Spaces:
Runtime error
Runtime error
mehdiabruee
commited on
Commit
·
7e0282a
1
Parent(s):
6041fbb
Upload cyclegan_inference.py
Browse files- cyclegan_inference.py +168 -0
cyclegan_inference.py
ADDED
@@ -0,0 +1,168 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# -*- coding: utf-8 -*-
|
2 |
+
"""cyclegan_inference.ipynb
|
3 |
+
|
4 |
+
Automatically generated by Colaboratory.
|
5 |
+
|
6 |
+
Original file is located at
|
7 |
+
https://colab.research.google.com/drive/12lelsBZXqNOe7xaXI724rEHAbppRt07y
|
8 |
+
"""
|
9 |
+
|
10 |
+
import gradio as gr
|
11 |
+
import torch
|
12 |
+
import torchvision
|
13 |
+
from torch import nn
|
14 |
+
from typing import List
|
15 |
+
|
16 |
+
def ifnone(a, b): # a fastai-specific (fastcore) function used below, redefined so it's independent
|
17 |
+
"`b` if `a` is None else `a`"
|
18 |
+
return b if a is None else a
|
19 |
+
|
20 |
+
class ConvBlock(torch.nn.Module):
|
21 |
+
def __init__(self,input_size,output_size,kernel_size=4,stride=2,padding=1,activation='relu',batch_norm=True):
|
22 |
+
super(ConvBlock,self).__init__()
|
23 |
+
self.conv = torch.nn.Conv2d(input_size,output_size,kernel_size,stride,padding)
|
24 |
+
self.batch_norm = batch_norm
|
25 |
+
self.bn = torch.nn.InstanceNorm2d(output_size)
|
26 |
+
self.activation = activation
|
27 |
+
self.relu = torch.nn.ReLU(True)
|
28 |
+
self.lrelu = torch.nn.LeakyReLU(0.2,True)
|
29 |
+
self.tanh = torch.nn.Tanh()
|
30 |
+
self.sigmoid = torch.nn.Sigmoid()
|
31 |
+
def forward(self,x):
|
32 |
+
if self.batch_norm:
|
33 |
+
out = self.bn(self.conv(x))
|
34 |
+
else:
|
35 |
+
out = self.conv(x)
|
36 |
+
|
37 |
+
if self.activation == 'relu':
|
38 |
+
return self.relu(out)
|
39 |
+
elif self.activation == 'lrelu':
|
40 |
+
return self.lrelu(out)
|
41 |
+
elif self.activation == 'tanh':
|
42 |
+
return self.tanh(out)
|
43 |
+
elif self.activation == 'no_act':
|
44 |
+
return out
|
45 |
+
elif self.activation =='sigmoid':
|
46 |
+
return self.sigmoid(out)
|
47 |
+
|
48 |
+
|
49 |
+
class ResnetBlock(torch.nn.Module):
|
50 |
+
def __init__(self,num_filter,kernel_size=3,stride=1,padding=0):
|
51 |
+
super(ResnetBlock,self).__init__()
|
52 |
+
conv1 = torch.nn.Conv2d(num_filter,num_filter,kernel_size,stride,padding)
|
53 |
+
conv2 = torch.nn.Conv2d(num_filter,num_filter,kernel_size,stride,padding)
|
54 |
+
bn = torch.nn.InstanceNorm2d(num_filter)
|
55 |
+
relu = torch.nn.ReLU(True)
|
56 |
+
pad = torch.nn.ReflectionPad2d(1)
|
57 |
+
|
58 |
+
self.resnet_block = torch.nn.Sequential(
|
59 |
+
pad,
|
60 |
+
conv1,
|
61 |
+
bn,
|
62 |
+
relu,
|
63 |
+
pad,
|
64 |
+
conv2,
|
65 |
+
bn
|
66 |
+
)
|
67 |
+
def forward(self,x):
|
68 |
+
out = self.resnet_block(x)
|
69 |
+
return out
|
70 |
+
def resnet_generator(ch_in:int, ch_out:int, n_ftrs:int=64, norm_layer:nn.Module=None,
|
71 |
+
dropout:float=0., n_blocks:int=9, pad_mode:str='reflection')->nn.Module:
|
72 |
+
norm_layer = ifnone(norm_layer, nn.InstanceNorm2d)
|
73 |
+
bias = (norm_layer == nn.InstanceNorm2d)
|
74 |
+
layers = pad_conv_norm_relu(ch_in, n_ftrs, 'reflection', norm_layer, pad=3, ks=7, bias=bias)
|
75 |
+
for i in range(2):
|
76 |
+
layers += pad_conv_norm_relu(n_ftrs, n_ftrs *2, 'zeros', norm_layer, stride=2, bias=bias)
|
77 |
+
n_ftrs *= 2
|
78 |
+
layers += [ResnetBlock(n_ftrs, pad_mode, norm_layer, dropout, bias) for _ in range(n_blocks)]
|
79 |
+
for i in range(2):
|
80 |
+
layers += convT_norm_relu(n_ftrs, n_ftrs//2, norm_layer, bias=bias)
|
81 |
+
n_ftrs //= 2
|
82 |
+
layers += [nn.ReflectionPad2d(3), nn.Conv2d(n_ftrs, ch_out, kernel_size=7, padding=0), nn.Tanh()]
|
83 |
+
return nn.Sequential(*layers)
|
84 |
+
|
85 |
+
class DeconvBlock(torch.nn.Module):
|
86 |
+
def __init__(self,input_size,output_size,kernel_size=4,stride=2,padding=1,activation='relu',batch_norm=True):
|
87 |
+
super(DeconvBlock,self).__init__()
|
88 |
+
self.deconv = torch.nn.ConvTranspose2d(input_size,output_size,kernel_size,stride,padding)
|
89 |
+
self.batch_norm = batch_norm
|
90 |
+
self.bn = torch.nn.InstanceNorm2d(output_size)
|
91 |
+
self.activation = activation
|
92 |
+
self.relu = torch.nn.ReLU(True)
|
93 |
+
self.tanh = torch.nn.Tanh()
|
94 |
+
def forward(self,x):
|
95 |
+
if self.batch_norm:
|
96 |
+
out = self.bn(self.deconv(x))
|
97 |
+
else:
|
98 |
+
out = self.deconv(x)
|
99 |
+
if self.activation == 'relu':
|
100 |
+
return self.relu(out)
|
101 |
+
elif self.activation == 'lrelu':
|
102 |
+
return self.lrelu(out)
|
103 |
+
elif self.activation == 'tanh':
|
104 |
+
return self.tanh(out)
|
105 |
+
elif self.activation == 'no_act':
|
106 |
+
return out
|
107 |
+
|
108 |
+
class Generator(torch.nn.Module):
|
109 |
+
def __init__(self,input_dim,num_filter,output_dim,num_resnet):
|
110 |
+
super(Generator,self).__init__()
|
111 |
+
|
112 |
+
#Reflection padding
|
113 |
+
#self.pad = torch.nn.ReflectionPad2d(3)
|
114 |
+
#Encoder
|
115 |
+
self.conv1 = ConvBlock(input_dim,num_filter,kernel_size=4,stride=2,padding=1)
|
116 |
+
self.conv2 = ConvBlock(num_filter,num_filter*2)
|
117 |
+
#self.conv3 = ConvBlock(num_filter*2,num_filter*4)
|
118 |
+
#Resnet blocks
|
119 |
+
self.resnet_blocks = []
|
120 |
+
for i in range(num_resnet):
|
121 |
+
self.resnet_blocks.append(ResnetBlock(num_filter*2))
|
122 |
+
self.resnet_blocks = torch.nn.Sequential(*self.resnet_blocks)
|
123 |
+
#Decoder
|
124 |
+
self.deconv1 = DeconvBlock(num_filter*2,num_filter)
|
125 |
+
self.deconv2 = DeconvBlock(num_filter,output_dim,activation='tanh')
|
126 |
+
#self.deconv3 = ConvBlock(num_filter,output_dim,kernel_size=7,stride=1,padding=0,activation='tanh',batch_norm=False)
|
127 |
+
|
128 |
+
def forward(self,x):
|
129 |
+
#Encoder
|
130 |
+
enc1 = self.conv1(x)
|
131 |
+
enc2 = self.conv2(enc1)
|
132 |
+
#enc3 = self.conv3(enc2)
|
133 |
+
#Resnet blocks
|
134 |
+
res = self.resnet_blocks(enc2)
|
135 |
+
#Decoder
|
136 |
+
dec1 = self.deconv1(res)
|
137 |
+
dec2 = self.deconv2(dec1)
|
138 |
+
#out = self.deconv3(self.pad(dec2))
|
139 |
+
return dec2
|
140 |
+
|
141 |
+
def normal_weight_init(self,mean=0.0,std=0.02):
|
142 |
+
for m in self.children():
|
143 |
+
if isinstance(m,ConvBlock):
|
144 |
+
torch.nn.init.normal_(m.conv.weight,mean,std)
|
145 |
+
if isinstance(m,DeconvBlock):
|
146 |
+
torch.nn.init.normal_(m.deconv.weight,mean,std)
|
147 |
+
if isinstance(m,ResnetBlock):
|
148 |
+
torch.nn.init.normal_(m.conv.weight,mean,std)
|
149 |
+
torch.nn.init.constant_(m.conv.bias,0)
|
150 |
+
|
151 |
+
model = G_A = Generator(3, 32, 3, 4).cuda() # input_dim, num_filter, output_dim, num_resnet
|
152 |
+
model.load_state_dict(torch.load('G_A_HW4_SAVE.pt',map_location=torch.device('cpu')))
|
153 |
+
model.eval()
|
154 |
+
|
155 |
+
|
156 |
+
totensor = torchvision.transforms.ToTensor()
|
157 |
+
normalize_fn = torchvision.transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
|
158 |
+
topilimage = torchvision.transforms.ToPILImage()
|
159 |
+
|
160 |
+
def predict(input):
|
161 |
+
im = normalize_fn(totensor(input))
|
162 |
+
print(im.shape)
|
163 |
+
preds = model(im.unsqueeze(0))/2 + 0.5
|
164 |
+
print(preds.shape)
|
165 |
+
return topilimage(preds.squeeze(0).detach())
|
166 |
+
|
167 |
+
gr_interface = gr.Interface(fn=predict, inputs=gr.inputs.Image(shape=(256, 256)), outputs="image", title='Horse-to-Zebra CycleGAN')
|
168 |
+
gr_interface.launch(inline=False,share=False)
|