-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathtest_minimax_mnist.py
187 lines (152 loc) · 7.06 KB
/
test_minimax_mnist.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
## Generative model of mnist digits.
## Jihun Hamm, 2017
import time
import numpy as np
import scipy.io
import tensorflow as tf
import matplotlib.pyplot as plt
import matplotlib.backends.backend_pdf
from mpl_toolkits.axes_grid1 import ImageGrid
#from __future__ import print_function
import keras
#from keras.datasets import cifar10, cifar100
from keras.datasets import mnist
from utils import *
import sys
##########################################################################################################################
def decoder(ins,scope,reuse=False,d=10): # z -> x
batch_size = tf.shape(ins)[0]
with tf.variable_scope(scope,reuse=reuse):
W1 = tf.get_variable('W1',[d,7*7*64],initializer=tf.random_normal_initializer(stddev=0.01))
b1 = tf.get_variable('b1',[7*7*64],initializer=tf.constant_initializer(0.0))
a1 = tf.nn.relu(tf.nn.bias_add(tf.matmul(ins,W1),b1))
W2 = tf.get_variable('W2',[5,5,32,64],initializer=tf.random_normal_initializer(stddev=0.01))
b2 = tf.get_variable('b2',[32],initializer=tf.constant_initializer(0.0))
a2 = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d_transpose(tf.reshape(a1,[-1,7,7,64]),W2,[batch_size,14,14,32],strides=[1,2,2,1],padding='SAME'),b2))
W3 = tf.get_variable('W3',[5,5,1,32],initializer=tf.random_normal_initializer(stddev=0.01))
b3 = tf.get_variable('b3',[1],initializer=tf.constant_initializer(0.0))
a3 = tf.nn.sigmoid(tf.nn.bias_add(tf.nn.conv2d_transpose(a2,W3,[batch_size,28,28,1],strides=[1,2,2,1],padding='SAME'),b3))
out = a3#tf.reshape(a3,[-1,28,28,1])
reg = tf.nn.l2_loss(W1) + tf.nn.l2_loss(W2) + tf.nn.l2_loss(W3)
return [out,reg]
def discriminator(ins,scope,reuse=False,nh=50):
with tf.variable_scope(scope,reuse=reuse):
W1 = tf.get_variable('W1',[5,5,1,16],initializer=tf.random_normal_initializer(stddev=0.01))
b1 = tf.get_variable('b1',[16],initializer=tf.constant_initializer(0.0))
c1 = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(ins, W1, strides=[1,1,1,1], padding='SAME'),b1))
p1 = tf.nn.max_pool(c1, ksize=[1,2,2,1],strides=[1,2,2,1], padding='SAME')
W2 = tf.get_variable('W2',[5,5,16,32],initializer=tf.random_normal_initializer(stddev=0.01))
b2 = tf.get_variable('b2',[32],initializer=tf.constant_initializer(0.0))
c2 = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(p1, W2, strides=[1,1,1,1], padding='SAME'), b2))
p2 = tf.nn.max_pool(c2, ksize=[1,2,2,1],strides=[1,2,2,1], padding='SAME')
a2 = tf.reshape(p2,[-1,7*7*32])
W3 = tf.get_variable('W3',[7*7*32,K],initializer=tf.random_normal_initializer(stddev=0.01))
b3 = tf.get_variable('b3',[K],initializer=tf.constant_initializer(0.0))
out = tf.nn.bias_add(tf.matmul(a2,W3),b3)
reg = tf.nn.l2_loss(W1) + tf.nn.l2_loss(W2) + tf.nn.l2_loss(W3)
return [out,reg]
def minimax_kbeam(sess, feed_dict):
# min-step
for it_min in range(min_step):
fs = sess.run(loss,feed_dict)
id_max = np.argmax(fs)
sess.run(optim_min[id_max],feed_dict)
# max-step
for it_max in range(max_step):
#fs = sess.run(loss,feed_dict)
#id_max = np.argmax(fs)
#sess.run(optim_max[id_max],feed_dict)
sess.run(optim_max,feed_dict)
###########################################################################################################################################
D = 28*28*1
d = 10
K = 5
max_step = 1
min_step = 1
batchsize = 128
lr_gen = 1E-3
lr_disc = 1E-3
niter = 10001
## Load data
(X1train, y1train), (X1test, y1test) = mnist.load_data()
X1train = X1train.reshape((-1,28,28,1)).astype('float32')/255.
X1test = X1test.reshape((-1,28,28,1)).astype('float32')/255.
y1train = keras.utils.to_categorical(y1train,10)
y1test = keras.utils.to_categorical(y1test,10)
## Create the model
x1 = tf.placeholder(tf.float32, [None, 28,28,1])
x2 = tf.placeholder(tf.float32, [None, d])
## Generator
dec,reg_dec = decoder(x2,'mnist_decoder',d=d)
vars_dec = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,scope='mnist_decoder')
## Discriminators
disc_real=[[] for i in range(K)]
disc_fake=[[] for i in range(K)]
loss=[[] for i in range(K)]
vars_disc=[[] for i in range(K)]
optim_max=[[] for i in range(K)]
optim_min=[[] for i in range(K)]
reg_disc=[[] for i in range(K)]
optimizer_min = tf.train.AdamOptimizer(lr_gen)
for i in range(K):
disc_real[i], reg_disc[i]= discriminator(x1,'disc'+str(i),False)
disc_fake[i],_ = discriminator(dec,'disc'+str(i),True)
loss[i] = -tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=disc_real[i], labels=tf.ones_like(disc_real[i]))) \
-tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=disc_fake[i], labels=tf.zeros_like(disc_fake[i])))
'''
eps = 1E-12
## original loss - doesn't work
loss[i] = tf.reduce_mean(tf.log(tf.maximum(eps,tf.minimum(tf.nn.sigmoid(disc_real[i]),1.-eps)))) \
+tf.reduce_mean(tf.log(tf.subtract(1.,tf.maximum(eps,tf.minimum(tf.nn.sigmoid(disc_fake[i]),1-eps)))))
'''
vars_disc[i] = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, 'disc'+str(i))
#optim_min[i] = tf.train.AdamOptimizer(lr_gen).minimize(loss[i],var_list=vars_dec)
optim_min[i] = optimizer_min.minimize(loss[i],var_list=vars_dec)
optim_max[i] = tf.train.AdamOptimizer(lr_disc).minimize(-loss[i],var_list=vars_disc[i])
## Misc
#saver = tf.train.Saver()
#########################################################################################################
## Init
sess = tf.Session()
sess.run(tf.global_variables_initializer())
#sess.run(tf.initialize_all_variables())
## Load pre-trained autoencoders
#saver_dec.restore(sess, dir_model+'/mnist_decoder.ckpt')
if True:
X2test = np.random.uniform(-1.,1.,size=((64,d)))
#np.save(dir_result+'/X2test.npy',X2test)
else:
pass
#X2test = np.load(dir_result+'/X2test.npy')
gen_batch = batch_generator([X1train, y1train], batchsize)
## Burn-in period for v, when using u from autoencoder
for it in range(0):
X1, Y1 = gen_batch.next()
X2 = np.random.uniform(-1.,1.,size=(batchsize,d))
feed_dict = {x1:X1, x2:X2}
sess.run(optim_max,feed_dict)
if it%100 == 0:
fs = sess.run(loss,feed_dict)
id_max = np.argmax(fs)
f = np.max(fs)
print 'step %d: f=%g, id_max=%d'%(it,f,id_max)
## Minimax
for it in range(niter):
X1, Y1 = gen_batch.next()
X2 = np.random.uniform(-1.,1.,size=(batchsize,d))
feed_dict = {x1:X1, x2:X2}
minimax_kbeam(sess, feed_dict)
if it%100 == 0:
fs = sess.run(loss,feed_dict)
id_max = np.argmax(fs)
f = np.max(fs)
print 'step %d: f=%g, id_max=%d'%(it,f,id_max)
if it%5000 == 0:
X2test_ = sess.run(dec,feed_dict={x2:X2test})
imshow_grid(X2test_[:64,:].reshape((64,28,28)),[8,8])
plt.suptitle('t=%d'%(it),size=12)
#plt.show(block=False)
plt.pause(3.)
#with matplotlib.backends.backend_pdf.PdfPages(dir_result+'/gan_mnist_K%d_max%d_%d.pdf'%(K,max_step,it)) as pdf:
# pdf.savefig(bbox_inches='tight')
raw_input('Press any key to continue')