/
train.py
271 lines (236 loc) · 13.9 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
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
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
import json
import os
from pprint import pprint
import numpy as np
import tensorflow as tf
from tensorflow.python.ops import control_flow_ops
from generator import Generator
from data.moving_mnist import read_data_sets
from discriminator import Discriminator
from utils import save_gif_from_sampled_videos, restore_vars, add_gradient_summaries, mkdir, \
cleanup_dir, add_variable_summaries
flags = tf.app.flags
FLAGS = flags.FLAGS
flags.DEFINE_string("loss_name", "WGAN-GP", "loss name: ALTERNATIVE, BASIC, WGAN-GP")
flags.DEFINE_float("gradient_policy_scale", 200.0, "labmda of gradient policy")
flags.DEFINE_integer("train_size", 10000, "size of train split")
flags.DEFINE_integer("batch_size", 64, "batch size")
flags.DEFINE_integer("total_batches", -1, "amount of batch run")
flags.DEFINE_integer("hidden_dim", 128, "hidden dim of LSTM")
flags.DEFINE_integer("sequence_length", 20, "number of frames in training")
flags.DEFINE_integer("num_channels", 1, "1 = grayscale, 3 = RGB...")
flags.DEFINE_integer("learning_rate_decay_schedule", 0, "number of iters for lr decay. put 0 to not decay")
flags.DEFINE_float("learning_rate_decay_rate", 0.96, "lr decay rate")
flags.DEFINE_float("d_lr", 0.0002, "lr of critic")
flags.DEFINE_float("adam_b1", 0.5, "adam b1")
flags.DEFINE_float("adam_b2", 0.9, "adam b2")
flags.DEFINE_float("g_lr_lstm", 0.0001, "lr of generator lstm")
flags.DEFINE_float("g_lr_conv", 0.0002, "lr of generator conv")
flags.DEFINE_float("g_momentum_lstm", 0.99, "momentum for SGD of generator LSTM")
flags.DEFINE_float("regularizer_scale", 0.0, "L2 regularizer, for critic only")
flags.DEFINE_integer("sample_every_n_batches", 200, "save gif")
flags.DEFINE_integer("save_every_n_batches", 200, "checkpoint iter")
flags.DEFINE_float("dropout_kept_prob", 0.8, "dropout")
flags.DEFINE_bool("d_use_batch_norm", False, "")
flags.DEFINE_bool("d_use_layer_norm", False, "") # Don't set this to True, currently has bugs
flags.DEFINE_bool("g_recurrent_use_layer_norm", True, "use layer norm in LSTM")
flags.DEFINE_bool("g_use_batch_norm", False, "")
flags.DEFINE_bool("g_use_layer_norm", True, "")
flags.DEFINE_bool("g_add_input_to_lstm", True, "add generated image to state of LSTM") # this is working now due to layer norm scale h*U and x*W to the same scale
flags.DEFINE_string("g_nonlinearity", "relu", "generator conv nonlinearity: relu or lrelu")
flags.DEFINE_integer("d_steps", 5, "")
flags.DEFINE_integer("g_steps", 1, "")
flags.DEFINE_float("d_threshold", float("inf"), "threshold to train d loss to under")
flags.DEFINE_float("g_threshold", float("inf"), "threshold to train g loss to under")
flags.DEFINE_bool("d_gradient_clip", True, "variable gradient clipping for d")
flags.DEFINE_float("d_clip_norm", 5.0, "norm for each var")
flags.DEFINE_bool("g_gradient_clip_conv", True, "variable gradient clipping for convolution of g")
flags.DEFINE_float("g_clip_norm_conv", 5.0, "norm for each var")
flags.DEFINE_bool("g_gradient_clip_lstm", True, "variable gradient clipping for lstm of g")
flags.DEFINE_float("g_clip_norm_lstm", 5.0, "norm for each var")
flags.DEFINE_bool("cleanup", False, "delete checkpoint and summary director")
flags.DEFINE_string("ckpt_dir", "checkpoints", "checkpoint dir to load/save")
flags.DEFINE_string("ckpt_fn", "", "name of file in checkpoint dir to load from. example: model-10000")
flags.DEFINE_string("log_dir", "logs", "dir to write summaries to")
flags.DEFINE_string("out_dir", "out", "dir to save generated gif files to")
flags.DEFINE_integer("random_seed", 1337, "random seed")
def main(_):
print("CONFIG: ")
pprint(FLAGS.__flags)
np.random.seed(FLAGS.random_seed)
tf.reset_default_graph()
tf.set_random_seed(FLAGS.random_seed)
save_dir = os.path.join(FLAGS.ckpt_dir)
generator = Generator(**FLAGS.__flags)
discriminator = Discriminator(**FLAGS.__flags)
global_step = tf.Variable(0, name="global_step", trainable=False)
noises = generator.get_noise_tensor()
x = tf.placeholder(dtype=tf.float32, shape=[FLAGS.batch_size, FLAGS.sequence_length,
64, 64, FLAGS.num_channels])
is_training = tf.placeholder(tf.bool, shape=())
dropout_kept_prob = tf.placeholder(tf.float32, shape=())
# x_noise = tf.placeholder(tf.float32, shape=[FLAGS.batch_size, FLAGS.sequence_length, 64, 64,
# FLAGS.num_channels])
# x_fake_noise = tf.placeholder(tf.float32, shape=[FLAGS.batch_size, FLAGS.sequence_length, 64, 64,
# FLAGS.num_channels])
x_fake = generator.g_output_unit(noises, is_training)
d_real, d_real_logit = discriminator.d_output_unit(x,
dropout_kept_prob, is_training)
d_fake, d_fake_logit = discriminator.d_output_unit(x_fake, dropout_kept_prob,
is_training)
if FLAGS.loss_name == "BASIC":
g_loss = -tf.reduce_mean(tf.log(d_fake))
d_loss_real = -tf.reduce_mean(tf.log(d_real))
d_loss_fake = -tf.reduce_mean(tf.log(1. - d_fake))
d_loss = -tf.reduce_mean(tf.log(d_real) + tf.log(1. - d_fake))
elif FLAGS.loss_name == "ALTERNATIVE":
g_loss = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(logits=d_fake_logit, labels=tf.ones_like(d_fake_logit)))
d_loss_real = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(logits=d_real_logit,
labels=tf.one_hot(tf.ones(shape=[FLAGS.batch_size], dtype=tf.uint8),
depth=2) *
tf.random_uniform([FLAGS.batch_size, 1], 0.7, 1.0)))
d_loss_fake = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(logits=d_fake_logit,
labels=tf.one_hot(tf.zeros(shape=[FLAGS.batch_size], dtype=tf.uint8),
depth=2) *
tf.random_uniform([FLAGS.batch_size, 1], 0.7, 1.0)))
d_loss = d_loss_real + d_loss_fake
elif FLAGS.loss_name == "WGAN-GP":
g_loss = -tf.reduce_mean(d_fake_logit)
d_loss_real = tf.reduce_mean(d_real_logit)
d_loss_fake = tf.reduce_mean(d_fake_logit)
d_loss = tf.reduce_mean(d_fake_logit) - tf.reduce_mean(d_real_logit)
epsilon = tf.random_uniform([FLAGS.batch_size, 1, 1, 1, 1], 0.0, 1.0)
x_hat = epsilon * x + (1 - epsilon) * x_fake
d_hat, d_hat_logit = discriminator.d_output_unit(x_hat, dropout_kept_prob, is_training)
ddx = tf.gradients(d_hat_logit, x_hat)[0]
ddx = tf.sqrt(tf.reduce_sum(tf.square(ddx), reduction_indices=[1, 2, 3, 4]))
ddx = tf.reduce_mean(tf.square(ddx - 1.0) * FLAGS.gradient_policy_scale)
d_loss = d_loss + ddx
if FLAGS.regularizer_scale > 0:
regularizer = tf.contrib.layers.l2_regularizer(scale=FLAGS.regularizer_scale)
reg_term = tf.contrib.layers.apply_regularization(regularizer, discriminator.get_weights)
d_loss += reg_term
if FLAGS.loss_name == "WGAN-GP":
d_batch_norm_update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS, scope="discriminator")
if d_batch_norm_update_ops:
d_loss = control_flow_ops.with_dependencies(d_batch_norm_update_ops, d_loss)
g_batch_norm_update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS, scope="while/generator")
if g_batch_norm_update_ops:
g_loss = control_flow_ops.with_dependencies(g_batch_norm_update_ops, g_loss)
d_loss_summary_op = tf.summary.scalar("d_loss/total", d_loss)
d_loss_real_summary_op = tf.summary.scalar("d_loss/real", d_loss_real)
d_loss_fake_summary_op = tf.summary.scalar("d_loss/fake", d_loss_fake)
d_loss_summary_op = tf.summary.merge([d_loss_real_summary_op, d_loss_fake_summary_op, d_loss_summary_op])
if FLAGS.loss_name == "WGAN-GP":
ddx_summary_op = tf.summary.scalar("d_loss/ddx", ddx)
d_loss_summary_op = tf.summary.merge([d_loss_summary_op, ddx_summary_op])
g_loss_summary_op = tf.summary.scalar("g_loss", g_loss)
add_variable_summaries(tf.identity(d_real_logit, name="d_activation/real"))
add_variable_summaries(tf.identity(d_fake_logit, name="d_activation/fake"))
increase_global_step = tf.assign(global_step, global_step + 1)
if FLAGS.learning_rate_decay_schedule != 0:
d_lr = tf.train.exponential_decay(FLAGS.d_lr, global_step, FLAGS.learning_rate_decay_schedule,
FLAGS.learning_rate_decay_rate, staircase=True)
g_lr_lstm = tf.train.exponential_decay(FLAGS.g_lr_lstm, global_step, FLAGS.learning_rate_decay_schedule,
FLAGS.learning_rate_decay_rate, staircase=True)
g_lr_conv = tf.train.exponential_decay(FLAGS.g_lr_conv, global_step, FLAGS.learning_rate_decay_schedule,
FLAGS.learning_rate_decay_rate, staircase=True)
else:
d_lr = FLAGS.d_lr
g_lr_lstm = FLAGS.g_lr_lstm
g_lr_conv = FLAGS.g_lr_conv
d_opt = tf.train.AdamOptimizer(d_lr, beta1=FLAGS.adam_b1, beta2=FLAGS.adam_b2)
g_opt_lstm = tf.train.MomentumOptimizer(g_lr_lstm, momentum=FLAGS.g_momentum_lstm)
#g_opt_lstm = tf.train.AdamOptimizer(g_lr_lstm, beta1=FLAGS.adam_b1, beta2=FLAGS.adam_b2)
g_opt_conv = tf.train.AdamOptimizer(g_lr_conv, beta1=FLAGS.adam_b1, beta2=FLAGS.adam_b2)
d_gvs = d_opt.compute_gradients(d_loss, var_list=list(discriminator.get_params()))
g_gvs_lstm = g_opt_lstm.compute_gradients(g_loss, var_list=list(generator.get_lstm_params()))
g_gvs_conv = g_opt_conv.compute_gradients(g_loss, var_list=list(generator.get_conv_params()))
# clip by individual norm
if FLAGS.d_gradient_clip:
d_gvs = [(tf.clip_by_norm(grad, clip_norm=FLAGS.d_clip_norm), var) for grad, var in d_gvs]
if FLAGS.g_gradient_clip_lstm:
#gradients, variables = list(zip(*g_gvs_lstm))
#gradients, _ = tf.clip_by_global_norm(gradients, clip_norm=FLAGS.g_clip_norm_lstm)
#g_gvs_lstm = [(grad, var) for grad, var in zip(gradients, variables)]
g_gvs_lstm = [(tf.clip_by_norm(grad, clip_norm=FLAGS.g_clip_norm_lstm), var) for grad, var in g_gvs_lstm]
if FLAGS.g_gradient_clip_conv:
g_gvs_conv = [(tf.clip_by_norm(grad, clip_norm=FLAGS.g_clip_norm_conv), var) for grad, var in g_gvs_conv]
add_gradient_summaries(d_gvs)
add_gradient_summaries(g_gvs_lstm)
add_gradient_summaries(g_gvs_conv)
d_solver = d_opt.apply_gradients(d_gvs)
g_solver_lstm = g_opt_lstm.apply_gradients(g_gvs_lstm)
g_solver_conv = g_opt_conv.apply_gradients(g_gvs_conv)
if FLAGS.loss_name in ["BASIC"]:
moving_mnist = read_data_sets("data/Moving_MNIST", train_size=FLAGS.train_size, rescale=False, num_channels=FLAGS.num_channels)
else:
moving_mnist = read_data_sets("data/Moving_MNIST", train_size=FLAGS.train_size, rescale=True, num_channels=FLAGS.num_channels)
mkdir(FLAGS.out_dir)
mkdir(FLAGS.log_dir)
if FLAGS.cleanup and FLAGS.ckpt_fn is None:
cleanup_dir(FLAGS.out_dir)
cleanup_dir(FLAGS.log_dir)
discriminator.create_summaries_for_variables()
generator.create_summaries_for_variables()
summary_op = tf.summary.merge_all()
sess = tf.Session()
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver(write_version=tf.train.SaverDef.V2)
writer = tf.summary.FileWriter(logdir=FLAGS.log_dir, graph=sess.graph)
restore_vars(saver, sess, save_dir, FLAGS.ckpt_fn)
iter = int(global_step.eval(sess))
start_iter = iter
json.dump(FLAGS.__flags, open(os.path.join(save_dir, "config.txt"), "w"))
summary_noises = generator.generate_noise()
num_iters_pretrain_d = int(25.0 * FLAGS.batch_size / FLAGS.train_size)
while True:
if iter > start_iter and iter % FLAGS.sample_every_n_batches == 0:
videos = sess.run(x_fake, feed_dict={noises: summary_noises, is_training: True})
save_gif_from_sampled_videos(videos, fn="{}/{}.gif".format(FLAGS.out_dir, str(iter).zfill(7)))
if iter > start_iter and iter % FLAGS.save_every_n_batches == 0:
print("Saving model begins...")
saver.save(sess, os.path.join(save_dir, "model"), global_step=iter)
print("Save completed!")
real_videos = moving_mnist.train.next_batch(FLAGS.batch_size)
if iter < num_iters_pretrain_d or iter % 500 == 0:
d_iters = 100
else:
d_iters = FLAGS.d_steps
g_iters = FLAGS.g_steps
for j in range(d_iters):
_, d_loss_curr, d_loss_summary = sess.run([d_solver, d_loss, d_loss_summary_op],
feed_dict={noises: generator.generate_noise(),
x: real_videos,
dropout_kept_prob: FLAGS.dropout_kept_prob,
is_training: True})
if d_loss_curr < FLAGS.d_threshold:
break
for j in range(g_iters):
_, _, g_loss_curr, g_loss_summary = sess.run([g_solver_lstm, g_solver_conv, g_loss, g_loss_summary_op],
feed_dict={noises: generator.generate_noise(),
dropout_kept_prob: FLAGS.dropout_kept_prob,
is_training: True})
if g_loss_curr < FLAGS.g_threshold:
break
print("Step {}: D loss {:.6f}, G loss {:.6f}".format(iter, d_loss_curr, g_loss_curr))
if iter > start_iter and iter % FLAGS.save_every_n_batches == 0:
summary = sess.run(summary_op, feed_dict={noises: summary_noises,
x: real_videos,
dropout_kept_prob: FLAGS.dropout_kept_prob,
is_training: True})
writer.add_summary(summary, global_step=iter)
writer.flush()
else:
writer.add_summary(d_loss_summary, global_step=iter)
writer.add_summary(g_loss_summary, global_step=iter)
sess.run([increase_global_step])
iter += 1
if iter == FLAGS.total_batches:
break
if __name__ == "__main__":
tf.logging.set_verbosity(tf.logging.INFO)
tf.app.run()