-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtransformer_net.py
342 lines (304 loc) · 12.3 KB
/
transformer_net.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
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
import torch
import torch.nn as nn
import numpy as np
import torch.nn.functional as F
from pad import *
class TransformerRNN(torch.nn.Module):
def __init__(self, pad_type="reflect-start", upsample="deconv"):
super(TransformerRNN, self).__init__()
self.pad_type = pad_type
# Initial convolution layers
self.conv1 = ConvLayer(3, 32, kernel_size=9, stride=1, pad_type=pad_type)
self.in1 = nn.InstanceNorm2d(32)
self.conv2 = ConvLayer(32, 64, kernel_size=4, stride=2, pad_type=pad_type)
self.in2 = nn.InstanceNorm2d(64)
self.conv3 = ConvLayer(64, 128, kernel_size=4, stride=2, pad_type=pad_type)
self.in3 = nn.InstanceNorm2d(128)
# Residual layers
self.res1 = ResidualBlock(128, pad_type)
self.res2 = ResidualBlock(128, pad_type)
self.res3 = ResidualBlock(128, pad_type)
self.res4 = ResidualBlock(128, pad_type)
self.res5 = ResidualBlock(128, pad_type)
# Upsampling Layers
self.deconv1 = UpsampleConvLayer(128, 64, 4, 2, pad_type=pad_type, upsample=upsample)
self.in4 = nn.InstanceNorm2d(64)
self.deconv2 = UpsampleConvLayer(64, 32, 4, 2, pad_type=pad_type, upsample=upsample)
self.in5 = nn.InstanceNorm2d(32)
self.deconv3 = ConvLayer(32, 3, kernel_size=9, stride=1, pad_type=pad_type)
# Non-linearities
self.relu = nn.ReLU()
self.setup_pad_input(256)
def pad(self, x):
if self.pad_type == "reflect-start":
return self.pad_input(x)
else:
return x
def setup_pad_input(self, size=256):
x = torch.zeros(1, 3, size, size).float()
y = self.relu(self.in1(self.conv1(x)))
y = self.relu(self.in2(self.conv2(y)))
y = self.relu(self.in3(self.conv3(y)))
y = self.res1(y)
y = self.res2(y)
y = self.res3(y)
y = self.res4(y)
y = self.res5(y)
y = self.relu(self.in4(self.deconv1(y)))
y = self.relu(self.in5(self.deconv2(y)))
y = self.deconv3(y)
diff_h = x.size(2) - y.size(2)
diff_w = x.size(3) - y.size(3)
ph, pw = diff_h // 2, diff_w // 2
self.pad_input = nn.ReflectionPad2d((ph, ph, pw, pw))
print(str(x.size()) + " => " + str(y.size()))
def forward(self, X, prev=None):
"""
Split the batch dim according to time and do rnn unroll.
"""
if prev is None: prev = torch.zeros_like(X[0:1])
out = []
for i in range(X.shape[0]):
y = torch.cat([prev, X[i:i+1]], 1)
y = self.pad(y)
y = self.relu(self.in1(self.conv1(y)))
y = self.relu(self.in2(self.conv2(y)))
y = self.relu(self.in3(self.conv3(y)))
y = self.res1(y)
y = self.res2(y)
y = self.res3(y)
y = self.res4(y)
y = self.res5(y)
y = self.relu(self.in4(self.deconv1(y)))
y = self.relu(self.in5(self.deconv2(y)))
y = self.deconv3(y)
# for evaluation, rnn dynamic BP graph is not maintained.
if not self.training: y = y.detach()
out.append(y)
# the batch dim is recursive number
return torch.cat(out)
class TransformerNet(torch.nn.Module):
def __init__(self, pad_type="reflect-start", upsample="deconv"):
super(TransformerNet, self).__init__()
self.pad_type = pad_type
# Initial convolution layers
self.conv1 = ConvLayer(3, 32, kernel_size=9, stride=1, pad_type=pad_type)
self.in1 = nn.InstanceNorm2d(32)
self.conv2 = ConvLayer(32, 64, kernel_size=4, stride=2, pad_type=pad_type)
self.in2 = nn.InstanceNorm2d(64)
self.conv3 = ConvLayer(64, 128, kernel_size=4, stride=2, pad_type=pad_type)
self.in3 = nn.InstanceNorm2d(128)
# Residual layers
self.res1 = ResidualBlock(128, pad_type)
self.res2 = ResidualBlock(128, pad_type)
self.res3 = ResidualBlock(128, pad_type)
self.res4 = ResidualBlock(128, pad_type)
self.res5 = ResidualBlock(128, pad_type)
# Upsampling Layers
self.deconv1 = UpsampleConvLayer(128, 64, 4, 2, pad_type=pad_type, upsample=upsample)
self.in4 = nn.InstanceNorm2d(64)
self.deconv2 = UpsampleConvLayer(64, 32, 4, 2, pad_type=pad_type, upsample=upsample)
self.in5 = nn.InstanceNorm2d(32)
self.deconv3 = ConvLayer(32, 3, kernel_size=9, stride=1, pad_type=pad_type)
# Non-linearities
self.relu = nn.ReLU()
self.setup_pad_input(256)
def record(self, x, name):
self.record_tensor.append(x)
self.record_name.append(name)
def reset_pad_type(self, pad_type):
self.pad_type = pad_type
self.conv1.set_pad_method(pad_type)
self.conv2.set_pad_method(pad_type)
self.conv3.set_pad_method(pad_type)
self.res1.pad_type = pad_type
self.res2.pad_type = pad_type
self.res3.pad_type = pad_type
self.res4.pad_type = pad_type
self.res5.pad_type = pad_type
self.res1.conv1.set_pad_method(pad_type)
self.res2.conv1.set_pad_method(pad_type)
self.res3.conv1.set_pad_method(pad_type)
self.res4.conv1.set_pad_method(pad_type)
self.res5.conv1.set_pad_method(pad_type)
self.res1.conv2.set_pad_method(pad_type)
self.res2.conv2.set_pad_method(pad_type)
self.res3.conv2.set_pad_method(pad_type)
self.res4.conv2.set_pad_method(pad_type)
self.res5.conv2.set_pad_method(pad_type)
self.deconv1.pad_type = pad_type
self.deconv2.pad_type = pad_type
self.deconv3.set_pad_method(pad_type)
def debug(self, X):
self.record_tensor = []
self.record_name = []
y = X
y = self.pad(y)
y = self.conv1(y)
self.record(y, "conv1")
y = self.relu(self.in1(y))
y = self.conv2(y)
self.record(y, "conv2")
y = self.relu(self.in2(y))
y = self.conv3(y)
self.record(y, "conv3")
y = self.relu(self.in3(y))
y = self.res1(y)
self.record(y, "res1")
y = self.res2(y)
self.record(y, "res2")
y = self.res3(y)
self.record(y, "res3")
y = self.res4(y)
self.record(y, "res4")
y = self.res5(y)
self.record(y, "res5")
y = self.deconv1(y)
self.record(y, "deconv1")
y = self.relu(self.in4(y))
y = self.deconv2(y)
self.record(y, "deconv2")
y = self.relu(self.in5(y))
y = self.deconv3(y)
return y
def print_shape(self):
for n,t in zip(self.record_name, self.record_tensor):
print("=> %s\t:%s" % (n, str(t.shape)))
def setup_pad_input(self, size=256):
x = torch.zeros(1, 3, size, size).float()
y = self.relu(self.in1(self.conv1(x)))
y = self.relu(self.in2(self.conv2(y)))
y = self.relu(self.in3(self.conv3(y)))
y = self.res1(y)
y = self.res2(y)
y = self.res3(y)
y = self.res4(y)
y = self.res5(y)
y = self.relu(self.in4(self.deconv1(y)))
y = self.relu(self.in5(self.deconv2(y)))
y = self.deconv3(y)
diff_h = x.size(2) - y.size(2)
diff_w = x.size(3) - y.size(3)
ph, pw = diff_h // 2, diff_w // 2
self.pad_input = nn.ReflectionPad2d((ph, ph, pw, pw))
self.diff_h, self.diff_w = diff_h, diff_w
print(str(x.size()) + " => " + str(y.size()))
def pad(self, x):
if "reflect-start" in self.pad_type:
return self.pad_input(x)
elif "resize-start" in self.pad_type:
n, c, h, w = x.shape
return F.upsample_bilinear(x, (self.diff_h + h, self.diff_w + w))
else:
return x
def forward(self, X):
y = X
y = self.pad(y)
y = self.relu(self.in1(self.conv1(y)))
y = self.relu(self.in2(self.conv2(y)))
y = self.relu(self.in3(self.conv3(y)))
y = self.res1(y)
y = self.res2(y)
y = self.res3(y)
y = self.res4(y)
y = self.res5(y)
y = self.relu(self.in4(self.deconv1(y)))
y = self.relu(self.in5(self.deconv2(y)))
y = self.deconv3(y)
return y
class ConvLayer(torch.nn.Module):
def __init__(self, in_channels, out_channels, kernel_size, stride, pad_type):
super(ConvLayer, self).__init__()
self.pad_type = pad_type
self.stride = stride
self.kernel_size = kernel_size
self.conv2d = nn.Conv2d(in_channels, out_channels, kernel_size, stride)
self.set_pad_method(pad_type)
def set_pad_method(self, pad_type):
padding = (self.kernel_size - self.stride, self.kernel_size - self.stride)
if pad_type == "none" or "start" in pad_type:
self.pad = None
else:
self.pad = Padding2d(padding, pad_type)
def forward(self, x):
if self.pad is not None:
x = self.pad(x)
out = self.conv2d(x)
return out
class ResidualBlock(torch.nn.Module):
"""ResidualBlock
introduced in: https://arxiv.org/abs/1512.03385
recommended architecture: http://torch.ch/blog/2016/02/04/resnets.html
"""
def __init__(self, channels, pad_type):
super(ResidualBlock, self).__init__()
self.pad_type = pad_type
self.conv1 = ConvLayer(channels, channels, kernel_size=3, stride=1, pad_type=pad_type)
self.in1 = nn.InstanceNorm2d(channels)
self.conv2 = ConvLayer(channels, channels, kernel_size=3, stride=1, pad_type=pad_type)
self.in2 = nn.InstanceNorm2d(channels)
self.relu = nn.ReLU()
def forward(self, x):
residual = x
if self.pad_type == "none" or "start" in self.pad_type:
residual = residual[:, :, 2:-2, 2:-2]
x = self.relu(self.in1(self.conv1(x)))
x = self.in2(self.conv2(x))
# no padding in the middle
return x + residual
class UpsampleConvLayer(torch.nn.Module):
"""UpsampleConvLayer
Upsamples the input and then does a convolution. This method gives better results
compared to ConvTranspose2d.
ref: http://distill.pub/2016/deconv-checkerboard/
"""
def __init__(self, in_channels, out_channels, kernel_size, stride, pad_type="none", upsample="deconv"):
super(UpsampleConvLayer, self).__init__()
self.upsample = upsample
self.kernel_size = kernel_size
if upsample == "nearest":
self.upsample_layer = torch.nn.Upsample(scale_factor=upsample)
self.conv2d = nn.Conv2d(in_channels, out_channels, kernel_size, stride)
elif upsample == "deconv":
self.deconv2d = nn.ConvTranspose2d(in_channels, out_channels, kernel_size, stride)
pad_size = int(np.floor(kernel_size / 2))
self.pad_type = pad_type
#self.reflection_pad = nn.ReflectionPad2d(reflection_padding)
def forward(self, x):
if self.upsample == "nearest":
x = self.upsample_layer(x)
x = self.reflection_pad(x)
x = self.conv2d(x)
elif self.upsample == "deconv":
x = self.deconv2d(x)
# crop output
if self.kernel_size == 3:
x = x[:, :, 1:, 1:]
elif self.kernel_size == 4:
x = x[:, :, 1:-1, 1:-1]
return x
"""
class nn.InstanceNorm2d(torch.nn.Module):
def __init__(self, dim, eps=1e-9):
super(nn.InstanceNorm2d, self).__init__()
self.scale = nn.Parameter(torch.FloatTensor(dim))
self.shift = nn.Parameter(torch.FloatTensor(dim))
self.eps = eps
self._reset_parameters()
def _reset_parameters(self):
self.scale.data.uniform_()
self.shift.data.zero_()
def forward(self, x):
n = x.size(2) * x.size(3)
t = x.view(x.size(0), x.size(1), n)
mean = torch.mean(t, 2).unsqueeze(2).unsqueeze(3).expand_as(x)
# Calculate the biased var. torch.var returns unbiased var
var = torch.var(t, 2).unsqueeze(2).unsqueeze(3).expand_as(x) * ((n - 1) / float(n))
scale_broadcast = self.scale.unsqueeze(1).unsqueeze(1).unsqueeze(0)
scale_broadcast = scale_broadcast.expand_as(x)
shift_broadcast = self.shift.unsqueeze(1).unsqueeze(1).unsqueeze(0)
shift_broadcast = shift_broadcast.expand_as(x)
out = (x - mean) / torch.sqrt(var + self.eps)
out = out * scale_broadcast + shift_broadcast
return out
"""