Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -21,7 +21,6 @@ os.environ["HF_MODULES_CACHE"] = os.path.join("/tmp/hf_cache", "modules")
|
|
21 |
import transformers
|
22 |
transformers.utils.move_cache()
|
23 |
|
24 |
-
|
25 |
torch.set_float32_matmul_precision('high')
|
26 |
torch.jit.script = lambda f: f
|
27 |
|
@@ -37,15 +36,11 @@ def refine_foreground(image, mask, r=90):
|
|
37 |
image_masked = Image.fromarray((estimated_foreground * 255.0).astype(np.uint8))
|
38 |
return image_masked
|
39 |
|
40 |
-
|
41 |
def FB_blur_fusion_foreground_estimator_2(image, alpha, r=90):
|
42 |
-
# Thanks to the source: https://github.com/Photoroom/fast-foreground-estimation
|
43 |
alpha = alpha[:, :, None]
|
44 |
-
F, blur_B = FB_blur_fusion_foreground_estimator(
|
45 |
-
image, image, image, alpha, r)
|
46 |
return FB_blur_fusion_foreground_estimator(image, F, blur_B, alpha, r=6)[0]
|
47 |
|
48 |
-
|
49 |
def FB_blur_fusion_foreground_estimator(image, F, B, alpha, r=90):
|
50 |
if isinstance(image, Image.Image):
|
51 |
image = np.array(image) / 255.0
|
@@ -56,15 +51,12 @@ def FB_blur_fusion_foreground_estimator(image, F, B, alpha, r=90):
|
|
56 |
|
57 |
blurred_B1A = cv2.blur(B * (1 - alpha), (r, r))
|
58 |
blurred_B = blurred_B1A / ((1 - blurred_alpha) + 1e-5)
|
59 |
-
F = blurred_F + alpha *
|
60 |
-
(image - alpha * blurred_F - (1 - alpha) * blurred_B)
|
61 |
F = np.clip(F, 0, 1)
|
62 |
return F, blurred_B
|
63 |
|
64 |
-
|
65 |
class ImagePreprocessor():
|
66 |
def __init__(self, resolution: Tuple[int, int] = (1024, 1024)) -> None:
|
67 |
-
# Input resolution is on WxH.
|
68 |
self.transform_image = transforms.Compose([
|
69 |
transforms.Resize(resolution[::-1]),
|
70 |
transforms.ToTensor(),
|
@@ -72,9 +64,7 @@ class ImagePreprocessor():
|
|
72 |
])
|
73 |
|
74 |
def proc(self, image: Image.Image) -> torch.Tensor:
|
75 |
-
|
76 |
-
return image
|
77 |
-
|
78 |
|
79 |
usage_to_weights_file = {
|
80 |
'General': 'BiRefNet',
|
@@ -94,17 +84,18 @@ usage_to_weights_file = {
|
|
94 |
'General-dynamic': 'BiRefNet_dynamic',
|
95 |
}
|
96 |
|
97 |
-
birefnet = transformers.AutoModelForImageSegmentation.from_pretrained(
|
|
|
|
|
|
|
98 |
birefnet.to(device)
|
99 |
birefnet.eval(); birefnet.half()
|
100 |
|
101 |
-
|
102 |
@spaces.GPU
|
103 |
def predict(images, resolution, weights_file):
|
104 |
-
assert
|
105 |
|
106 |
global birefnet
|
107 |
-
# Load BiRefNet with chosen weights
|
108 |
_weights_file = '/'.join(('zhengpeng7', usage_to_weights_file[weights_file] if weights_file is not None else usage_to_weights_file['General']))
|
109 |
print('Using weights: {}.'.format(_weights_file))
|
110 |
birefnet = transformers.AutoModelForImageSegmentation.from_pretrained(_weights_file, trust_remote_code=True)
|
@@ -129,7 +120,6 @@ def predict(images, resolution, weights_file):
|
|
129 |
print('Invalid resolution input. Automatically changed to 1024x1024 / 2048x2048 / 2560x1440.')
|
130 |
|
131 |
if isinstance(images, list):
|
132 |
-
# For tab_batch
|
133 |
save_paths = []
|
134 |
save_dir = 'preds-BiRefNet'
|
135 |
if not os.path.exists(save_dir):
|
@@ -151,21 +141,17 @@ def predict(images, resolution, weights_file):
|
|
151 |
image_ori = Image.fromarray(image_src)
|
152 |
|
153 |
image = image_ori.convert('RGB')
|
154 |
-
# Preprocess the image
|
155 |
if resolution is None:
|
156 |
resolution_div_by_32 = [int(int(reso)//32*32) for reso in image.size]
|
157 |
if resolution_div_by_32 != resolution:
|
158 |
resolution = resolution_div_by_32
|
159 |
image_preprocessor = ImagePreprocessor(resolution=tuple(resolution))
|
160 |
-
image_proc = image_preprocessor.proc(image)
|
161 |
-
image_proc = image_proc.unsqueeze(0)
|
162 |
|
163 |
-
# Prediction
|
164 |
with torch.no_grad():
|
165 |
preds = birefnet(image_proc.to(device).half())[-1].sigmoid().cpu()
|
166 |
pred = preds[0].squeeze()
|
167 |
|
168 |
-
# Show Results
|
169 |
pred_pil = transforms.ToPILImage()(pred)
|
170 |
image_masked = refine_foreground(image, pred_pil)
|
171 |
image_masked.putalpha(pred_pil.resize(image.size))
|
@@ -184,32 +170,13 @@ def predict(images, resolution, weights_file):
|
|
184 |
zipf.write(file, os.path.basename(file))
|
185 |
return save_paths, zip_file_path
|
186 |
else:
|
187 |
-
return
|
188 |
-
|
189 |
-
|
190 |
-
examples = [[_] for _ in glob('examples/*')][:]
|
191 |
-
# Add the option of resolution in a text box.
|
192 |
-
for idx_example, example in enumerate(examples):
|
193 |
-
if 'My_' in example[0]:
|
194 |
-
example_resolution = '2048x2048'
|
195 |
-
else:
|
196 |
-
example_resolution = '1024x1024'
|
197 |
-
examples[idx_example].append(example_resolution)
|
198 |
-
examples.append(examples[-1].copy())
|
199 |
-
examples[-1][1] = '512x512'
|
200 |
-
|
201 |
-
examples_url = [
|
202 |
-
['https://hips.hearstapps.com/hmg-prod/images/gettyimages-1229892983-square.jpg'],
|
203 |
-
]
|
204 |
-
for idx_example_url, example_url in enumerate(examples_url):
|
205 |
-
examples_url[idx_example_url].append('1024x1024')
|
206 |
|
207 |
descriptions = (
|
208 |
"Upload a picture, and we'll remove the background!\n"
|
209 |
"The resolution used is `1024x1024`\n"
|
210 |
)
|
211 |
|
212 |
-
|
213 |
tab_image = gr.Interface(
|
214 |
fn=predict,
|
215 |
inputs=[
|
@@ -218,7 +185,6 @@ tab_image = gr.Interface(
|
|
218 |
gr.Radio(list(usage_to_weights_file.keys()), value='General', label="Weights", info="Choose the weights you want.")
|
219 |
],
|
220 |
outputs=gr.ImageSlider(label="BiRefNet's prediction", type="pil", format='png'),
|
221 |
-
examples=examples,
|
222 |
api_name="image",
|
223 |
description=descriptions,
|
224 |
)
|
@@ -231,9 +197,8 @@ tab_text = gr.Interface(
|
|
231 |
gr.Radio(list(usage_to_weights_file.keys()), value='General', label="Weights", info="Choose the weights you want.")
|
232 |
],
|
233 |
outputs=gr.ImageSlider(label="BiRefNet's prediction", type="pil", format='png'),
|
234 |
-
examples=examples_url,
|
235 |
api_name="URL",
|
236 |
-
description=descriptions+'\nTab-URL is partially modified from https://huggingface.co/spaces/not-lain/background-removal, thanks to this great work!',
|
237 |
)
|
238 |
|
239 |
tab_batch = gr.Interface(
|
@@ -245,7 +210,7 @@ tab_batch = gr.Interface(
|
|
245 |
],
|
246 |
outputs=[gr.Gallery(label="BiRefNet's predictions"), gr.File(label="Download masked images.")],
|
247 |
api_name="batch",
|
248 |
-
description=descriptions+'\nTab-batch is partially modified from https://huggingface.co/spaces/NegiTurkey/Multi_Birefnetfor_Background_Removal, thanks to this great work!',
|
249 |
)
|
250 |
|
251 |
demo = gr.TabbedInterface(
|
|
|
21 |
import transformers
|
22 |
transformers.utils.move_cache()
|
23 |
|
|
|
24 |
torch.set_float32_matmul_precision('high')
|
25 |
torch.jit.script = lambda f: f
|
26 |
|
|
|
36 |
image_masked = Image.fromarray((estimated_foreground * 255.0).astype(np.uint8))
|
37 |
return image_masked
|
38 |
|
|
|
39 |
def FB_blur_fusion_foreground_estimator_2(image, alpha, r=90):
|
|
|
40 |
alpha = alpha[:, :, None]
|
41 |
+
F, blur_B = FB_blur_fusion_foreground_estimator(image, image, image, alpha, r)
|
|
|
42 |
return FB_blur_fusion_foreground_estimator(image, F, blur_B, alpha, r=6)[0]
|
43 |
|
|
|
44 |
def FB_blur_fusion_foreground_estimator(image, F, B, alpha, r=90):
|
45 |
if isinstance(image, Image.Image):
|
46 |
image = np.array(image) / 255.0
|
|
|
51 |
|
52 |
blurred_B1A = cv2.blur(B * (1 - alpha), (r, r))
|
53 |
blurred_B = blurred_B1A / ((1 - blurred_alpha) + 1e-5)
|
54 |
+
F = blurred_F + alpha * (image - alpha * blurred_F - (1 - alpha) * blurred_B)
|
|
|
55 |
F = np.clip(F, 0, 1)
|
56 |
return F, blurred_B
|
57 |
|
|
|
58 |
class ImagePreprocessor():
|
59 |
def __init__(self, resolution: Tuple[int, int] = (1024, 1024)) -> None:
|
|
|
60 |
self.transform_image = transforms.Compose([
|
61 |
transforms.Resize(resolution[::-1]),
|
62 |
transforms.ToTensor(),
|
|
|
64 |
])
|
65 |
|
66 |
def proc(self, image: Image.Image) -> torch.Tensor:
|
67 |
+
return self.transform_image(image)
|
|
|
|
|
68 |
|
69 |
usage_to_weights_file = {
|
70 |
'General': 'BiRefNet',
|
|
|
84 |
'General-dynamic': 'BiRefNet_dynamic',
|
85 |
}
|
86 |
|
87 |
+
birefnet = transformers.AutoModelForImageSegmentation.from_pretrained(
|
88 |
+
'/'.join(('zhengpeng7', usage_to_weights_file['General'])),
|
89 |
+
trust_remote_code=True
|
90 |
+
)
|
91 |
birefnet.to(device)
|
92 |
birefnet.eval(); birefnet.half()
|
93 |
|
|
|
94 |
@spaces.GPU
|
95 |
def predict(images, resolution, weights_file):
|
96 |
+
assert images is not None, 'AssertionError: images cannot be None.'
|
97 |
|
98 |
global birefnet
|
|
|
99 |
_weights_file = '/'.join(('zhengpeng7', usage_to_weights_file[weights_file] if weights_file is not None else usage_to_weights_file['General']))
|
100 |
print('Using weights: {}.'.format(_weights_file))
|
101 |
birefnet = transformers.AutoModelForImageSegmentation.from_pretrained(_weights_file, trust_remote_code=True)
|
|
|
120 |
print('Invalid resolution input. Automatically changed to 1024x1024 / 2048x2048 / 2560x1440.')
|
121 |
|
122 |
if isinstance(images, list):
|
|
|
123 |
save_paths = []
|
124 |
save_dir = 'preds-BiRefNet'
|
125 |
if not os.path.exists(save_dir):
|
|
|
141 |
image_ori = Image.fromarray(image_src)
|
142 |
|
143 |
image = image_ori.convert('RGB')
|
|
|
144 |
if resolution is None:
|
145 |
resolution_div_by_32 = [int(int(reso)//32*32) for reso in image.size]
|
146 |
if resolution_div_by_32 != resolution:
|
147 |
resolution = resolution_div_by_32
|
148 |
image_preprocessor = ImagePreprocessor(resolution=tuple(resolution))
|
149 |
+
image_proc = image_preprocessor.proc(image).unsqueeze(0)
|
|
|
150 |
|
|
|
151 |
with torch.no_grad():
|
152 |
preds = birefnet(image_proc.to(device).half())[-1].sigmoid().cpu()
|
153 |
pred = preds[0].squeeze()
|
154 |
|
|
|
155 |
pred_pil = transforms.ToPILImage()(pred)
|
156 |
image_masked = refine_foreground(image, pred_pil)
|
157 |
image_masked.putalpha(pred_pil.resize(image.size))
|
|
|
170 |
zipf.write(file, os.path.basename(file))
|
171 |
return save_paths, zip_file_path
|
172 |
else:
|
173 |
+
return image_masked, image_ori
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
174 |
|
175 |
descriptions = (
|
176 |
"Upload a picture, and we'll remove the background!\n"
|
177 |
"The resolution used is `1024x1024`\n"
|
178 |
)
|
179 |
|
|
|
180 |
tab_image = gr.Interface(
|
181 |
fn=predict,
|
182 |
inputs=[
|
|
|
185 |
gr.Radio(list(usage_to_weights_file.keys()), value='General', label="Weights", info="Choose the weights you want.")
|
186 |
],
|
187 |
outputs=gr.ImageSlider(label="BiRefNet's prediction", type="pil", format='png'),
|
|
|
188 |
api_name="image",
|
189 |
description=descriptions,
|
190 |
)
|
|
|
197 |
gr.Radio(list(usage_to_weights_file.keys()), value='General', label="Weights", info="Choose the weights you want.")
|
198 |
],
|
199 |
outputs=gr.ImageSlider(label="BiRefNet's prediction", type="pil", format='png'),
|
|
|
200 |
api_name="URL",
|
201 |
+
description=descriptions + '\nTab-URL is partially modified from https://huggingface.co/spaces/not-lain/background-removal, thanks to this great work!',
|
202 |
)
|
203 |
|
204 |
tab_batch = gr.Interface(
|
|
|
210 |
],
|
211 |
outputs=[gr.Gallery(label="BiRefNet's predictions"), gr.File(label="Download masked images.")],
|
212 |
api_name="batch",
|
213 |
+
description=descriptions + '\nTab-batch is partially modified from https://huggingface.co/spaces/NegiTurkey/Multi_Birefnetfor_Background_Removal, thanks to this great work!',
|
214 |
)
|
215 |
|
216 |
demo = gr.TabbedInterface(
|