-
Notifications
You must be signed in to change notification settings - Fork 57
/
GANs.py
233 lines (211 loc) · 13 KB
/
GANs.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
import tensorflow as tf
import scipy.io as sio
import numpy as np
from PIL import Image
img_H = 64
img_W = 64
img_C = 3
GAN_type = "SNGAN" # DCGAN, WGAN, WGAN-GP, SNGAN, LSGAN, RSGAN, RaSGAN
batchsize = 128
epsilon = 1e-14#if epsilon is too big, training of DCGAN is failure.
def deconv(inputs, shape, strides, out_num, is_sn=False):
filters = tf.get_variable("kernel", shape=shape, initializer=tf.random_normal_initializer(stddev=0.02))
bias = tf.get_variable("bias", shape=[shape[-2]], initializer=tf.constant_initializer([0]))
if is_sn:
return tf.nn.conv2d_transpose(inputs, spectral_norm("sn", filters), out_num, strides) + bias
else:
return tf.nn.conv2d_transpose(inputs, filters, out_num, strides) + bias
def conv(inputs, shape, strides, is_sn=False):
filters = tf.get_variable("kernel", shape=shape, initializer=tf.random_normal_initializer(stddev=0.02))
bias = tf.get_variable("bias", shape=[shape[-1]], initializer=tf.constant_initializer([0]))
if is_sn:
return tf.nn.conv2d(inputs, spectral_norm("sn", filters), strides, "SAME") + bias
else:
return tf.nn.conv2d(inputs, filters, strides, "SAME") + bias
def fully_connected(inputs, num_out, is_sn=False):
W = tf.get_variable("W", [inputs.shape[-1], num_out], initializer=tf.random_normal_initializer(stddev=0.02))
b = tf.get_variable("b", [num_out], initializer=tf.constant_initializer([0]))
if is_sn:
return tf.matmul(inputs, spectral_norm("sn", W)) + b
else:
return tf.matmul(inputs, W) + b
def leaky_relu(inputs, slope=0.2):
return tf.maximum(slope*inputs, inputs)
def spectral_norm(name, w, iteration=1):
#Spectral normalization which was published on ICLR2018,please refer to "https://www.researchgate.net/publication/318572189_Spectral_Normalization_for_Generative_Adversarial_Networks"
#This function spectral_norm is forked from "https://github.com/taki0112/Spectral_Normalization-Tensorflow"
w_shape = w.shape.as_list()
w = tf.reshape(w, [-1, w_shape[-1]])
with tf.variable_scope(name, reuse=False):
u = tf.get_variable("u", [1, w_shape[-1]], initializer=tf.truncated_normal_initializer(), trainable=False)
u_hat = u
v_hat = None
def l2_norm(v, eps=1e-12):
return v / (tf.reduce_sum(v ** 2) ** 0.5 + eps)
for i in range(iteration):
v_ = tf.matmul(u_hat, tf.transpose(w))
v_hat = l2_norm(v_)
u_ = tf.matmul(v_hat, w)
u_hat = l2_norm(u_)
sigma = tf.matmul(tf.matmul(v_hat, w), tf.transpose(u_hat))
w_norm = w / sigma
with tf.control_dependencies([u.assign(u_hat)]):
w_norm = tf.reshape(w_norm, w_shape)
return w_norm
def mapping(x):
max = np.max(x)
min = np.min(x)
return (x - min) * 255.0 / (max - min + epsilon)
def instanceNorm(inputs):
mean, var = tf.nn.moments(inputs, axes=[1, 2], keep_dims=True)
scale = tf.get_variable("scale", shape=mean.shape[-1], initializer=tf.constant_initializer([1.0]))
shift = tf.get_variable("shift", shape=mean.shape[-1], initializer=tf.constant_initializer([0.0]))
return (inputs - mean) * scale / (tf.sqrt(var + epsilon)) + shift
class Generator:
def __init__(self, name):
self.name = name
def __call__(self, Z):
with tf.variable_scope(name_or_scope=self.name, reuse=False):
with tf.variable_scope(name_or_scope="linear"):
inputs = tf.reshape(tf.nn.relu((fully_connected(Z, 4*4*512))), [batchsize, 4, 4, 512])
with tf.variable_scope(name_or_scope="deconv1"):
inputs = tf.nn.relu(instanceNorm(deconv(inputs, [5, 5, 256, 512], [1, 2, 2, 1], [batchsize, 8, 8, 256])))
with tf.variable_scope(name_or_scope="deconv2"):
inputs = tf.nn.relu(instanceNorm(deconv(inputs, [5, 5, 128, 256], [1, 2, 2, 1], [batchsize, 16, 16, 128])))
with tf.variable_scope(name_or_scope="deconv3"):
inputs = tf.nn.relu(instanceNorm(deconv(inputs, [5, 5, 64, 128], [1, 2, 2, 1], [batchsize, 32, 32, 64])))
if img_H == 32:
with tf.variable_scope(name_or_scope="deconv4"):
inputs = tf.nn.tanh(deconv(inputs, [5, 5, img_C, 64], [1, 1, 1, 1], [batchsize, img_H, img_W, img_C]))
return inputs
with tf.variable_scope(name_or_scope="deconv4"):
inputs = tf.nn.tanh(deconv(inputs, [5, 5, img_C, 64], [1, 2, 2, 1], [batchsize, img_H, img_W, img_C]))
if img_H == 64:
return inputs
@property
def var(self):
return tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, self.name)
class Discriminator:
def __init__(self, name):
self.name = name
def __call__(self, inputs, reuse=False, is_sn=False):
with tf.variable_scope(name_or_scope=self.name, reuse=reuse):
with tf.variable_scope("conv1"):
inputs = leaky_relu(conv(inputs, [5, 5, img_C, 64], [1, 2, 2, 1], is_sn))
with tf.variable_scope("conv2"):
inputs = leaky_relu(instanceNorm(conv(inputs, [5, 5, 64, 128], [1, 2, 2, 1], is_sn)))
with tf.variable_scope("conv3"):
inputs = leaky_relu(instanceNorm(conv(inputs, [5, 5, 128, 256], [1, 2, 2, 1], is_sn)))
with tf.variable_scope("conv4"):
inputs = leaky_relu(instanceNorm(conv(inputs, [5, 5, 256, 512], [1, 2, 2, 1], is_sn)))
inputs = tf.layers.flatten(inputs)
return fully_connected(inputs, 1, is_sn)
@property
def var(self):
return tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope=self.name)
class GAN:
#Architecture of generator and discriminator just like DCGAN.
def __init__(self):
self.Z = tf.placeholder("float", [batchsize, 100])
self.img = tf.placeholder("float", [batchsize, img_H, img_W, img_C])
D = Discriminator("discriminator")
G = Generator("generator")
self.fake_img = G(self.Z)
if GAN_type == "DCGAN":
#DCGAN, paper: UNSUPERVISED REPRESENTATION LEARNING WITH DEEP CONVOLUTIONAL GENERATIVE ADVERSARIAL NETWORKS
self.fake_logit = tf.nn.sigmoid(D(self.fake_img))
self.real_logit = tf.nn.sigmoid(D(self.img, reuse=True))
self.d_loss = - (tf.reduce_mean(tf.log(self.real_logit + epsilon)) + tf.reduce_mean(tf.log(1 - self.fake_logit + epsilon)))
self.g_loss = - tf.reduce_mean(tf.log(self.fake_logit + epsilon))
self.opt_D = tf.train.AdamOptimizer(2e-4, beta1=0.5).minimize(self.d_loss, var_list=D.var)
self.opt_G = tf.train.AdamOptimizer(2e-4, beta1=0.5).minimize(self.g_loss, var_list=G.var)
elif GAN_type == "WGAN":
#WGAN, paper: Wasserstein GAN
self.fake_logit = D(self.fake_img)
self.real_logit = D(self.img, reuse=True)
self.d_loss = -tf.reduce_mean(self.real_logit) + tf.reduce_mean(self.fake_logit)
self.g_loss = -tf.reduce_mean(self.fake_logit)
self.clip = []
for _, var in enumerate(D.var):
self.clip.append(var.assign(tf.clip_by_value(var, -0.01, 0.01)))
self.opt_D = tf.train.RMSPropOptimizer(5e-5).minimize(self.d_loss, var_list=D.var)
self.opt_G = tf.train.RMSPropOptimizer(5e-5).minimize(self.g_loss, var_list=G.var)
elif GAN_type == "WGAN-GP":
#WGAN-GP, paper: Improved Training of Wasserstein GANs
self.fake_logit = D(self.fake_img)
self.real_logit = D(self.img, reuse=True)
e = tf.random_uniform([batchsize, 1, 1, 1], 0, 1)
x_hat = e * self.img + (1 - e) * self.fake_img
grad = tf.gradients(D(x_hat, reuse=True), x_hat)[0]
self.d_loss = tf.reduce_mean(self.fake_logit - self.real_logit) + 10 * tf.reduce_mean(tf.square(tf.sqrt(tf.reduce_sum(tf.square(grad), axis=[1, 2, 3])) - 1))
self.g_loss = tf.reduce_mean(-self.fake_logit)
self.opt_D = tf.train.AdamOptimizer(1e-4, beta1=0., beta2=0.9).minimize(self.d_loss, var_list=D.var)
self.opt_G = tf.train.AdamOptimizer(1e-4, beta1=0., beta2=0.9).minimize(self.g_loss, var_list=G.var)
elif GAN_type == "LSGAN":
#LSGAN, paper: Least Squares Generative Adversarial Networks
self.fake_logit = D(self.fake_img)
self.real_logit = D(self.img, reuse=True)
self.d_loss = tf.reduce_mean(0.5 * tf.square(self.real_logit - 1) + 0.5 * tf.square(self.fake_logit))
self.g_loss = tf.reduce_mean(0.5 * tf.square(self.fake_logit - 1))
self.opt_D = tf.train.AdamOptimizer(2e-4, beta1=0.5).minimize(self.d_loss, var_list=D.var)
self.opt_G = tf.train.AdamOptimizer(2e-4, beta1=0.5).minimize(self.g_loss, var_list=G.var)
elif GAN_type == "SNGAN":
#SNGAN, paper: SPECTRAL NORMALIZATION FOR GENERATIVE ADVERSARIAL NETWORKS
self.fake_logit = tf.nn.sigmoid(D(self.fake_img, is_sn=True))
self.real_logit = tf.nn.sigmoid(D(self.img, reuse=True, is_sn=True))
self.d_loss = - (tf.reduce_mean(tf.log(self.real_logit + epsilon) + tf.log(1 - self.fake_logit + epsilon)))
self.g_loss = - tf.reduce_mean(tf.log(self.fake_logit + epsilon))
self.opt_D = tf.train.AdamOptimizer(2e-4, beta1=0.5).minimize(self.d_loss, var_list=D.var)
self.opt_G = tf.train.AdamOptimizer(2e-4, beta1=0.5).minimize(self.g_loss, var_list=G.var)
elif GAN_type == "RSGAN":
#RSGAN, paper: The relativistic discriminator: a key element missing from standard GAN
self.fake_logit = D(self.fake_img)
self.real_logit = D(self.img, reuse=True)
self.d_loss = - tf.reduce_mean(tf.log(tf.nn.sigmoid(self.real_logit - self.fake_logit) + epsilon))
self.g_loss = - tf.reduce_mean(tf.log(tf.nn.sigmoid(self.fake_logit - self.real_logit) + epsilon))
self.opt_D = tf.train.AdamOptimizer(2e-4, beta1=0.5).minimize(self.d_loss, var_list=D.var)
self.opt_G = tf.train.AdamOptimizer(2e-4, beta1=0.5).minimize(self.g_loss, var_list=G.var)
elif GAN_type == "RaSGAN":
#RaSGAN, paper: The relativistic discriminator: a key element missing from standard GAN
self.fake_logit = D(self.fake_img)
self.real_logit = D(self.img, reuse=True)
self.avg_fake_logit = tf.reduce_mean(self.fake_logit)
self.avg_real_logit = tf.reduce_mean(self.real_logit)
self.D_r_tilde = tf.nn.sigmoid(self.real_logit - self.avg_fake_logit)
self.D_f_tilde = tf.nn.sigmoid(self.fake_logit - self.avg_real_logit)
self.d_loss = - tf.reduce_mean(tf.log(self.D_r_tilde + epsilon)) - tf.reduce_mean(tf.log(1 - self.D_f_tilde + epsilon))
self.g_loss = - tf.reduce_mean(tf.log(self.D_f_tilde + epsilon)) - tf.reduce_mean(tf.log(1 - self.D_r_tilde + epsilon))
self.opt_D = tf.train.AdamOptimizer(2e-4, beta1=0.5).minimize(self.d_loss, var_list=D.var)
self.opt_G = tf.train.AdamOptimizer(2e-4, beta1=0.5).minimize(self.g_loss, var_list=G.var)
self.sess = tf.Session()
self.sess.run(tf.global_variables_initializer())
def __call__(self):
saver = tf.train.Saver()
epoch_nums = 200
facedata = sio.loadmat("./TrainingSet//facedata.mat")["data"]
#For face data, i random select about 10,000 images from CelebA and resize them to 64x64 by Matlab.
for epoch in range(epoch_nums):
for i in range(facedata.__len__()//batchsize-1):
batch = facedata[i*batchsize:i*batchsize+batchsize, :, :, :] / 255.0
z = np.random.standard_normal([batchsize, 100])
d_loss = self.sess.run(self.d_loss, feed_dict={self.img: batch, self.Z: z})
g_loss = self.sess.run(self.g_loss, feed_dict={self.img: batch, self.Z: z})
self.sess.run(self.opt_D, feed_dict={self.img: batch, self.Z: z})
if GAN_type == "WGAN":
self.sess.run(self.clip)#WGAN weight clipping
self.sess.run(self.opt_G, feed_dict={self.img: batch, self.Z: z})
if i % 10 == 0:
print("epoch: %d, step: %d, d_loss: %g, g_loss: %g"%(epoch, i, d_loss, g_loss))
z = np.random.standard_normal([batchsize, 100])
imgs = self.sess.run(self.fake_img, feed_dict={self.img: batch, self.Z: z})
for j in range(batchsize):
if img_C == 1:
Image.fromarray(np.reshape(np.uint8(mapping(imgs[j, :, :, :])), [img_H, img_W])).save(
"./result//" + str(epoch) + "_" + str(j) + ".jpg")
else:
Image.fromarray(np.uint8(mapping(imgs[j, :, :, :]))).save(
"./result//" + str(epoch) + "_" + str(j) + ".jpg")
saver.save(self.sess, "./para//model.ckpt")
if __name__ == "__main__":
gan = GAN()
gan()