-
Notifications
You must be signed in to change notification settings - Fork 9
/
Copy pathinfogan.py
197 lines (163 loc) · 9.76 KB
/
infogan.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
# coding:utf-8
import matplotlib, os
matplotlib.use("Agg")
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
import tensorflow as tf
import numpy as np
from utils import BasicTrainFramework, one_hot_encode
from datamanager import datamanager_mnist
from discriminators import Discriminator_CNN, Discriminator_MLP
from generators import Generator_CNN, Generator_MLP
from classifier import Classifier_MLP
from matplotlib import pyplot as plt
class InfoGAN(BasicTrainFramework):
def __init__(self,
optim_type='adam',
net_type='cnn',
batch_size=64,
noise_dim=50,
learning_rate=2e-4,
optim_num=0.5,
critic_iter=1,
plot_iter=5,
verbose=True):
self.noise_dim = noise_dim
self.class_num = 10
self.lr = learning_rate
self.optim_num = optim_num
self.critic_iter = critic_iter
self.plot_iter = plot_iter
self.verbose = verbose
super(InfoGAN, self).__init__(batch_size, "infogan_"+net_type)
# code
self.len_discrete_code = self.class_num # categorical distribution (i.e. label)
self.len_continuous_code = 2 # gaussian distribution (e.g. rotation, thickness)
self.optim_type = optim_type
self.SUPERVISED = True
self.data = datamanager_mnist(train_ratio=1.0, fold_k=None, norm=True, expand_dim=True, seed=23333)
sample_data = datamanager_mnist(train_ratio=1.0, fold_k=None, norm=True, expand_dim=True, seed=23333)
self.sample_data = sample_data(self.batch_size, var_list=["data", "labels"])
if net_type == 'cnn':
self.generator = Generator_CNN(name='cnn_generator')
self.discriminator = Discriminator_CNN(name='cnn_discriminator')
elif net_type == 'mlp':
self.generator = Generator_MLP(name='mlp_generator')
self.discriminator = Discriminator_CNN(name='discriminator')
self.classifier = Classifier_MLP(self.len_discrete_code + self.len_continuous_code, name='classifier')
self.build_placeholder()
self.build_gan()
self.build_optimizer(optim_type)
self.build_summary()
self.build_sess()
self.build_dirs()
def build_placeholder(self):
self.noise = tf.placeholder(shape=(self.batch_size, self.noise_dim), dtype=tf.float32)
self.source = tf.placeholder(shape=(self.batch_size, 28, 28, 1), dtype=tf.float32)
self.labels = tf.placeholder(shape=(self.batch_size, self.len_discrete_code + self.len_continuous_code), dtype=tf.float32)
def build_gan(self):
self.G = self.generator(self.noise, self.labels, is_training=True, reuse=False)
self.G_test = self.generator(self.noise, self.labels, is_training=False, reuse=True)
self.logit_real, self.conv_real = self.discriminator(self.source, is_training=True, reuse=False)
self.logit_fake, self.conv_fake = self.discriminator(self.G, is_training=True, reuse=True)
self.logit_cls, self.softmax_cls = self.classifier(self.conv_fake, is_training=True)
def build_optimizer(self, optim_type='adam'):
self.D_loss_real = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=self.logit_real, labels=tf.ones_like(self.logit_real)))
self.D_loss_fake = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=self.logit_fake, labels=tf.zeros_like(self.logit_fake)))
self.D_loss = self.D_loss_real + self.D_loss_fake
self.G_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=self.logit_fake, labels=tf.ones_like(self.logit_fake)))
# discrete code : categorical
disc_code_est = self.softmax_cls[:, :self.len_discrete_code]
disc_code_tg = self.labels[:, :self.len_discrete_code]
q_disc_loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=disc_code_est, labels=disc_code_tg))
# continuous code : gaussian
cont_code_est = self.softmax_cls[:, self.len_discrete_code:]
cont_code_tg = self.labels[:, self.len_discrete_code:]
q_cont_loss = tf.reduce_mean(tf.reduce_sum(tf.square(cont_code_tg - cont_code_est), axis=1))
self.Q_loss = q_disc_loss + q_cont_loss
if optim_type == 'adam':
self.D_solver = tf.train.AdamOptimizer(learning_rate=self.lr, beta1=self.optim_num).minimize(self.D_loss, var_list=self.discriminator.vars)
self.G_solver = tf.train.AdamOptimizer(learning_rate=5*self.lr, beta1=self.optim_num).minimize(self.G_loss, var_list=self.generator.vars)
elif optim_type == 'rmsprop':
self.D_solver = tf.train.RMSPropOptimizer(learning_rate=self.lr).minimize(self.D_loss, var_list=self.discriminator.vars)
self.G_solver = tf.train.RMSPropOptimizer(learning_rate=5*self.lr).minimize(self.G_loss, var_list=self.generator.vars)
self.Q_solver = tf.train.AdamOptimizer(learning_rate=5*self.lr, beta1=self.optim_num).minimize(self.Q_loss, var_list=self.classifier.vars + self.generator.vars + self.discriminator.vars)
def build_summary(self):
D_sum = tf.summary.scalar("D_loss", self.D_loss)
D_sum_real = tf.summary.scalar("D_loss_real", self.D_loss_real)
D_sum_fake = tf.summary.scalar("D_loss_fake", self.D_loss_fake)
G_sum = tf.summary.scalar("G_loss", self.G_loss)
Q_sum = tf.summary.scalar("Q_loss", self.Q_loss)
self.summary = tf.summary.merge([D_sum, D_sum_real, D_sum_fake, G_sum, Q_sum])
def sample(self, epoch):
# batch_size=64
labels = np.repeat(np.arange(10).reshape((10,1)), 10)
labels = one_hot_encode(labels, 10) # [100, 10]
test_codes = np.concatenate([labels, np.tile(np.linspace(-1, 1, 10), 10)[:, None], np.zeros((100, 1))], 1)
test_codes = np.concatenate([test_codes, test_codes], 0)
feed_dict = {
self.labels: test_codes[:self.batch_size, :],
self.noise: np.random.uniform(size=(self.batch_size, self.noise_dim), low=-1.0, high=1.0)
}
G1 = self.sess.run(self.G_test, feed_dict=feed_dict)
feed_dict = {
self.labels: test_codes[self.batch_size:2*self.batch_size, :],
self.noise: np.random.uniform(size=(self.batch_size, self.noise_dim), low=-1.0, high=1.0)
}
G1 = np.concatenate([G1, self.sess.run(self.G_test, feed_dict=feed_dict)], 0)
G1 = np.concatenate([np.concatenate([seq[:,:,0] for seq in G1[10*i:10*(i+1)]], 1) for i in range(10)], 0)
plt.imshow(G1, cmap=plt.cm.gray)
plt.savefig(os.path.join(self.fig_dir, "fake_{}_1.png".format(epoch)))
plt.clf()
test_codes = np.concatenate([labels, np.zeros((100, 1)), np.tile(np.linspace(-1, 1, 10), 10)[:, None]], 1)
test_codes = np.concatenate([test_codes, test_codes], 0)
feed_dict = {
self.labels: test_codes[:self.batch_size, :],
self.noise: np.random.uniform(size=(self.batch_size, self.noise_dim), low=-1.0, high=1.0)
}
G2 = self.sess.run(self.G_test, feed_dict=feed_dict)
feed_dict = {
self.labels: test_codes[self.batch_size:2*self.batch_size, :],
self.noise: np.random.uniform(size=(self.batch_size, self.noise_dim), low=-1.0, high=1.0)
}
G2 = np.concatenate([G2, self.sess.run(self.G_test, feed_dict=feed_dict)], 0)
G2 = np.concatenate([np.concatenate([seq[:,:,0] for seq in G2[10*i:10*(i+1)]], 1) for i in range(10)], 0)
plt.imshow(G2, cmap=plt.cm.gray)
plt.savefig(os.path.join(self.fig_dir, "fake_{}_2.png".format(epoch)))
plt.clf()
def train(self, epoches=1):
self.writer = tf.summary.FileWriter(self.log_dir, self.sess.graph)
batches_per_epoch = self.data.train_num // self.batch_size
for epoch in range(epoches):
self.data.shuffle_train(seed=epoch)
for idx in range(batches_per_epoch):
cnt = epoch * batches_per_epoch + idx
data = self.data(self.batch_size, var_list=["data", "labels"])
# [bz, 10]
if self.SUPERVISED:
batch_labels = data["labels"]
else:
batch_labels = np.random.multinomial(1, self.len_discrete_code * [float(1.0/self.len_discrete_code)], size=[self.batch_size])
# [bz, 12]
batch_codes = np.concatenate((batch_labels, np.random.uniform(-1,1,size=(self.batch_size, 2))), axis=1)
feed_dict = {
self.source: data["data"],
self.labels: batch_codes,
self.noise: np.random.uniform(size=(self.batch_size, self.noise_dim), low=-1.0, high=1.0)
}
# train D
self.sess.run(self.D_solver, feed_dict=feed_dict)
# train G
if (cnt-1) % self.critic_iter == 0:
self.sess.run(self.G_solver, feed_dict=feed_dict)
# train Q
self.sess.run(self.Q_solver, feed_dict=feed_dict)
if cnt % 10 == 0:
d_loss, d_loss_r, d_loss_f, g_loss, q_loss, sum_str = self.sess.run([self.D_loss, self.D_loss_real, self.D_loss_fake, self.G_loss, self.Q_loss, self.summary], feed_dict=feed_dict)
if self.verbose:
print self.version + " epoch [%3d/%3d] iter [%3d/%3d] D=%.3f Dr=%.3f Df=%.3f G=%.3f Q=%.3f" % \
(epoch, epoches, idx, batches_per_epoch, d_loss, d_loss_r, d_loss_f, g_loss, q_loss)
self.writer.add_summary(sum_str, cnt)
if epoch % self.plot_iter == 0:
self.sample(epoch)
self.sample(epoch)
self.saver.save(self.sess, os.path.join(self.model_dir, 'model.ckpt'), global_step=cnt)