Caroline Mai Chan
commited on
Commit
·
41bee7b
1
Parent(s):
72026fd
add model to code
Browse files
app.py
CHANGED
@@ -3,8 +3,69 @@ import torch
|
|
3 |
import torch.nn as nn
|
4 |
import gradio as gr
|
5 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
6 |
|
7 |
def sepia(input_img):
|
|
|
8 |
sepia_filter = np.array(
|
9 |
[[0.393, 0.769, 0.189], [0.349, 0.686, 0.168], [0.272, 0.534, 0.131]]
|
10 |
)
|
|
|
3 |
import torch.nn as nn
|
4 |
import gradio as gr
|
5 |
|
6 |
+
class Generator(nn.Module):
|
7 |
+
def __init__(self, input_nc, output_nc, n_residual_blocks=9, sigmoid=True):
|
8 |
+
super(Generator, self).__init__()
|
9 |
+
|
10 |
+
# Initial convolution block
|
11 |
+
model0 = [ nn.ReflectionPad2d(3),
|
12 |
+
nn.Conv2d(input_nc, 64, 7),
|
13 |
+
norm_layer(64),
|
14 |
+
nn.ReLU(inplace=True) ]
|
15 |
+
self.model0 = nn.Sequential(*model0)
|
16 |
+
|
17 |
+
# Downsampling
|
18 |
+
model1 = []
|
19 |
+
in_features = 64
|
20 |
+
out_features = in_features*2
|
21 |
+
for _ in range(2):
|
22 |
+
model1 += [ nn.Conv2d(in_features, out_features, 3, stride=2, padding=1),
|
23 |
+
norm_layer(out_features),
|
24 |
+
nn.ReLU(inplace=True) ]
|
25 |
+
in_features = out_features
|
26 |
+
out_features = in_features*2
|
27 |
+
self.model1 = nn.Sequential(*model1)
|
28 |
+
|
29 |
+
model2 = []
|
30 |
+
# Residual blocks
|
31 |
+
for _ in range(n_residual_blocks):
|
32 |
+
model2 += [ResidualBlock(in_features)]
|
33 |
+
self.model2 = nn.Sequential(*model2)
|
34 |
+
|
35 |
+
# Upsampling
|
36 |
+
model3 = []
|
37 |
+
out_features = in_features//2
|
38 |
+
for _ in range(2):
|
39 |
+
model3 += [ nn.ConvTranspose2d(in_features, out_features, 3, stride=2, padding=1, output_padding=1),
|
40 |
+
norm_layer(out_features),
|
41 |
+
nn.ReLU(inplace=True) ]
|
42 |
+
in_features = out_features
|
43 |
+
out_features = in_features//2
|
44 |
+
self.model3 = nn.Sequential(*model3)
|
45 |
+
|
46 |
+
# Output layer
|
47 |
+
model4 = [ nn.ReflectionPad2d(3),
|
48 |
+
nn.Conv2d(64, output_nc, 7)]
|
49 |
+
if sigmoid:
|
50 |
+
model4 += [nn.Sigmoid()]
|
51 |
+
|
52 |
+
self.model4 = nn.Sequential(*model4)
|
53 |
+
|
54 |
+
def forward(self, x, cond=None):
|
55 |
+
out = self.model0(x)
|
56 |
+
out = self.model1(out)
|
57 |
+
out = self.model2(out)
|
58 |
+
out = self.model3(out)
|
59 |
+
out = self.model4(out)
|
60 |
+
|
61 |
+
return out
|
62 |
+
|
63 |
+
# model = Generator(3, 1, 3)
|
64 |
+
# model.load_state_dict(torch.load('model.pth'))
|
65 |
+
# model.eval()
|
66 |
|
67 |
def sepia(input_img):
|
68 |
+
print(input_img.shape, np.max(input_img), np.min(input_img))
|
69 |
sepia_filter = np.array(
|
70 |
[[0.393, 0.769, 0.189], [0.349, 0.686, 0.168], [0.272, 0.534, 0.131]]
|
71 |
)
|