_mode = abs(original_aspect_ratio - 240 / 320)
|
16 |
portrait_mode = abs(original_aspect_ratio - 320 / 240)
|
17 |
-
|
18 |
best_mode = min(square_mode, landscape_mode, portrait_mode)
|
19 |
-
|
20 |
if best_mode == square_mode:
|
21 |
return (320, 320)
|
22 |
elif best_mode == landscape_mode:
|
@@ -26,19 +23,13 @@ def get_target_shape(original_shape):
|
|
26 |
|
27 |
def preprocess_input(input_image, target_shape):
|
28 |
input_tensor = tf.expand_dims(input_image, axis=0)
|
29 |
-
|
30 |
-
input_tensor = tf.image.resize(
|
31 |
-
input_tensor, target_shape, preserve_aspect_ratio=True
|
32 |
-
)
|
33 |
-
|
34 |
vertical_padding = target_shape[0] - input_tensor.shape[1]
|
35 |
horizontal_padding = target_shape[1] - input_tensor.shape[2]
|
36 |
-
|
37 |
vertical_padding_1 = vertical_padding // 2
|
38 |
vertical_padding_2 = vertical_padding - vertical_padding_1
|
39 |
horizontal_padding_1 = horizontal_padding // 2
|
40 |
horizontal_padding_2 = horizontal_padding - horizontal_padding_1
|
41 |
-
|
42 |
input_tensor = tf.pad(
|
43 |
input_tensor,
|
44 |
[
|
@@ -48,12 +39,7 @@ def preprocess_input(input_image, target_shape):
|
|
48 |
[0, 0],
|
49 |
],
|
50 |
)
|
51 |
-
|
52 |
-
return (
|
53 |
-
input_tensor,
|
54 |
-
[vertical_padding_1, vertical_padding_2],
|
55 |
-
[horizontal_padding_1, horizontal_padding_2],
|
56 |
-
)
|
57 |
|
58 |
def postprocess_output(output_tensor, vertical_padding, horizontal_padding, original_shape):
|
59 |
output_tensor = output_tensor[
|
@@ -62,45 +48,54 @@ def postprocess_output(output_tensor, vertical_padding, horizontal_padding, orig
|
|
62 |
horizontal_padding[0] : output_tensor.shape[2] - horizontal_padding[1],
|
63 |
:,
|
64 |
]
|
65 |
-
|
66 |
output_tensor = tf.image.resize(output_tensor, original_shape)
|
67 |
-
return output_tensor.numpy().squeeze()
|
68 |
|
69 |
def process_image(input_image):
|
70 |
input_image = np.array(input_image, dtype=np.float32)
|
71 |
original_shape = input_image.shape[:2]
|
72 |
target_shape = get_target_shape(original_shape)
|
73 |
-
|
74 |
input_tensor, vertical_padding, horizontal_padding = preprocess_input(input_image, target_shape)
|
75 |
output_tensor = model(input_tensor)["output"]
|
76 |
saliency_gray = postprocess_output(output_tensor, vertical_padding, horizontal_padding, original_shape)
|
77 |
total_saliency = np.sum(saliency_gray)
|
78 |
-
|
79 |
saliency_rgb = plt.cm.inferno(saliency_gray)[..., :3]
|
80 |
alpha = 0.9
|
81 |
blended_image = alpha * saliency_rgb + (1 - alpha) * input_image / 255
|
82 |
-
|
83 |
return blended_image, f"Total grayscale saliency: {total_saliency:.2f}"
|
84 |
|
85 |
-
def
|
|
|
|
|
|
|
86 |
result1_img, result1_val = process_image(image1)
|
87 |
result2_img, result2_val = process_image(image2)
|
88 |
return result1_img, result1_val, result2_img, result2_val
|
89 |
|
90 |
-
|
91 |
-
|
92 |
-
|
93 |
-
gr.
|
94 |
-
|
95 |
-
|
96 |
-
|
97 |
-
|
98 |
-
|
99 |
-
|
100 |
-
|
101 |
-
|
102 |
-
|
103 |
-
|
104 |
-
)
|
105 |
-
|
106 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
10 |
|
11 |
def get_target_shape(original_shape):
|
12 |
original_aspect_ratio = original_shape[0] / original_shape[1]
|
|
|
13 |
square_mode = abs(original_aspect_ratio - 1.0)
|
14 |
landscape_mode = abs(original_aspect_ratio - 240 / 320)
|
15 |
portrait_mode = abs(original_aspect_ratio - 320 / 240)
|
|
|
16 |
best_mode = min(square_mode, landscape_mode, portrait_mode)
|
|
|
17 |
if best_mode == square_mode:
|
18 |
return (320, 320)
|
19 |
elif best_mode == landscape_mode:
|
|
|
23 |
|
24 |
def preprocess_input(input_image, target_shape):
|
25 |
input_tensor = tf.expand_dims(input_image, axis=0)
|
26 |
+
input_tensor = tf.image.resize(input_tensor, target_shape, preserve_aspect_ratio=True)
|
|
|
|
|
|
|
|
|
27 |
vertical_padding = target_shape[0] - input_tensor.shape[1]
|
28 |
horizontal_padding = target_shape[1] - input_tensor.shape[2]
|
|
|
29 |
vertical_padding_1 = vertical_padding // 2
|
30 |
vertical_padding_2 = vertical_padding - vertical_padding_1
|
31 |
horizontal_padding_1 = horizontal_padding // 2
|
32 |
horizontal_padding_2 = horizontal_padding - horizontal_padding_1
|
|
|
33 |
input_tensor = tf.pad(
|
34 |
input_tensor,
|
35 |
[
|
|
|
39 |
[0, 0],
|
40 |
],
|
41 |
)
|
42 |
+
return input_tensor, [vertical_padding_1, vertical_padding_2], [horizontal_padding_1, horizontal_padding_2]
|
|
|
|
|
|
|
|
|
|
|
43 |
|
44 |
def postprocess_output(output_tensor, vertical_padding, horizontal_padding, original_shape):
|
45 |
output_tensor = output_tensor[
|
|
|
48 |
horizontal_padding[0] : output_tensor.shape[2] - horizontal_padding[1],
|
49 |
:,
|
50 |
]
|
|
|
51 |
output_tensor = tf.image.resize(output_tensor, original_shape)
|
52 |
+
return output_tensor.numpy().squeeze()
|
53 |
|
54 |
def process_image(input_image):
|
55 |
input_image = np.array(input_image, dtype=np.float32)
|
56 |
original_shape = input_image.shape[:2]
|
57 |
target_shape = get_target_shape(original_shape)
|
|
|
58 |
input_tensor, vertical_padding, horizontal_padding = preprocess_input(input_image, target_shape)
|
59 |
output_tensor = model(input_tensor)["output"]
|
60 |
saliency_gray = postprocess_output(output_tensor, vertical_padding, horizontal_padding, original_shape)
|
61 |
total_saliency = np.sum(saliency_gray)
|
|
|
62 |
saliency_rgb = plt.cm.inferno(saliency_gray)[..., :3]
|
63 |
alpha = 0.9
|
64 |
blended_image = alpha * saliency_rgb + (1 - alpha) * input_image / 255
|
|
|
65 |
return blended_image, f"Total grayscale saliency: {total_saliency:.2f}"
|
66 |
|
67 |
+
def predict_single(image):
|
68 |
+
return process_image(image)
|
69 |
+
|
70 |
+
def predict_dual(image1, image2):
|
71 |
result1_img, result1_val = process_image(image1)
|
72 |
result2_img, result2_val = process_image(image2)
|
73 |
return result1_img, result1_val, result2_img, result2_val
|
74 |
|
75 |
+
with gr.Blocks(title="MSI-Net Saliency App") as demo:
|
76 |
+
gr.Markdown("## MSI-Net Saliency Map Viewer")
|
77 |
+
with gr.Tabs():
|
78 |
+
with gr.Tab("Single Image"):
|
79 |
+
gr.Markdown("### Upload an image to see its saliency map and total grayscale saliency value.")
|
80 |
+
with gr.Row():
|
81 |
+
input_image_single = gr.Image(type="pil", label="Input Image")
|
82 |
+
with gr.Row():
|
83 |
+
output_image_single = gr.Image(type="numpy", label="Saliency Map")
|
84 |
+
output_text_single = gr.Textbox(label="Grayscale Sum")
|
85 |
+
submit_single = gr.Button("Generate Saliency")
|
86 |
+
submit_single.click(fn=predict_single, inputs=input_image_single, outputs=[output_image_single, output_text_single])
|
87 |
+
|
88 |
+
with gr.Tab("Compare Two Images"):
|
89 |
+
gr.Markdown("### Upload two images to compare their saliency maps and grayscale saliency values.")
|
90 |
+
with gr.Row():
|
91 |
+
input_image1 = gr.Image(type="pil", label="Image 1")
|
92 |
+
input_image2 = gr.Image(type="pil", label="Image 2")
|
93 |
+
with gr.Row():
|
94 |
+
output_image1 = gr.Image(type="numpy", label="Saliency Map 1")
|
95 |
+
output_text1 = gr.Textbox(label="Grayscale Sum 1")
|
96 |
+
output_image2 = gr.Image(type="numpy", label="Saliency Map 2")
|
97 |
+
output_text2 = gr.Textbox(label="Grayscale Sum 2")
|
98 |
+
submit_dual = gr.Button("Compare Saliency")
|
99 |
+
submit_dual.click(fn=predict_dual, inputs=[input_image1, input_image2], outputs=[output_image1, output_text1, output_image2, output_text2])
|
100 |
+
|
101 |
+
demo.launch(share=True)
|