forked from Automanmm/PPGAN
/
train.py
175 lines (149 loc) · 8.82 KB
/
train.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
"""
Train GAN model.
"""
import os
import tensorflow as tf
import tensorflow.contrib as tc
from data import Loader
from PPGAN import Generator
from discriminator import Discriminator
# import tester
tf.app.flags.DEFINE_string("tables", "", "tables info")
tf.app.flags.DEFINE_string("buckets", "", "buckets info")
tf.app.flags.DEFINE_string("outputs", "", "odps outputs info")
tf.app.flags.DEFINE_string('data_click0_table', './data/click0.csv', 'Directory in which data is stored.')
tf.app.flags.DEFINE_string('data_click1_table', './data/click1.csv', 'Directory in which data is stored.')
tf.app.flags.DEFINE_string('data_dir', './data', 'Directory in which data is stored.')
tf.app.flags.DEFINE_string('vocab_path', './data/vocab.csv', 'Where to save vocabulary.')
tf.app.flags.DEFINE_string('save_dir', './models', 'Where to save checkpoint models.')
tf.app.flags.DEFINE_integer('pn_batch_size', 100, 'Batch size of pointer networks.')
tf.app.flags.DEFINE_integer('ctr_batch_size', 100, 'Batch size of ctr networks.')
tf.app.flags.DEFINE_integer('seq_length', 20, 'The default lengths of long title sequences.')
tf.app.flags.DEFINE_integer('train_epochs', 1000, 'Number of epochs to run.')
tf.app.flags.DEFINE_integer('G_pretrain_epochs', 10, 'The epochs of Generator pre_training.')
tf.app.flags.DEFINE_float('learning_rate', 0.001, 'Learning rate for Adam optimizer.')
tf.app.flags.DEFINE_float('dropout_keep_prob', 0.5, 'Dropout keep probability (default: 0.5).')
tf.app.flags.DEFINE_integer('d_iters', 5, 'The iterations of Discriminator training each epoch.')
tf.app.flags.DEFINE_integer('g_iters', 5, 'The iterations of Generation training each epoch.')
FLAGS = tf.app.flags.FLAGS
def pretrain_G(sess, saver, generator, trainSet, epochs, pre_train_step):
# pre_train Generator
print('Start pre_training generator...')
# tempData = {'output_prob': [], 'x_batch': [], 'u_batch': [], 'x_lengths': [], 'y_hot_batch': []}
for i in range(epochs):
G_loss0, G_acc0 = 0, 0
for itr in range(trainSet.n_batches):
x_batch, u_batch, x_lengths, y_batch, y_hot_batch = trainSet.next_pn_batch()
train_dict = {generator.encoder_inputs: x_batch,
generator.user_inputs: u_batch,
generator.input_lengths: x_lengths,
generator.pointer_labels: y_batch,
generator.pointer_hot_labels: y_hot_batch}
loss, _, output_prob, pre_labels = sess.run([generator.loss,
pre_train_step,
generator.pointer_prob,
generator.rank_pointers],
feed_dict=train_dict)
jishu = 0
for j, line in enumerate(pre_labels):
for word in line:
if word in y_batch[j]:
jishu = jishu + 1
acc = jishu*1.0 / (FLAGS.pn_batch_size * 5)
G_loss0 += loss
G_acc0 += acc
print("Pre_train Generator: epoch:{}, loss:{:.2f}, acc:{:.2f}".format(
i, G_loss0 / trainSet.n_batches, G_acc0 / trainSet.n_batches))
saver.save(sess, os.path.join(FLAGS.save_dir, 'pre_G/pretrain_generator'))
def main(args):
# load data
click1_Set = Loader(os.path.join(FLAGS.data_dir, 'click1.csv'), FLAGS.vocab_path,
FLAGS.pn_batch_size, FLAGS.ctr_batch_size, FLAGS.seq_length)
click0_Set = Loader(os.path.join(FLAGS.data_dir, 'click0.csv'), FLAGS.vocab_path,
FLAGS.pn_batch_size, FLAGS.ctr_batch_size, FLAGS.seq_length)
# pretrain graph
generator = Generator(click1_Set.vocab_size)
pre_optimize = tf.train.AdamOptimizer(FLAGS.learning_rate)
pre_train_step = pre_optimize.minimize(generator.loss, var_list=generator.vars)
saver = tf.train.Saver()
title_inputs = tf.placeholder(tf.int32, [None, FLAGS.seq_length], name='title_inputs')
title_distribution = tf.placeholder(tf.float32, [None, FLAGS.seq_length], name='title_distribution')
# formal train graph
generator_fake = Generator(click0_Set.vocab_size)
real_discriminator = Discriminator(click1_Set.vocab_size, title_inputs, title_distribution)
fake_discriminator = Discriminator(click0_Set.vocab_size, generator_fake.encoder_inputs,
generator_fake.pointer_prob)
reg = tc.layers.apply_regularization(
tc.layers.l1_regularizer(2.5e-5),
weights_list=[var for var in tf.global_variables() if 'kernel' or 'W1' or 'W2' in var.name]
)
D_real_loss = tf.reduce_mean(real_discriminator.predictions)
D_fake_loss = tf.reduce_mean(fake_discriminator.predictions)
D_loss = D_fake_loss - D_real_loss
D_loss_reg = D_loss + reg
D_optimize = tf.train.RMSPropOptimizer(FLAGS.learning_rate)
# WGAN lipschitz-penalty
alpha = tf.random_uniform(
shape=[tf.shape(title_distribution)[0], 1, 1],
minval=0.,
maxval=1.
)
differences = generator_fake.pointer_prob_hot - title_distribution
interpolates = title_distribution + (alpha * differences)
gradients = tf.gradients(Discriminator(click0_Set.vocab_size, generator_fake.encoder_inputs,
interpolates).predictions, [interpolates])[0]
slopes = tf.sqrt(tf.reduce_sum(tf.square(gradients), reduction_indices=[1, 2]))
gradient_penalty = tf.reduce_mean((slopes - 1.) ** 2)
D_loss_reg = D_loss + 10 * gradient_penalty
D_train_step = D_optimize.minimize(D_loss_reg, var_list=fake_discriminator.vars)
G_loss = - tf.reduce_mean(fake_discriminator.predictions)
G_loss_reg = G_loss + reg
G_optimize = tf.train.RMSPropOptimizer(FLAGS.learning_rate)
G_train_step = G_optimize.minimize(G_loss_reg, var_list=generator_fake.vars)
saver_G = tf.train.Saver(var_list=generator.vars)
saver_D = tf.train.Saver(var_list=real_discriminator.vars)
# for var in tf.global_variables():
# print(var.name)
# training precess
init = tf.global_variables_initializer()
with tf.Session() as sess:
sess.run(init)
pretrain_G(sess, saver, generator, click1_Set, FLAGS.G_pretrain_epochs, pre_train_step)
print('Start training...')
for i in range(FLAGS.train_epochs):
d_iters = FLAGS.d_iters
g_iters = FLAGS.g_iters
for _ in range(0, d_iters):
x_batch_fake, u_batch_fake, x_lengths_fake, y_batch_fake, y_hot_batch_fake = click0_Set.next_pn_batch()
x_batch_real, u_batch_real, x_lengths_real, y_batch_real, y_hot_batch_real = click1_Set.next_pn_batch()
D_dict = {generator_fake.encoder_inputs: x_batch_fake,
generator_fake.user_inputs: u_batch_fake,
generator_fake.input_lengths: x_lengths_fake,
generator_fake.pointer_labels: y_batch_fake,
generator_fake.pointer_hot_labels: y_hot_batch_fake,
title_inputs: x_batch_real,
title_distribution: y_hot_batch_real,
real_discriminator.user_inputs: u_batch_real,
fake_discriminator.user_inputs: u_batch_fake}
sess.run(fake_discriminator.d_clip)
loss_Real, loss_Fake, lossD, _ = sess.run([D_real_loss, D_fake_loss, D_loss, D_train_step], feed_dict=D_dict)
for _ in range(0, g_iters):
x_batch_fake, u_batch_fake, x_lengths_fake, y_batch_fake, y_hot_batch_fake = click0_Set.next_pn_batch()
x_batch_real, u_batch_real, x_lengths_real, y_batch_real, y_hot_batch_real = click1_Set.next_pn_batch()
D_dict = {generator_fake.encoder_inputs: x_batch_fake,
generator_fake.user_inputs: u_batch_fake,
generator_fake.input_lengths: x_lengths_fake,
generator_fake.pointer_labels: y_batch_fake,
generator_fake.pointer_hot_labels: y_hot_batch_fake,
title_inputs: x_batch_real,
title_distribution: y_hot_batch_real,
real_discriminator.user_inputs: u_batch_real,
fake_discriminator.user_inputs: u_batch_fake}
lossG, _ = sess.run([G_loss, G_train_step], feed_dict=D_dict)
print("epoch:{}, D_loss:{:.2f}, G_loss:{:.2f}, loss_Real:{:.2f}, loss_Fake:{:.2f}, Sum_loss:{:.2f}"
.format(i, lossD, lossG, loss_Real, loss_Fake, lossD+lossG))
saver_G.save(sess, os.path.join(FLAGS.save_dir, 'G/train_generator'))
saver_D.save(sess, os.path.join(FLAGS.save_dir, 'D/train_discriminator'))
# tester.generator_predict(FLAGS)
if __name__ == '__main__':
tf.app.run()