-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathbinary_gan.py
190 lines (156 loc) · 7.35 KB
/
binary_gan.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
# pylint: disable=E0401
import torchvision.datasets as ds
import torchvision.transforms as transforms
from torch.autograd import Variable
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
import numpy as np
import torchvision.utils as vutils
transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,))])
batch_size = 200
gen_noise_size = 100
epochs = 30
noise_deviaiton = 0.2
d_iters_target = 5
image_size = 28
clamp_lower = -0.01
clamp_upper = 0.01
nc = 1
sw = 4
noise = torch.FloatTensor(batch_size, gen_noise_size).cuda()
fixed_noise = torch.FloatTensor(batch_size, gen_noise_size).normal_(0, 1).cuda()
num_weights = 64
model_out_folder = "./checkpoints"
pic_out_folder = "./sample_images"
class Generator(nn.Module):
def __init__(self):
super(Generator, self).__init__()
self.fc1 = nn.Linear(gen_noise_size, 4*num_weights*sw*sw)
self.norm1 = nn.BatchNorm2d(4*num_weights)
self.dc1 = nn.ConvTranspose2d(num_weights*4, num_weights*2, 4, stride=1, padding=0)
self.norm2 = nn.BatchNorm2d(2*num_weights)
self.dc2 = nn.ConvTranspose2d(num_weights*2, num_weights, 4, stride=2, padding=1)
self.norm3 = nn.BatchNorm2d(num_weights)
self.dc3 = nn.ConvTranspose2d(num_weights, nc, 4, stride=2, padding=1)
def forward(self, x):
x = self.norm1(F.relu(self.fc1(x)).view(batch_size, num_weights*4, sw, sw))
x = self.norm2(F.relu(self.dc1(x)))
x = self.norm3(F.relu(self.dc2(x)))
x = self.dc3(x)
return F.tanh(x)
class Discriminator(nn.Module):
def __init__(self):
super(Discriminator, self).__init__()
self.conv1 = nn.Conv2d(nc, num_weights, kernel_size=4, stride=2, padding=1)
self.norm1 = nn.BatchNorm2d(num_weights)
self.conv2 = nn.Conv2d(num_weights, num_weights*2, kernel_size=4, stride=2, padding=1)
self.norm2 = nn.BatchNorm2d(2*num_weights)
self.conv3 = nn.Conv2d(num_weights*2, num_weights*4, kernel_size=4, stride=1, padding=0)
self.norm3 = nn.BatchNorm2d(4*num_weights)
self.fc1 = nn.Linear(4*num_weights*sw*sw, 1)
def forward(self, x):
x = self.norm1(F.leaky_relu(self.conv1(x)))
x = self.norm2(F.leaky_relu(self.conv2(x)))
x = self.norm3(F.leaky_relu(self.conv3(x)))
x = self.fc1(x.view(batch_size, -1))
return F.sigmoid(x)
generator = Generator()
generator = generator.cuda()
discriminator = Discriminator()
discriminator = discriminator.cuda()
# discriminator.load_state_dict(torch.load("%s/discriminator_epoch_39.pth" % model_out_folder))
# generator.load_state_dict(torch.load("%s/generator_epoch_39.pth" % model_out_folder))
gen_optimizer = optim.Adam(generator.parameters(), lr=0.001, betas=(.1, .999))
dis_optimizer = optim.Adam(discriminator.parameters(), lr=0.001, betas=(.1, .999))
trainset = ds.MNIST("./mnist", download=True, train=True, transform=transform)
testset = ds.MNIST("./mnist", download=True, train=False, transform=transform)
# trainset = torch.index_select(train_mnist.train_data, 0, torch.LongTensor(np.where(train_mnist.train_labels.numpy() == 2)[0])).type(torch.FloatTensor)
# testset = torch.index_select(test_mnist.test_data, 0, torch.LongTensor(np.where(test_mnist.test_labels.numpy() == 2)[0])).type(torch.FloatTensor)
train_feeder = torch.utils.data.DataLoader(
trainset, batch_size=batch_size, shuffle=True, num_workers=2, drop_last=True)
test_feeder = torch.utils.data.DataLoader(
testset, batch_size=batch_size, shuffle=False, num_workers=2, drop_last=True)
input = torch.FloatTensor(batch_size, 1, image_size, image_size).cuda()
one = torch.FloatTensor([1]*batch_size).cuda()
mone = one * -1
gen_iterations = 0
for epoch in range(epochs): # loop over the dataset multiple times
data_iter = iter(train_feeder)
i = 0
while i < len(train_feeder):
############################
# (1) Update D network
###########################
for p in discriminator.parameters(): # reset requires_grad
p.requires_grad = True # they are set to False below in generator update
# train the discriminator Diters times
if gen_iterations < 25 or gen_iterations % 500 == 0:
d_iters = 100
else:
d_iters = d_iters_target
j = 0
while j < d_iters and i < len(train_feeder):
j += 1
# clamp parameters to a cube
for p in discriminator.parameters():
p.data.clamp_(clamp_lower, clamp_upper)
data = data_iter.next()
i += 1
# train with real
real_cpu, _ = data
discriminator.zero_grad()
batch_size = real_cpu.size(0)
real_cpu = real_cpu.cuda()
input.resize_as_(real_cpu).copy_(real_cpu)
inputv = Variable(input)
errD_real = discriminator(inputv)
errD_real.backward(one)
# train with fake
noise.resize_(batch_size, gen_noise_size).normal_(0, 1)
noisev = Variable(noise, volatile = True) # totally freeze generator
fake = Variable(generator(noisev).data)
inputv = fake
errD_fake = discriminator(inputv)
errD_fake.backward(mone)
errD = errD_real - errD_fake
dis_optimizer.step()
############################
# (2) Update G network
###########################
for p in discriminator.parameters():
p.requires_grad = False # to avoid computation
generator.zero_grad()
# in case our last batch was the tail batch of the dataloader,
# make sure we feed a full batch of noise
noise.resize_(batch_size, gen_noise_size).normal_(0, 1)
noisev = Variable(noise).cuda()
fake = generator(noisev)
errG = discriminator(fake)
errG.backward(one)
gen_optimizer.step()
gen_iterations += 1
if i % 100 == 0:
print('[%d/%d][%d/%d][%d] Loss_D: %f Loss_G: %f Loss_D_real: %f Loss_D_fake %f'
% (epoch, epochs, i, len(train_feeder), gen_iterations,
errD.data[0][0], errG.data[0][0], errD_real.data[0][0], errD_fake.data[0][0]))
if gen_iterations % 100 == 0:
real_cpu = real_cpu.mul(0.5).add(0.5)
vutils.save_image(real_cpu, '%s/real_samples.png' % pic_out_folder)
fake = generator(Variable(fixed_noise, volatile=True))
fake.data = fake.data.mul(0.5).add(0.5)
vutils.save_image(fake.data, '%s/fake_samples_%d.png' % (pic_out_folder, gen_iterations))
torch.save(generator.state_dict(), '%s/generator_epoch_%d.pth' % (model_out_folder, epoch))
torch.save(discriminator.state_dict(), '%s/discriminator_epoch_%d.pth' % (model_out_folder, epoch))
print('Finished Training')
# correct = 0
# total = 0
# for data in test_feeder:
# images, labels = data
# outputs = generator(Variable(images.cuda()))
# _, predicted = torch.max(outputs.data, 1)
# total += labels.size(0)
# correct += (predicted == labels.cuda()).sum()
# print('Accuracy of the network on the 10000 test images: %d %%' % (
# 100 * correct / total))