Spaces:
Running
on
Zero
Running
on
Zero
Commit
·
ce634b2
1
Parent(s):
e391bd4
Acceleration on the refine_foreground.
Browse files- app.py +90 -25
- app_local.py +90 -25
app.py
CHANGED
@@ -9,6 +9,7 @@ from glob import glob
|
|
9 |
from typing import Tuple
|
10 |
|
11 |
from PIL import Image
|
|
|
12 |
from torchvision import transforms
|
13 |
|
14 |
import requests
|
@@ -27,39 +28,103 @@ torch.jit.script = lambda f: f
|
|
27 |
|
28 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
29 |
|
30 |
-
### image_proc.py
|
31 |
-
def refine_foreground(image, mask, r=90):
|
32 |
-
if mask.size != image.size:
|
33 |
-
mask = mask.resize(image.size)
|
34 |
-
image = np.array(image) / 255.0
|
35 |
-
mask = np.array(mask) / 255.0
|
36 |
-
estimated_foreground = FB_blur_fusion_foreground_estimator_2(image, mask, r=r)
|
37 |
-
image_masked = Image.fromarray((estimated_foreground * 255.0).astype(np.uint8))
|
38 |
-
return image_masked
|
39 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
40 |
|
41 |
-
|
|
|
42 |
# Thanks to the source: https://github.com/Photoroom/fast-foreground-estimation
|
43 |
alpha = alpha[:, :, None]
|
44 |
-
|
45 |
-
|
46 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
47 |
|
|
|
48 |
|
49 |
-
|
50 |
-
if isinstance(image, Image.Image):
|
51 |
-
image = np.array(image) / 255.0
|
52 |
-
blurred_alpha = cv2.blur(alpha, (r, r))[:, :, None]
|
53 |
|
54 |
-
|
55 |
-
|
56 |
|
57 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
58 |
blurred_B = blurred_B1A / ((1 - blurred_alpha) + 1e-5)
|
59 |
-
|
60 |
-
|
61 |
-
|
62 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
63 |
|
64 |
|
65 |
class ImagePreprocessor():
|
@@ -167,7 +232,7 @@ def predict(images, resolution, weights_file):
|
|
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))
|
172 |
|
173 |
torch.cuda.empty_cache()
|
|
|
9 |
from typing import Tuple
|
10 |
|
11 |
from PIL import Image
|
12 |
+
import torch
|
13 |
from torchvision import transforms
|
14 |
|
15 |
import requests
|
|
|
28 |
|
29 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
30 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
31 |
|
32 |
+
## CPU version refinement
|
33 |
+
def FB_blur_fusion_foreground_estimator_cpu(image, FG, B, alpha, r=90):
|
34 |
+
if isinstance(image, Image.Image):
|
35 |
+
image = np.array(image) / 255.0
|
36 |
+
blurred_alpha = cv2.blur(alpha, (r, r))[:, :, None]
|
37 |
+
|
38 |
+
blurred_FGA = cv2.blur(FG * alpha, (r, r))
|
39 |
+
blurred_FG = blurred_FGA / (blurred_alpha + 1e-5)
|
40 |
+
|
41 |
+
blurred_B1A = cv2.blur(B * (1 - alpha), (r, r))
|
42 |
+
blurred_B = blurred_B1A / ((1 - blurred_alpha) + 1e-5)
|
43 |
+
FG = blurred_FG + alpha * (image - alpha * blurred_FG - (1 - alpha) * blurred_B)
|
44 |
+
FG = np.clip(FG, 0, 1)
|
45 |
+
return FG, blurred_B
|
46 |
|
47 |
+
|
48 |
+
def FB_blur_fusion_foreground_estimator_cpu_2(image, alpha, r=90):
|
49 |
# Thanks to the source: https://github.com/Photoroom/fast-foreground-estimation
|
50 |
alpha = alpha[:, :, None]
|
51 |
+
FG, blur_B = FB_blur_fusion_foreground_estimator_cpu(image, image, image, alpha, r)
|
52 |
+
return FB_blur_fusion_foreground_estimator_cpu(image, FG, blur_B, alpha, r=6)[0]
|
53 |
+
|
54 |
+
|
55 |
+
## GPU version refinement
|
56 |
+
def mean_blur(x, kernel_size):
|
57 |
+
"""
|
58 |
+
equivalent to cv.blur
|
59 |
+
x: [B, C, H, W]
|
60 |
+
"""
|
61 |
+
if kernel_size % 2 == 0:
|
62 |
+
pad_l = kernel_size // 2 - 1
|
63 |
+
pad_r = kernel_size // 2
|
64 |
+
pad_t = kernel_size // 2 - 1
|
65 |
+
pad_b = kernel_size // 2
|
66 |
+
else:
|
67 |
+
pad_l = pad_r = pad_t = pad_b = kernel_size // 2
|
68 |
|
69 |
+
x_padded = torch.nn.functional.pad(x, (pad_l, pad_r, pad_t, pad_b), mode='replicate')
|
70 |
|
71 |
+
return torch.nn.functional.avg_pool2d(x_padded, kernel_size=(kernel_size, kernel_size), stride=1, count_include_pad=False)
|
|
|
|
|
|
|
72 |
|
73 |
+
def FB_blur_fusion_foreground_estimator_gpu(image, FG, B, alpha, r=90):
|
74 |
+
as_dtype = lambda x, dtype: x.to(dtype) if x.dtype != dtype else x
|
75 |
|
76 |
+
input_dtype = image.dtype
|
77 |
+
# convert image to float to avoid overflow
|
78 |
+
image = as_dtype(image, torch.float32)
|
79 |
+
FG = as_dtype(FG, torch.float32)
|
80 |
+
B = as_dtype(B, torch.float32)
|
81 |
+
alpha = as_dtype(alpha, torch.float32)
|
82 |
+
|
83 |
+
blurred_alpha = mean_blur(alpha, kernel_size=r)
|
84 |
+
|
85 |
+
blurred_FGA = mean_blur(FG * alpha, kernel_size=r)
|
86 |
+
blurred_FG = blurred_FGA / (blurred_alpha + 1e-5)
|
87 |
+
|
88 |
+
blurred_B1A = mean_blur(B * (1 - alpha), kernel_size=r)
|
89 |
blurred_B = blurred_B1A / ((1 - blurred_alpha) + 1e-5)
|
90 |
+
|
91 |
+
FG_output = blurred_FG + alpha * (image - alpha * blurred_FG - (1 - alpha) * blurred_B)
|
92 |
+
FG_output = torch.clamp(FG_output, 0, 1)
|
93 |
+
|
94 |
+
return as_dtype(FG_output, input_dtype), as_dtype(blurred_B, input_dtype)
|
95 |
+
|
96 |
+
|
97 |
+
def FB_blur_fusion_foreground_estimator_gpu_2(image, alpha, r=90):
|
98 |
+
# Thanks to the source: https://github.com/ZhengPeng7/BiRefNet/issues/226#issuecomment-3016433728
|
99 |
+
FG, blur_B = FB_blur_fusion_foreground_estimator_gpu(image, image, image, alpha, r)
|
100 |
+
return FB_blur_fusion_foreground_estimator_gpu(image, FG, blur_B, alpha, r=6)[0]
|
101 |
+
|
102 |
+
|
103 |
+
def refine_foreground(image, mask, r=90, device='cuda'):
|
104 |
+
"""both image and mask are in range of [0, 1]"""
|
105 |
+
if mask.size != image.size:
|
106 |
+
mask = mask.resize(image.size)
|
107 |
+
|
108 |
+
if device == 'cuda':
|
109 |
+
image = transforms.functional.to_tensor(image).float().cuda()
|
110 |
+
mask = transforms.functional.to_tensor(mask).float().cuda()
|
111 |
+
image = image.unsqueeze(0)
|
112 |
+
mask = mask.unsqueeze(0)
|
113 |
+
|
114 |
+
estimated_foreground = FB_blur_fusion_foreground_estimator_gpu_2(image, mask, r=r)
|
115 |
+
|
116 |
+
estimated_foreground = estimated_foreground.squeeze()
|
117 |
+
estimated_foreground = (estimated_foreground.mul(255.0)).to(torch.uint8)
|
118 |
+
estimated_foreground = estimated_foreground.permute(1, 2, 0).contiguous().cpu().numpy().astype(np.uint8)
|
119 |
+
else:
|
120 |
+
image = np.array(image, dtype=np.float32) / 255.0
|
121 |
+
mask = np.array(mask, dtype=np.float32) / 255.0
|
122 |
+
estimated_foreground = FB_blur_fusion_foreground_estimator_cpu_2(image, mask, r=r)
|
123 |
+
estimated_foreground = (estimated_foreground * 255.0).astype(np.uint8)
|
124 |
+
|
125 |
+
estimated_foreground = Image.fromarray(np.ascontiguousarray(estimated_foreground))
|
126 |
+
|
127 |
+
return estimated_foreground
|
128 |
|
129 |
|
130 |
class ImagePreprocessor():
|
|
|
232 |
|
233 |
# Show Results
|
234 |
pred_pil = transforms.ToPILImage()(pred)
|
235 |
+
image_masked = refine_foreground(image, pred_pil, device=device)
|
236 |
image_masked.putalpha(pred_pil.resize(image.size))
|
237 |
|
238 |
torch.cuda.empty_cache()
|
app_local.py
CHANGED
@@ -11,6 +11,7 @@ from typing import Tuple
|
|
11 |
from PIL import Image
|
12 |
# from gradio_imageslider import ImageSlider
|
13 |
import transformers
|
|
|
14 |
from torchvision import transforms
|
15 |
|
16 |
import requests
|
@@ -23,39 +24,103 @@ torch.set_float32_matmul_precision('high')
|
|
23 |
|
24 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
25 |
|
26 |
-
### image_proc.py
|
27 |
-
def refine_foreground(image, mask, r=90):
|
28 |
-
if mask.size != image.size:
|
29 |
-
mask = mask.resize(image.size)
|
30 |
-
image = np.array(image) / 255.0
|
31 |
-
mask = np.array(mask) / 255.0
|
32 |
-
estimated_foreground = FB_blur_fusion_foreground_estimator_2(image, mask, r=r)
|
33 |
-
image_masked = Image.fromarray((estimated_foreground * 255.0).astype(np.uint8))
|
34 |
-
return image_masked
|
35 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
36 |
|
37 |
-
|
|
|
38 |
# Thanks to the source: https://github.com/Photoroom/fast-foreground-estimation
|
39 |
alpha = alpha[:, :, None]
|
40 |
-
|
41 |
-
|
42 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
43 |
|
|
|
44 |
|
45 |
-
|
46 |
-
if isinstance(image, Image.Image):
|
47 |
-
image = np.array(image) / 255.0
|
48 |
-
blurred_alpha = cv2.blur(alpha, (r, r))[:, :, None]
|
49 |
|
50 |
-
|
51 |
-
|
52 |
|
53 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
54 |
blurred_B = blurred_B1A / ((1 - blurred_alpha) + 1e-5)
|
55 |
-
|
56 |
-
|
57 |
-
|
58 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
59 |
|
60 |
|
61 |
class ImagePreprocessor():
|
@@ -163,7 +228,7 @@ def predict(images, resolution, weights_file):
|
|
163 |
|
164 |
# Show Results
|
165 |
pred_pil = transforms.ToPILImage()(pred)
|
166 |
-
image_masked = refine_foreground(image, pred_pil)
|
167 |
image_masked.putalpha(pred_pil.resize(image.size))
|
168 |
|
169 |
torch.cuda.empty_cache()
|
|
|
11 |
from PIL import Image
|
12 |
# from gradio_imageslider import ImageSlider
|
13 |
import transformers
|
14 |
+
import torch
|
15 |
from torchvision import transforms
|
16 |
|
17 |
import requests
|
|
|
24 |
|
25 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
26 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
27 |
|
28 |
+
## CPU version refinement
|
29 |
+
def FB_blur_fusion_foreground_estimator_cpu(image, FG, B, alpha, r=90):
|
30 |
+
if isinstance(image, Image.Image):
|
31 |
+
image = np.array(image) / 255.0
|
32 |
+
blurred_alpha = cv2.blur(alpha, (r, r))[:, :, None]
|
33 |
+
|
34 |
+
blurred_FGA = cv2.blur(FG * alpha, (r, r))
|
35 |
+
blurred_FG = blurred_FGA / (blurred_alpha + 1e-5)
|
36 |
+
|
37 |
+
blurred_B1A = cv2.blur(B * (1 - alpha), (r, r))
|
38 |
+
blurred_B = blurred_B1A / ((1 - blurred_alpha) + 1e-5)
|
39 |
+
FG = blurred_FG + alpha * (image - alpha * blurred_FG - (1 - alpha) * blurred_B)
|
40 |
+
FG = np.clip(FG, 0, 1)
|
41 |
+
return FG, blurred_B
|
42 |
|
43 |
+
|
44 |
+
def FB_blur_fusion_foreground_estimator_cpu_2(image, alpha, r=90):
|
45 |
# Thanks to the source: https://github.com/Photoroom/fast-foreground-estimation
|
46 |
alpha = alpha[:, :, None]
|
47 |
+
FG, blur_B = FB_blur_fusion_foreground_estimator_cpu(image, image, image, alpha, r)
|
48 |
+
return FB_blur_fusion_foreground_estimator_cpu(image, FG, blur_B, alpha, r=6)[0]
|
49 |
+
|
50 |
+
|
51 |
+
## GPU version refinement
|
52 |
+
def mean_blur(x, kernel_size):
|
53 |
+
"""
|
54 |
+
equivalent to cv.blur
|
55 |
+
x: [B, C, H, W]
|
56 |
+
"""
|
57 |
+
if kernel_size % 2 == 0:
|
58 |
+
pad_l = kernel_size // 2 - 1
|
59 |
+
pad_r = kernel_size // 2
|
60 |
+
pad_t = kernel_size // 2 - 1
|
61 |
+
pad_b = kernel_size // 2
|
62 |
+
else:
|
63 |
+
pad_l = pad_r = pad_t = pad_b = kernel_size // 2
|
64 |
|
65 |
+
x_padded = torch.nn.functional.pad(x, (pad_l, pad_r, pad_t, pad_b), mode='replicate')
|
66 |
|
67 |
+
return torch.nn.functional.avg_pool2d(x_padded, kernel_size=(kernel_size, kernel_size), stride=1, count_include_pad=False)
|
|
|
|
|
|
|
68 |
|
69 |
+
def FB_blur_fusion_foreground_estimator_gpu(image, FG, B, alpha, r=90):
|
70 |
+
as_dtype = lambda x, dtype: x.to(dtype) if x.dtype != dtype else x
|
71 |
|
72 |
+
input_dtype = image.dtype
|
73 |
+
# convert image to float to avoid overflow
|
74 |
+
image = as_dtype(image, torch.float32)
|
75 |
+
FG = as_dtype(FG, torch.float32)
|
76 |
+
B = as_dtype(B, torch.float32)
|
77 |
+
alpha = as_dtype(alpha, torch.float32)
|
78 |
+
|
79 |
+
blurred_alpha = mean_blur(alpha, kernel_size=r)
|
80 |
+
|
81 |
+
blurred_FGA = mean_blur(FG * alpha, kernel_size=r)
|
82 |
+
blurred_FG = blurred_FGA / (blurred_alpha + 1e-5)
|
83 |
+
|
84 |
+
blurred_B1A = mean_blur(B * (1 - alpha), kernel_size=r)
|
85 |
blurred_B = blurred_B1A / ((1 - blurred_alpha) + 1e-5)
|
86 |
+
|
87 |
+
FG_output = blurred_FG + alpha * (image - alpha * blurred_FG - (1 - alpha) * blurred_B)
|
88 |
+
FG_output = torch.clamp(FG_output, 0, 1)
|
89 |
+
|
90 |
+
return as_dtype(FG_output, input_dtype), as_dtype(blurred_B, input_dtype)
|
91 |
+
|
92 |
+
|
93 |
+
def FB_blur_fusion_foreground_estimator_gpu_2(image, alpha, r=90):
|
94 |
+
# Thanks to the source: https://github.com/ZhengPeng7/BiRefNet/issues/226#issuecomment-3016433728
|
95 |
+
FG, blur_B = FB_blur_fusion_foreground_estimator_gpu(image, image, image, alpha, r)
|
96 |
+
return FB_blur_fusion_foreground_estimator_gpu(image, FG, blur_B, alpha, r=6)[0]
|
97 |
+
|
98 |
+
|
99 |
+
def refine_foreground(image, mask, r=90, device='cuda'):
|
100 |
+
"""both image and mask are in range of [0, 1]"""
|
101 |
+
if mask.size != image.size:
|
102 |
+
mask = mask.resize(image.size)
|
103 |
+
|
104 |
+
if device == 'cuda':
|
105 |
+
image = transforms.functional.to_tensor(image).float().cuda()
|
106 |
+
mask = transforms.functional.to_tensor(mask).float().cuda()
|
107 |
+
image = image.unsqueeze(0)
|
108 |
+
mask = mask.unsqueeze(0)
|
109 |
+
|
110 |
+
estimated_foreground = FB_blur_fusion_foreground_estimator_gpu_2(image, mask, r=r)
|
111 |
+
|
112 |
+
estimated_foreground = estimated_foreground.squeeze()
|
113 |
+
estimated_foreground = (estimated_foreground.mul(255.0)).to(torch.uint8)
|
114 |
+
estimated_foreground = estimated_foreground.permute(1, 2, 0).contiguous().cpu().numpy().astype(np.uint8)
|
115 |
+
else:
|
116 |
+
image = np.array(image, dtype=np.float32) / 255.0
|
117 |
+
mask = np.array(mask, dtype=np.float32) / 255.0
|
118 |
+
estimated_foreground = FB_blur_fusion_foreground_estimator_cpu_2(image, mask, r=r)
|
119 |
+
estimated_foreground = (estimated_foreground * 255.0).astype(np.uint8)
|
120 |
+
|
121 |
+
estimated_foreground = Image.fromarray(np.ascontiguousarray(estimated_foreground))
|
122 |
+
|
123 |
+
return estimated_foreground
|
124 |
|
125 |
|
126 |
class ImagePreprocessor():
|
|
|
228 |
|
229 |
# Show Results
|
230 |
pred_pil = transforms.ToPILImage()(pred)
|
231 |
+
image_masked = refine_foreground(image, pred_pil, device=device)
|
232 |
image_masked.putalpha(pred_pil.resize(image.size))
|
233 |
|
234 |
torch.cuda.empty_cache()
|