Spaces:
Running
Running
HaoFeng2019
commited on
Commit
•
4cbf82a
1
Parent(s):
1b147a3
Update app.py
Browse files
app.py
CHANGED
@@ -14,11 +14,13 @@ import os
|
|
14 |
from PIL import Image
|
15 |
import argparse
|
16 |
import warnings
|
|
|
17 |
warnings.filterwarnings('ignore')
|
18 |
|
19 |
import gradio as gr
|
20 |
|
21 |
-
example_img_list = ['51_1 copy.png','48_2 copy.png','25.jpg']
|
|
|
22 |
|
23 |
def reload_model(model, path=""):
|
24 |
if not bool(path):
|
@@ -49,51 +51,54 @@ def reload_segmodel(model, path=""):
|
|
49 |
|
50 |
return model
|
51 |
|
|
|
52 |
class GeoTr_Seg(nn.Module):
|
53 |
def __init__(self):
|
54 |
super(GeoTr_Seg, self).__init__()
|
55 |
self.msk = U2NETP(3, 1)
|
56 |
self.GeoTr = GeoTr(num_attn_layers=6)
|
57 |
-
|
58 |
def forward(self, x):
|
59 |
-
msk, _1,_2,_3,_4,_5,_6 = self.msk(x)
|
60 |
msk = (msk > 0.5).float()
|
61 |
x = msk * x
|
62 |
|
63 |
bm = self.GeoTr(x)
|
64 |
bm = (2 * (bm / 286.8) - 1) * 0.99
|
65 |
-
|
66 |
return bm
|
67 |
|
68 |
|
69 |
# Initialize models
|
70 |
GeoTr_Seg_model = GeoTr_Seg()
|
71 |
-
#IllTr_model = IllTr()
|
72 |
|
73 |
# Load models only once
|
74 |
reload_segmodel(GeoTr_Seg_model.msk, './model_pretrained/seg.pth')
|
75 |
reload_model(GeoTr_Seg_model.GeoTr, './model_pretrained/geotr.pth')
|
76 |
-
#reload_model(IllTr_model, './model_pretrained/illtr.pth')
|
77 |
|
78 |
# Compile models (assuming PyTorch 2.0)
|
79 |
GeoTr_Seg_model = torch.compile(GeoTr_Seg_model)
|
80 |
-
|
|
|
|
|
81 |
|
82 |
def process_image(input_image):
|
83 |
GeoTr_Seg_model.eval()
|
|
|
84 |
|
85 |
im_ori = np.array(input_image)[:, :, :3] / 255.
|
86 |
h, w, _ = im_ori.shape
|
87 |
-
|
88 |
-
im = cv2.resize(im_ori, (288, new_height))
|
89 |
im = im.transpose(2, 0, 1)
|
90 |
im = torch.from_numpy(im).float().unsqueeze(0)
|
91 |
|
92 |
with torch.no_grad():
|
93 |
bm = GeoTr_Seg_model(im)
|
94 |
bm = bm.cpu()
|
95 |
-
bm0 = cv2.resize(bm[0, 0].numpy(), (
|
96 |
-
bm1 = cv2.resize(bm[0, 1].numpy(), (
|
97 |
bm0 = cv2.blur(bm0, (3, 3))
|
98 |
bm1 = cv2.blur(bm1, (3, 3))
|
99 |
lbl = torch.from_numpy(np.stack([bm0, bm1], axis=2)).unsqueeze(0)
|
@@ -114,5 +119,6 @@ def process_image(input_image):
|
|
114 |
input_image = gr.inputs.Image()
|
115 |
output_image = gr.outputs.Image(type='pil')
|
116 |
|
117 |
-
iface = gr.Interface(fn=process_image, inputs=input_image, outputs=output_image, title="DocTr",
|
118 |
-
|
|
|
|
14 |
from PIL import Image
|
15 |
import argparse
|
16 |
import warnings
|
17 |
+
|
18 |
warnings.filterwarnings('ignore')
|
19 |
|
20 |
import gradio as gr
|
21 |
|
22 |
+
example_img_list = ['51_1 copy.png', '48_2 copy.png', '25.jpg']
|
23 |
+
|
24 |
|
25 |
def reload_model(model, path=""):
|
26 |
if not bool(path):
|
|
|
51 |
|
52 |
return model
|
53 |
|
54 |
+
|
55 |
class GeoTr_Seg(nn.Module):
|
56 |
def __init__(self):
|
57 |
super(GeoTr_Seg, self).__init__()
|
58 |
self.msk = U2NETP(3, 1)
|
59 |
self.GeoTr = GeoTr(num_attn_layers=6)
|
60 |
+
|
61 |
def forward(self, x):
|
62 |
+
msk, _1, _2, _3, _4, _5, _6 = self.msk(x)
|
63 |
msk = (msk > 0.5).float()
|
64 |
x = msk * x
|
65 |
|
66 |
bm = self.GeoTr(x)
|
67 |
bm = (2 * (bm / 286.8) - 1) * 0.99
|
68 |
+
|
69 |
return bm
|
70 |
|
71 |
|
72 |
# Initialize models
|
73 |
GeoTr_Seg_model = GeoTr_Seg()
|
74 |
+
# IllTr_model = IllTr()
|
75 |
|
76 |
# Load models only once
|
77 |
reload_segmodel(GeoTr_Seg_model.msk, './model_pretrained/seg.pth')
|
78 |
reload_model(GeoTr_Seg_model.GeoTr, './model_pretrained/geotr.pth')
|
79 |
+
# reload_model(IllTr_model, './model_pretrained/illtr.pth')
|
80 |
|
81 |
# Compile models (assuming PyTorch 2.0)
|
82 |
GeoTr_Seg_model = torch.compile(GeoTr_Seg_model)
|
83 |
+
|
84 |
+
|
85 |
+
# IllTr_model = torch.compile(IllTr_model)
|
86 |
|
87 |
def process_image(input_image):
|
88 |
GeoTr_Seg_model.eval()
|
89 |
+
# IllTr_model.eval()
|
90 |
|
91 |
im_ori = np.array(input_image)[:, :, :3] / 255.
|
92 |
h, w, _ = im_ori.shape
|
93 |
+
im = cv2.resize(im_ori, (288, 288))
|
|
|
94 |
im = im.transpose(2, 0, 1)
|
95 |
im = torch.from_numpy(im).float().unsqueeze(0)
|
96 |
|
97 |
with torch.no_grad():
|
98 |
bm = GeoTr_Seg_model(im)
|
99 |
bm = bm.cpu()
|
100 |
+
bm0 = cv2.resize(bm[0, 0].numpy(), (w, h))
|
101 |
+
bm1 = cv2.resize(bm[0, 1].numpy(), (w, h))
|
102 |
bm0 = cv2.blur(bm0, (3, 3))
|
103 |
bm1 = cv2.blur(bm1, (3, 3))
|
104 |
lbl = torch.from_numpy(np.stack([bm0, bm1], axis=2)).unsqueeze(0)
|
|
|
119 |
input_image = gr.inputs.Image()
|
120 |
output_image = gr.outputs.Image(type='pil')
|
121 |
|
122 |
+
iface = gr.Interface(fn=process_image, inputs=input_image, outputs=output_image, title="DocTr",
|
123 |
+
examples=example_img_list)
|
124 |
+
iface.launch()
|