Spaces:
Sleeping
Sleeping
Upload folder using huggingface_hub
Browse files- app.py +31 -20
- generated-all.jpg +0 -0
- generated-bg.jpg +0 -0
- inference.py +9 -3
app.py
CHANGED
@@ -41,7 +41,7 @@ for style_name, style_img_path in style_options.items():
|
|
41 |
|
42 |
@spaces.GPU(duration=12)
|
43 |
def run(content_image, style_name, style_strength=5):
|
44 |
-
yield None
|
45 |
content_img, original_size = preprocess_img(content_image, img_size)
|
46 |
content_img = content_img.to(device)
|
47 |
|
@@ -55,32 +55,41 @@ def run(content_image, style_name, style_strength=5):
|
|
55 |
|
56 |
st = time.time()
|
57 |
|
58 |
-
|
59 |
-
|
|
|
60 |
|
61 |
-
def
|
62 |
with torch.cuda.stream(stream):
|
63 |
-
return
|
64 |
-
|
65 |
-
|
66 |
-
|
67 |
-
|
68 |
-
|
69 |
-
|
70 |
-
|
|
|
|
|
|
|
71 |
|
72 |
with ThreadPoolExecutor() as executor:
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
-
|
|
|
|
|
|
|
|
|
77 |
|
78 |
et = time.time()
|
79 |
print('TIME TAKEN:', et-st)
|
80 |
|
81 |
yield (
|
82 |
(content_image, postprocess_img(generated_img_all, original_size)),
|
83 |
-
(content_image, postprocess_img(generated_img_bg, original_size))
|
|
|
84 |
)
|
85 |
|
86 |
def set_slider(value):
|
@@ -115,7 +124,9 @@ with gr.Blocks(css=css) as demo:
|
|
115 |
with gr.Column():
|
116 |
output_image_all = ImageSlider(position=0.15, label='Styled Image', type='pil', interactive=False, show_download_button=False)
|
117 |
download_button_1 = gr.DownloadButton(label='Download Styled Image', visible=False)
|
118 |
-
|
|
|
|
|
119 |
download_button_2 = gr.DownloadButton(label='Download Styled Background', visible=False)
|
120 |
|
121 |
def save_image(img_tuple1, img_tuple2):
|
@@ -132,7 +143,7 @@ with gr.Blocks(css=css) as demo:
|
|
132 |
submit_button.click(
|
133 |
fn=run,
|
134 |
inputs=[content_image, style_dropdown, style_strength_slider],
|
135 |
-
outputs=[output_image_all, output_image_background]
|
136 |
).then(
|
137 |
fn=save_image,
|
138 |
inputs=[output_image_all, output_image_background],
|
@@ -144,4 +155,4 @@ with gr.Blocks(css=css) as demo:
|
|
144 |
|
145 |
demo.queue = False
|
146 |
demo.config['queue'] = False
|
147 |
-
demo.launch(show_api=False)
|
|
|
41 |
|
42 |
@spaces.GPU(duration=12)
|
43 |
def run(content_image, style_name, style_strength=5):
|
44 |
+
yield [None] * 3
|
45 |
content_img, original_size = preprocess_img(content_image, img_size)
|
46 |
content_img = content_img.to(device)
|
47 |
|
|
|
55 |
|
56 |
st = time.time()
|
57 |
|
58 |
+
if device == 'cuda':
|
59 |
+
stream_all = torch.cuda.Stream()
|
60 |
+
stream_bg = torch.cuda.Stream()
|
61 |
|
62 |
+
def run_inference_cuda(apply_to_background, stream):
|
63 |
with torch.cuda.stream(stream):
|
64 |
+
return run_inference(apply_to_background)
|
65 |
+
|
66 |
+
def run_inference(apply_to_background):
|
67 |
+
return inference(
|
68 |
+
model=model,
|
69 |
+
segmentation_model=segmentation_model,
|
70 |
+
content_image=content_img,
|
71 |
+
style_features=style_features,
|
72 |
+
lr=lrs[style_strength-1],
|
73 |
+
apply_to_background=apply_to_background
|
74 |
+
)
|
75 |
|
76 |
with ThreadPoolExecutor() as executor:
|
77 |
+
if device == 'cuda':
|
78 |
+
future_all = executor.submit(run_inference_cuda, False, stream_all)
|
79 |
+
future_bg = executor.submit(run_inference_cuda, True, stream_bg)
|
80 |
+
else:
|
81 |
+
future_all = executor.submit(run_inference, False)
|
82 |
+
future_bg = executor.submit(run_inference, True)
|
83 |
+
generated_img_all, _ = future_all.result()
|
84 |
+
generated_img_bg, bg_ratio = future_bg.result()
|
85 |
|
86 |
et = time.time()
|
87 |
print('TIME TAKEN:', et-st)
|
88 |
|
89 |
yield (
|
90 |
(content_image, postprocess_img(generated_img_all, original_size)),
|
91 |
+
(content_image, postprocess_img(generated_img_bg, original_size)),
|
92 |
+
f'{bg_ratio:.2f}'
|
93 |
)
|
94 |
|
95 |
def set_slider(value):
|
|
|
124 |
with gr.Column():
|
125 |
output_image_all = ImageSlider(position=0.15, label='Styled Image', type='pil', interactive=False, show_download_button=False)
|
126 |
download_button_1 = gr.DownloadButton(label='Download Styled Image', visible=False)
|
127 |
+
with gr.Group():
|
128 |
+
output_image_background = ImageSlider(position=0.15, label='Styled Background', type='pil', interactive=False, show_download_button=False)
|
129 |
+
bg_ratio_label = gr.Label(label='Background Ratio')
|
130 |
download_button_2 = gr.DownloadButton(label='Download Styled Background', visible=False)
|
131 |
|
132 |
def save_image(img_tuple1, img_tuple2):
|
|
|
143 |
submit_button.click(
|
144 |
fn=run,
|
145 |
inputs=[content_image, style_dropdown, style_strength_slider],
|
146 |
+
outputs=[output_image_all, output_image_background, bg_ratio_label]
|
147 |
).then(
|
148 |
fn=save_image,
|
149 |
inputs=[output_image_all, output_image_background],
|
|
|
155 |
|
156 |
demo.queue = False
|
157 |
demo.config['queue'] = False
|
158 |
+
demo.launch(show_api=False)
|
generated-all.jpg
ADDED
generated-bg.jpg
ADDED
inference.py
CHANGED
@@ -52,13 +52,19 @@ def inference(
|
|
52 |
with torch.no_grad():
|
53 |
content_features = model(content_image)
|
54 |
|
55 |
-
resized_bg_masks = []
|
|
|
56 |
if apply_to_background:
|
57 |
segmentation_output = segmentation_model(content_image)['out']
|
58 |
segmentation_mask = segmentation_output.argmax(dim=1)
|
59 |
|
60 |
background_mask = (segmentation_mask == 0).float()
|
61 |
foreground_mask = (segmentation_mask != 0).float()
|
|
|
|
|
|
|
|
|
|
|
62 |
|
63 |
for cf in content_features:
|
64 |
_, _, h_i, w_i = cf.shape
|
@@ -83,6 +89,6 @@ def inference(
|
|
83 |
foreground_mask_resized = F.interpolate(foreground_mask.unsqueeze(1), size=generated_image.shape[2:], mode='nearest')
|
84 |
generated_image.data = generated_image.data * (1 - foreground_mask_resized) + content_image.data * foreground_mask_resized
|
85 |
|
86 |
-
if iter % 10 == 0: print(f'Loss ({iter}):', min_losses[iter])
|
87 |
|
88 |
-
return generated_image
|
|
|
52 |
with torch.no_grad():
|
53 |
content_features = model(content_image)
|
54 |
|
55 |
+
resized_bg_masks = []
|
56 |
+
background_ratio = None
|
57 |
if apply_to_background:
|
58 |
segmentation_output = segmentation_model(content_image)['out']
|
59 |
segmentation_mask = segmentation_output.argmax(dim=1)
|
60 |
|
61 |
background_mask = (segmentation_mask == 0).float()
|
62 |
foreground_mask = (segmentation_mask != 0).float()
|
63 |
+
|
64 |
+
background_pixel_count = background_mask.sum().item()
|
65 |
+
total_pixel_count = segmentation_mask.numel()
|
66 |
+
background_ratio = background_pixel_count / total_pixel_count
|
67 |
+
print(f'Background Detected: {background_ratio * 100:.2f}%')
|
68 |
|
69 |
for cf in content_features:
|
70 |
_, _, h_i, w_i = cf.shape
|
|
|
89 |
foreground_mask_resized = F.interpolate(foreground_mask.unsqueeze(1), size=generated_image.shape[2:], mode='nearest')
|
90 |
generated_image.data = generated_image.data * (1 - foreground_mask_resized) + content_image.data * foreground_mask_resized
|
91 |
|
92 |
+
if iter % 10 == 0: print(f'[{'Background' if apply_to_background else 'Image'}] Loss ({iter}):', min_losses[iter])
|
93 |
|
94 |
+
return generated_image, background_ratio
|