Spaces:
Runtime error
Runtime error
Update app.py
Browse files
app.py
CHANGED
@@ -1,7 +1,145 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import gradio as gr
|
2 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
3 |
def main(image, backbone, style):
|
|
|
|
|
4 |
isize = image.size
|
|
|
5 |
s = f"The output image ({str(image.size)}) is processed by {backbone} based on input image ({str(isize)}) . <br> Please <b>rate</b> the generated image through the <b>Flag</b> button below!"
|
6 |
return image, s
|
7 |
|
@@ -20,6 +158,7 @@ gr.Interface(
|
|
20 |
# []
|
21 |
# ],
|
22 |
# live = True, # the interface will recalculate as soon as the user input changes.
|
|
|
23 |
flagging_options = ["Excellect", "Moderate", "Bad"],
|
24 |
flagging_dir = "flagged",
|
25 |
allow_screenshot = False,
|
|
|
1 |
+
from huggingface_hub import hf_hub_download
|
2 |
+
|
3 |
+
|
4 |
+
Rain_Princess = hf_hub_download(repo_id="maze/FastStyleTransfer", filename="Rain_Princess.pth")
|
5 |
+
#modelarcanev3 = hf_hub_download(repo_id="akhaliq/ArcaneGANv0.3", filename="ArcaneGANv0.3.jit")
|
6 |
+
#modelarcanev2 = hf_hub_download(repo_id="akhaliq/ArcaneGANv0.2", filename="ArcaneGANv0.2.jit")
|
7 |
+
|
8 |
+
|
9 |
+
import numpy as np
|
10 |
+
from PIL import Image
|
11 |
import gradio as gr
|
12 |
|
13 |
+
import torch
|
14 |
+
import torch.nn as nn
|
15 |
+
|
16 |
+
import torchvision.transforms as transforms
|
17 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
18 |
+
|
19 |
+
|
20 |
+
class TransformerNetwork(nn.Module):
|
21 |
+
def __init__(self, tanh_multiplier=None):
|
22 |
+
super(TransformerNetwork, self).__init__()
|
23 |
+
self.ConvBlock = nn.Sequential(
|
24 |
+
ConvLayer(3, 32, 9, 1),
|
25 |
+
nn.ReLU(),
|
26 |
+
ConvLayer(32, 64, 3, 2),
|
27 |
+
nn.ReLU(),
|
28 |
+
ConvLayer(64, 128, 3, 2),
|
29 |
+
nn.ReLU()
|
30 |
+
)
|
31 |
+
self.ResidualBlock = nn.Sequential(
|
32 |
+
ResidualLayer(128, 3),
|
33 |
+
ResidualLayer(128, 3),
|
34 |
+
ResidualLayer(128, 3),
|
35 |
+
ResidualLayer(128, 3),
|
36 |
+
ResidualLayer(128, 3)
|
37 |
+
)
|
38 |
+
self.DeconvBlock = nn.Sequential(
|
39 |
+
DeconvLayer(128, 64, 3, 2, 1),
|
40 |
+
nn.ReLU(),
|
41 |
+
DeconvLayer(64, 32, 3, 2, 1),
|
42 |
+
nn.ReLU(),
|
43 |
+
ConvLayer(32, 3, 9, 1, norm="None")
|
44 |
+
)
|
45 |
+
self.tanh_multiplier = tanh_multiplier
|
46 |
+
|
47 |
+
def forward(self, x):
|
48 |
+
x = self.ConvBlock(x)
|
49 |
+
x = self.ResidualBlock(x)
|
50 |
+
x = self.DeconvBlock(x)
|
51 |
+
if isinstance(self.tanh_multiplier, int):
|
52 |
+
x = self.tanh_multiplier * F.tanh(x)
|
53 |
+
return x
|
54 |
+
|
55 |
+
class ConvLayer(nn.Module):
|
56 |
+
def __init__(self, in_channels, out_channels, kernel_size, stride, norm="instance"):
|
57 |
+
super(ConvLayer, self).__init__()
|
58 |
+
padding_size = kernel_size // 2
|
59 |
+
self.pad = nn.ReflectionPad2d(padding_size)
|
60 |
+
self.conv = nn.Conv2d(in_channels, out_channels, kernel_size, stride)
|
61 |
+
if norm == "instance":
|
62 |
+
self.norm = nn.InstanceNorm2d(out_channels, affine=True)
|
63 |
+
elif norm == "batch":
|
64 |
+
self.norm = nn.BatchNorm2d(out_channels, affine=True)
|
65 |
+
else:
|
66 |
+
self.norm = nn.Identity()
|
67 |
+
|
68 |
+
def forward(self, x):
|
69 |
+
x = self.pad(x)
|
70 |
+
x = self.conv(x)
|
71 |
+
x = self.norm(x)
|
72 |
+
return x
|
73 |
+
|
74 |
+
class ResidualLayer(nn.Module):
|
75 |
+
def __init__(self, channels=128, kernel_size=3):
|
76 |
+
super(ResidualLayer, self).__init__()
|
77 |
+
self.conv1 = ConvLayer(channels, channels, kernel_size, stride=1)
|
78 |
+
self.relu = nn.ReLU()
|
79 |
+
self.conv2 = ConvLayer(channels, channels, kernel_size, stride=1)
|
80 |
+
|
81 |
+
def forward(self, x):
|
82 |
+
identity = x
|
83 |
+
out = self.relu(self.conv1(x))
|
84 |
+
out = self.conv2(out)
|
85 |
+
out = out + identity
|
86 |
+
return out
|
87 |
+
|
88 |
+
class DeconvLayer(nn.Module):
|
89 |
+
def __init__(self, in_channels, out_channels, kernel_size, stride, output_padding, norm="instance"):
|
90 |
+
super(DeconvLayer, self).__init__()
|
91 |
+
|
92 |
+
padding_size = kernel_size // 2
|
93 |
+
self.conv_transpose = nn.ConvTranspose2d(in_channels, out_channels, kernel_size, stride, padding_size, output_padding)
|
94 |
+
if norm == "instance":
|
95 |
+
self.norm = nn.InstanceNorm2d(out_channels, affine=True)
|
96 |
+
elif norm == "batch":
|
97 |
+
self.norm = nn.BatchNorm2d(out_channels, affine=True)
|
98 |
+
else:
|
99 |
+
self.norm = nn.Identity()
|
100 |
+
|
101 |
+
def forward(self, x):
|
102 |
+
x = self.conv_transpose(x)
|
103 |
+
out = self.norm(x)
|
104 |
+
return out
|
105 |
+
|
106 |
+
|
107 |
+
mean = np.array([0.485, 0.456, 0.406])
|
108 |
+
std = np.array([0.229, 0.224, 0.225])
|
109 |
+
|
110 |
+
transformer = TransformerNetwork().to(device)
|
111 |
+
|
112 |
+
transformer.eval()
|
113 |
+
|
114 |
+
transform = transforms.Compose([
|
115 |
+
transforms.Resize(256),
|
116 |
+
transforms.ToTensor(),
|
117 |
+
transforms.Normalize(mean, std),
|
118 |
+
])
|
119 |
+
|
120 |
+
denormalize = transforms.Normalize(
|
121 |
+
mean= [-m/s for m, s in zip(mean, std)],
|
122 |
+
std= [1/s for s in std]
|
123 |
+
)
|
124 |
+
tensor2Image = transforms.ToPILImage()
|
125 |
+
|
126 |
+
@torch.no_grad()
|
127 |
+
def process(image, model):
|
128 |
+
image = transform(image).to(device)
|
129 |
+
image = image.unsqueeze(dim=0)
|
130 |
+
|
131 |
+
image = denormalize(model(image)).cpu()
|
132 |
+
image = torch.clamp(image.squeeze(dim=0), 0, 1)
|
133 |
+
image = tensor2Image(image)
|
134 |
+
|
135 |
+
return image
|
136 |
+
|
137 |
+
|
138 |
def main(image, backbone, style):
|
139 |
+
transformer.load_state_dict(torch.load(Rain_Princess))
|
140 |
+
image = Image.fromarray(image)
|
141 |
isize = image.size
|
142 |
+
image = process(image, transformer)
|
143 |
s = f"The output image ({str(image.size)}) is processed by {backbone} based on input image ({str(isize)}) . <br> Please <b>rate</b> the generated image through the <b>Flag</b> button below!"
|
144 |
return image, s
|
145 |
|
|
|
158 |
# []
|
159 |
# ],
|
160 |
# live = True, # the interface will recalculate as soon as the user input changes.
|
161 |
+
allow_flagging = "manual",
|
162 |
flagging_options = ["Excellect", "Moderate", "Bad"],
|
163 |
flagging_dir = "flagged",
|
164 |
allow_screenshot = False,
|