forked from yunjey/mnist-svhn-transfer
-
Notifications
You must be signed in to change notification settings - Fork 0
/
solver.py
228 lines (191 loc) · 8.47 KB
/
solver.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
import torch
import torch.nn as nn
import torchvision
import os
import pickle
import scipy.io
import numpy as np
from torch.autograd import Variable
from torch import optim
from model import G12, G21
from model import D1, D2
class Solver(object):
def __init__(self, config, svhn_loader, mnist_loader):
self.svhn_loader = svhn_loader
self.mnist_loader = mnist_loader
self.g12 = None
self.g21 = None
self.d1 = None
self.d2 = None
self.g_optimizer = None
self.d_optimizer = None
self.use_reconst_loss = config.use_reconst_loss
self.use_labels = config.use_labels
self.num_classes = config.num_classes
self.beta1 = config.beta1
self.beta2 = config.beta2
self.g_conv_dim = config.g_conv_dim
self.d_conv_dim = config.d_conv_dim
self.train_iters = config.train_iters
self.batch_size = config.batch_size
self.lr = config.lr
self.log_step = config.log_step
self.sample_step = config.sample_step
self.sample_path = config.sample_path
self.model_path = config.model_path
self.build_model()
def build_model(self):
"""Builds a generator and a discriminator."""
self.g12 = G12(conv_dim=self.g_conv_dim)
self.g21 = G21(conv_dim=self.g_conv_dim)
self.d1 = D1(conv_dim=self.d_conv_dim, use_labels=self.use_labels)
self.d2 = D2(conv_dim=self.d_conv_dim, use_labels=self.use_labels)
g_params = list(self.g12.parameters()) + list(self.g21.parameters())
d_params = list(self.d1.parameters()) + list(self.d2.parameters())
self.g_optimizer = optim.Adam(g_params, self.lr, [self.beta1, self.beta2])
self.d_optimizer = optim.Adam(d_params, self.lr, [self.beta1, self.beta2])
if torch.cuda.is_available():
self.g12.cuda()
self.g21.cuda()
self.d1.cuda()
self.d2.cuda()
def merge_images(self, sources, targets, k=10):
_, _, h, w = sources.shape
row = int(np.sqrt(self.batch_size))
merged = np.zeros([3, row*h, row*w*2])
for idx, (s, t) in enumerate(zip(sources, targets)):
i = idx // row
j = idx % row
merged[:, i*h:(i+1)*h, (j*2)*h:(j*2+1)*h] = s
merged[:, i*h:(i+1)*h, (j*2+1)*h:(j*2+2)*h] = t
return merged.transpose(1, 2, 0)
def to_var(self, x):
"""Converts numpy to variable."""
if torch.cuda.is_available():
x = x.cuda()
return Variable(x)
def to_data(self, x):
"""Converts variable to numpy."""
if torch.cuda.is_available():
x = x.cpu()
return x.data.numpy()
def reset_grad(self):
"""Zeros the gradient buffers."""
self.g_optimizer.zero_grad()
self.d_optimizer.zero_grad()
def train(self):
svhn_iter = iter(self.svhn_loader)
mnist_iter = iter(self.mnist_loader)
iter_per_epoch = min(len(svhn_iter), len(mnist_iter))
# fixed mnist and svhn for sampling
fixed_svhn = self.to_var(svhn_iter.next()[0])
fixed_mnist = self.to_var(mnist_iter.next()[0])
# loss if use_labels = True
criterion = nn.CrossEntropyLoss()
for step in range(self.train_iters+1):
# reset data_iter for each epoch
if (step+1) % iter_per_epoch == 0:
mnist_iter = iter(self.mnist_loader)
svhn_iter = iter(self.svhn_loader)
# load svhn and mnist dataset
svhn, s_labels = svhn_iter.next()
svhn, s_labels = self.to_var(svhn), self.to_var(s_labels).long().squeeze()
mnist, m_labels = mnist_iter.next()
mnist, m_labels = self.to_var(mnist), self.to_var(m_labels)
if self.use_labels:
mnist_fake_labels = self.to_var(
torch.Tensor([self.num_classes]*svhn.size(0)).long())
svhn_fake_labels = self.to_var(
torch.Tensor([self.num_classes]*mnist.size(0)).long())
#============ train D ============#
# train with real images
self.reset_grad()
out = self.d1(mnist)
if self.use_labels:
d1_loss = criterion(out, m_labels)
else:
d1_loss = torch.mean((out-1)**2)
out = self.d2(svhn)
if self.use_labels:
d2_loss = criterion(out, s_labels)
else:
d2_loss = torch.mean((out-1)**2)
d_mnist_loss = d1_loss
d_svhn_loss = d2_loss
d_real_loss = d1_loss + d2_loss
d_real_loss.backward()
self.d_optimizer.step()
# train with fake images
self.reset_grad()
fake_svhn = self.g12(mnist)
out = self.d2(fake_svhn)
if self.use_labels:
d2_loss = criterion(out, svhn_fake_labels)
else:
d2_loss = torch.mean(out**2)
fake_mnist = self.g21(svhn)
out = self.d1(fake_mnist)
if self.use_labels:
d1_loss = criterion(out, mnist_fake_labels)
else:
d1_loss = torch.mean(out**2)
d_fake_loss = d1_loss + d2_loss
d_fake_loss.backward()
self.d_optimizer.step()
#============ train G ============#
# train mnist-svhn-mnist cycle
self.reset_grad()
fake_svhn = self.g12(mnist)
out = self.d2(fake_svhn)
reconst_mnist = self.g21(fake_svhn)
if self.use_labels:
g_loss = criterion(out, m_labels)
else:
g_loss = torch.mean((out-1)**2)
if self.use_reconst_loss:
g_loss += torch.mean((mnist - reconst_mnist)**2)
g_loss.backward()
self.g_optimizer.step()
# train svhn-mnist-svhn cycle
self.reset_grad()
fake_mnist = self.g21(svhn)
out = self.d1(fake_mnist)
reconst_svhn = self.g12(fake_mnist)
if self.use_labels:
g_loss = criterion(out, s_labels)
else:
g_loss = torch.mean((out-1)**2)
if self.use_reconst_loss:
g_loss += torch.mean((svhn - reconst_svhn)**2)
g_loss.backward()
self.g_optimizer.step()
# print the log info
if (step+1) % self.log_step == 0:
print('Step [%d/%d], d_real_loss: %.4f, d_mnist_loss: %.4f, d_svhn_loss: %.4f, '
'd_fake_loss: %.4f, g_loss: %.4f'
%(step+1, self.train_iters, d_real_loss.data[0], d_mnist_loss.data[0],
d_svhn_loss.data[0], d_fake_loss.data[0], g_loss.data[0]))
# save the sampled images
if (step+1) % self.sample_step == 0:
fake_svhn = self.g12(fixed_mnist)
fake_mnist = self.g21(fixed_svhn)
mnist, fake_mnist = self.to_data(fixed_mnist), self.to_data(fake_mnist)
svhn , fake_svhn = self.to_data(fixed_svhn), self.to_data(fake_svhn)
merged = self.merge_images(mnist, fake_svhn)
path = os.path.join(self.sample_path, 'sample-%d-m-s.png' %(step+1))
scipy.misc.imsave(path, merged)
print ('saved %s' %path)
merged = self.merge_images(svhn, fake_mnist)
path = os.path.join(self.sample_path, 'sample-%d-s-m.png' %(step+1))
scipy.misc.imsave(path, merged)
print ('saved %s' %path)
if (step+1) % 5000 == 0:
# save the model parameters for each epoch
g12_path = os.path.join(self.model_path, 'g12-%d.pkl' %(step+1))
g21_path = os.path.join(self.model_path, 'g21-%d.pkl' %(step+1))
d1_path = os.path.join(self.model_path, 'd1-%d.pkl' %(step+1))
d2_path = os.path.join(self.model_path, 'd2-%d.pkl' %(step+1))
torch.save(self.g12.state_dict(), g12_path)
torch.save(self.g21.state_dict(), g21_path)
torch.save(self.d1.state_dict(), d1_path)
torch.save(self.d2.state_dict(), d2_path)