-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathgenerator.py
188 lines (159 loc) · 8.88 KB
/
generator.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
import torch
from torch import nn
import torch.nn.functional as F
from functions import weights_init, get_norm_layer
def define_G(input_nc, output_nc, ngf, netG, n_downsample_global=3, n_blocks_global=9, n_local_enhancers=1,
n_blocks_local=3, norm='instance', gpu_ids=[]):
norm_layer = get_norm_layer(norm_type=norm)
if netG == 'global':
netG = GlobalGenerator(input_nc, output_nc, ngf, n_downsample_global, n_blocks_global, norm_layer)
elif netG == 'local':
netG = LocalEnhancer(input_nc, output_nc, ngf, n_downsample_global, n_blocks_global,
n_local_enhancers, n_blocks_local, norm_layer)
elif netG == 'encoder':
netG = Encoder(input_nc, output_nc, ngf, n_downsample_global, norm_layer)
else:
raise('generator not implemented!')
print(netG)
if len(gpu_ids) > 0:
assert(torch.cuda.is_available())
netG.cuda(gpu_ids[0])
netG.apply(weights_init)
return netG
# Define a resnet block
class ResnetBlock(nn.Module):
def __init__(self, dim, padding_type, norm_layer, activation=nn.ReLU(True), use_dropout=False):
super(ResnetBlock, self).__init__()
self.conv_block = self.build_conv_block(dim, padding_type, norm_layer, activation, use_dropout)
def build_conv_block(self, dim, padding_type, norm_layer, activation, use_dropout):
conv_block = []
p = 0
if padding_type == 'reflect':
conv_block += [nn.ReflectionPad2d(1)]
elif padding_type == 'replicate':
conv_block += [nn.ReplicationPad2d(1)]
elif padding_type == 'zero':
p = 1
else:
raise NotImplementedError('padding [%s] is not implemented' % padding_type)
conv_block += [nn.Conv2d(dim, dim, kernel_size=3, padding=p),
norm_layer(dim),
activation]
if use_dropout:
conv_block += [nn.Dropout(0.5)]
p = 0
if padding_type == 'reflect':
conv_block += [nn.ReflectionPad2d(1)]
elif padding_type == 'replicate':
conv_block += [nn.ReplicationPad2d(1)]
elif padding_type == 'zero':
p = 1
else:
raise NotImplementedError('padding [%s] is not implemented' % padding_type)
conv_block += [nn.Conv2d(dim, dim, kernel_size=3, padding=p),
norm_layer(dim)]
return nn.Sequential(*conv_block)
def forward(self, x):
out = x + self.conv_block(x)
return out
class GlobalGenerator(nn.Module):
def __init__(self, input_nc, output_nc, ngf=64, n_downsampling=3, n_blocks=9, norm_layer=nn.BatchNorm2d,
padding_type='reflect'):
assert(n_blocks >= 0)
super(GlobalGenerator, self).__init__()
activation = nn.ReLU(True)
model = [nn.ReflectionPad2d(3), nn.Conv2d(input_nc, ngf, kernel_size=7, padding=0), norm_layer(ngf), activation]
### downsample
for i in range(n_downsampling):
mult = 2**i
model += [nn.Conv2d(ngf * mult, ngf * mult * 2, kernel_size=3, stride=2, padding=1),
norm_layer(ngf * mult * 2), activation]
### resnet blocks
mult = 2**n_downsampling
for i in range(n_blocks):
model += [ResnetBlock(ngf * mult, padding_type=padding_type, activation=activation, norm_layer=norm_layer)]
### upsample
for i in range(n_downsampling):
mult = 2**(n_downsampling - i)
model += [nn.ConvTranspose2d(ngf * mult, int(ngf * mult / 2), kernel_size=3, stride=2, padding=1, output_padding=1),
norm_layer(int(ngf * mult / 2)), activation]
model += [nn.ReflectionPad2d(3), nn.Conv2d(ngf, output_nc, kernel_size=7, padding=0), nn.Tanh()]
self.model = nn.Sequential(*model)
def forward(self, input):
return self.model(input)
class LocalEnhancer(nn.Module):
def __init__(self, input_nc, output_nc, ngf=32, n_downsample_global=3, n_blocks_global=9,
n_local_enhancers=1, n_blocks_local=3, norm_layer=nn.BatchNorm2d, padding_type='reflect'):
super(LocalEnhancer, self).__init__()
self.n_local_enhancers = n_local_enhancers
###### global generator model #####
ngf_global = ngf * (2**n_local_enhancers)
model_global = GlobalGenerator(input_nc, output_nc, ngf_global, n_downsample_global, n_blocks_global, norm_layer).model
model_global = [model_global[i] for i in range(len(model_global)-3)] # get rid of final convolution layers
self.model = nn.Sequential(*model_global)
###### local enhancer layers #####
for n in range(1, n_local_enhancers+1):
### downsample
ngf_global = ngf * (2**(n_local_enhancers-n))
model_downsample = [nn.ReflectionPad2d(3), SNConv(input_nc, ngf_global, kernel_size=7, padding=0),
norm_layer(ngf_global), nn.ReLU(True),
SNConv(ngf_global, ngf_global * 2, kernel_size=3, stride=2, padding=1),
norm_layer(ngf_global * 2), nn.ReLU(True)]
### residual blocks
model_upsample = []
for i in range(n_blocks_local):
model_upsample += [ResnetBlock(ngf_global * 2, padding_type=padding_type, norm_layer=norm_layer)]
### upsample
model_upsample += [nn.ConvTranspose2d(ngf_global * 2, ngf_global, kernel_size=3, stride=2, padding=1, output_padding=1),
norm_layer(ngf_global), nn.ReLU(True)]
### final convolution
if n == n_local_enhancers:
model_upsample += [nn.ReflectionPad2d(3), SNConv(ngf, output_nc, kernel_size=7, padding=0), nn.Tanh()]
setattr(self, 'model'+str(n)+'_1', nn.Sequential(*model_downsample))
setattr(self, 'model'+str(n)+'_2', nn.Sequential(*model_upsample))
self.downsample = nn.AvgPool2d(3, stride=2, padding=[1, 1], count_include_pad=False)
def forward(self, input):
### create input pyramid
input_downsampled = [input]
for i in range(self.n_local_enhancers):
input_downsampled.append(self.downsample(input_downsampled[-1]))
### output at coarest level
output_prev = self.model(input_downsampled[-1])
### build up one layer at a time
for n_local_enhancers in range(1, self.n_local_enhancers+1):
model_downsample = getattr(self, 'model'+str(n_local_enhancers)+'_1')
model_upsample = getattr(self, 'model'+str(n_local_enhancers)+'_2')
input_i = input_downsampled[self.n_local_enhancers-n_local_enhancers]
output_prev = model_upsample(model_downsample(input_i) + output_prev)
return output_prev
class Encoder(nn.Module):
def __init__(self, input_nc, output_nc, ngf=32, n_downsampling=4, norm_layer=nn.BatchNorm2d):
super(Encoder, self).__init__()
self.output_nc = output_nc
model = [nn.ReflectionPad2d(3), SNConv(input_nc, ngf, kernel_size=7, padding=0),
norm_layer(ngf), nn.ReLU(True)]
### downsample
for i in range(n_downsampling):
mult = 2**i
model += [SNConv(ngf * mult, ngf * mult * 2, kernel_size=3, stride=2, padding=1),
norm_layer(ngf * mult * 2), nn.ReLU(True)]
### upsample
for i in range(n_downsampling):
mult = 2**(n_downsampling - i)
model += [nn.ConvTranspose2d(ngf * mult, int(ngf * mult / 2), kernel_size=3, stride=2, padding=1, output_padding=1),
norm_layer(int(ngf * mult / 2)), nn.ReLU(True)]
model += [nn.ReflectionPad2d(3), SNConv(ngf, output_nc, kernel_size=7, padding=0), nn.Tanh()]
self.model = nn.Sequential(*model)
def forward(self, input, inst):
outputs = self.model(input)
# instance-wise average pooling
outputs_mean = outputs.clone()
inst_list = np.unique(inst.cpu().numpy().astype(int))
for i in inst_list:
for b in range(input.size()[0]):
indices = (inst[b:b+1] == int(i)).nonzero() # n x 4
for j in range(self.output_nc):
output_ins = outputs[indices[:,0] + b, indices[:,1] + j, indices[:,2], indices[:,3]]
mean_feat = torch.mean(output_ins).expand_as(output_ins)
outputs_mean[indices[:,0] + b, indices[:,1] + j, indices[:,2], indices[:,3]] = mean_feat
return outputs_mean