/
main.py
executable file
·224 lines (178 loc) · 9.78 KB
/
main.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
# -*- coding: utf-8 -*-
import os
import sys
import time
import argparse
import numpy as np
import matplotlib.pyplot as plt
import matplotlib.gridspec as gridspec
import tensorflow as tf
from pprint import pprint
from tensorflow.keras import optimizers
from models import get_generator, get_discriminator
from dataset import get_dataset_and_info
# Define both loss function
def hinge_loss_g(generated_output):
return -tf.reduce_mean(generated_output)
def hinge_loss_d(real_output, generated_output):
real_loss = tf.reduce_mean(tf.nn.relu(1.0 - real_output))
generated_loss = tf.reduce_mean(tf.nn.relu(1 + generated_output))
return real_loss + generated_loss
def cross_entropy_g(generated_output):
return tf.reduce_mean(tf.losses.binary_crossentropy(tf.ones_like(generated_output), generated_output))
def cross_entropy_d(real_output, generated_output):
real_loss = tf.reduce_mean(
tf.losses.binary_crossentropy(tf.ones_like(real_output), real_output))
generated_loss = tf.reduce_mean(
tf.losses.binary_crossentropy(tf.zeros_like(generated_output), generated_output))
total_loss = real_loss + generated_loss
return total_loss
class Trainer(object):
def __init__(self, config):
self.config = config
self.ds_train, self.info = get_dataset_and_info(self.config)
self.steps_per_epoch = self.info['num_records'] // self.config['batch_size']
self.generator = get_generator(self.info["num_classes"])
self.discriminator = get_discriminator(self.info["num_classes"])
if self.config['loss'] == "cross_entropy":
print("use ce loss")
self.gloss_fn = cross_entropy_g
self.dloss_fn = cross_entropy_d
elif self.config['loss'] == "hinge_loss":
print("use hinge loss")
self.gloss_fn = hinge_loss_g
self.dloss_fn = hinge_loss_d
else:
raise ValueError('Unsupported loss type')
lr_fn_G = tf.optimizers.schedules.ExponentialDecay(1e-4, self.steps_per_epoch, decay_rate=0.99, staircase=True)
lr_fn_D = tf.optimizers.schedules.ExponentialDecay(4e-4, self.steps_per_epoch * self.config['update_ratio'], decay_rate=0.99, staircase=True)
self.generator_optimizer = optimizers.Adam(learning_rate=lr_fn_G, beta_1=0.)
self.discriminator_optimizer = optimizers.Adam(learning_rate=lr_fn_D, beta_1=0.)
# build model to get target the name of tensors
self.generator.build(input_shape=[(self.config['batch_size'], self.config['z_dim']), (self.config['batch_size'])])
self.var_name_list = [var.name for var in self.generator.trainable_variables]
# metrics
self.metrics = {}
self.metrics['G_loss'] = tf.keras.metrics.Mean(
'generator_loss', dtype=tf.float32)
self.metrics['D_loss'] = tf.keras.metrics.Mean(
'discriminator_loss', dtype=tf.float32)
for name in self.var_name_list:
self.metrics[name] = tf.keras.metrics.Mean(
name, dtype=tf.float32)
self.random_vector = tf.random.normal([config['batch_size'], config['z_dim']])
self.fix_label = tf.random.uniform((self.config['batch_size'],), 0, self.info['num_classes'], dtype=tf.int32)
@tf.function
def train_step(self, images, labels):
# Update D. n times per update of G.
average_loss = tf.constant(0., dtype=tf.float32)
for _ in range(self.config['update_ratio']):
noise = tf.random.normal([images.shape[0], self.config['z_dim']])
fake_labels = tf.random.uniform((labels.shape[0],), 0, self.info['num_classes'], dtype=tf.int32)
generated_images = self.generator([noise, fake_labels], training=True)
with tf.GradientTape() as disc_tape:
real_output = self.discriminator([images, labels], training=True)
generated_output = self.discriminator([generated_images, fake_labels], training=True)
disc_loss = self.dloss_fn(real_output, generated_output)
# for computing average loss of this train_step
average_loss = average_loss + disc_loss
gradients_of_discriminator = disc_tape.gradient(
disc_loss, self.discriminator.trainable_variables)
self.discriminator_optimizer.apply_gradients(
zip(gradients_of_discriminator, self.discriminator.trainable_variables))
# Update G.
noise = tf.random.normal([labels.shape[0], self.config['z_dim']])
fake_labels = tf.random.uniform((labels.shape[0],), 0, self.info['num_classes'], dtype=tf.int32)
with tf.GradientTape() as gen_tape:
generated_images = self.generator([noise, fake_labels], training=True)
generated_output = self.discriminator([generated_images, fake_labels], training=True)
gen_loss = self.gloss_fn(generated_output)
gradients_of_generator = gen_tape.gradient(
gen_loss, self.generator.trainable_variables)
self.generator_optimizer.apply_gradients(zip(gradients_of_generator, self.generator.trainable_variables))
# for track gradients of specific tensors
zip_of_generator = list(zip(gradients_of_generator, self.generator.trainable_variables))
name_to_grads = {var[1].name: var[0] for var in zip_of_generator}
self.metrics['G_loss'](gen_loss)
self.metrics['D_loss'](average_loss / self.config['update_ratio'])
for name in self.var_name_list:
self.metrics[name](name_to_grads[name])
def train(self):
log_dir = 'logs/{}'.format(self.config['path_root'])
for epoch in range(self.config['epoch']):
start_time = time.time()
for i, (images, labels) in enumerate(self.ds_train):
self.train_step(images, labels)
with tf.summary.create_file_writer(log_dir).as_default():
tf.summary.scalar(
'Generator Loss', self.metrics['G_loss'].result(), step=epoch)
tf.summary.scalar(
'Discriminator Loss', self.metrics['D_loss'].result(), step=epoch)
for name in self.var_name_list:
tf.summary.scalar('grads_norm/{}'.format(name), self.metrics[name].result(), step=epoch)
# save checkpoints every 20 epochs
if (epoch+1) % 10 == 0 and not self.config['debug']:
print("save checkpoint ...")
ckpt_path = 'checkpoints/{}/epoch_{}'.format(self.config['path_root'], epoch)
self.generator.save_weights(ckpt_path, save_format='tf')
if (epoch+1) % 5 == 0:
self.save_sample_images(epoch)
template = 'Epoch({:.2f} sec): {}, gen_loss: {}, disc_loss: {}'
print(template.format(time.time()-start_time, epoch+1, self.metrics['G_loss'].result(), self.metrics['D_loss'].result()))
sys.stdout.flush()
self.metrics['G_loss'].reset_states()
self.metrics['D_loss'].reset_states()
for name in self.var_name_list:
self.metrics[name].reset_states()
def save_sample_images(self, epoch):
img_path = os.path.abspath('images/{}/'.format(self.config['path_root']))
if not os.path.exists(img_path):
os.makedirs(img_path)
sample_img = self.generator(
[self.random_vector, self.fix_label], training=False)
samples = np.uint8(sample_img*127.5+128).clip(0, 255)
fig = plt.figure(figsize=(8, 8))
gs = gridspec.GridSpec(8, 8)
gs.update(wspace=0.05, hspace=0.05)
for i, sample in enumerate(samples[:self.config['num_sample']]):
ax = plt.subplot(gs[i])
plt.axis('off')
ax.set_xticklabels([])
ax.set_yticklabels([])
ax.set_aspect('equal')
plt.imshow(sample)
plt.savefig(os.path.join(img_path, 'epoch-{}.png'.format(str(epoch+1).zfill(3))), bbox_inches='tight')
plt.close(fig)
def main(config):
trainer = Trainer(config)
trainer.train()
if __name__ == '__main__':
# Handle cuDNN failure issue.
physical_devices = tf.config.experimental.list_physical_devices('GPU')
assert len(physical_devices) > 0, "Not enough GPU hardware devices available"
for physical_device in physical_devices:
tf.config.experimental.set_memory_growth(physical_device, True)
parser = argparse.ArgumentParser(description='Experiment parameters')
parser.add_argument("--debug", action='store_true', default=False,
help="whether to use debug mode")
parser.add_argument("--path_root", default='test',
help="path root of images, checkpoints, and logs")
parser.add_argument("--data_path", default='/home/yct/data/ILSVRC2017_CLS-LOC/ILSVRC/Data/CLS-LOC', help="path to the dataset")
parser.add_argument("--z_dim", type=int, default=128,
help="dimension of noise")
parser.add_argument('-b', "--batch_size", type=int, default=64,
help="Batch size")
parser.add_argument('-l', "--loss", default="hinge_loss",
help="loss function")
parser.add_argument('-e', '--epoch', type=int, default=5,
help="training epochs")
parser.add_argument('-u', '--update_ratio', type=int, default=1,
help="updating ratio of Discriminator to Generator")
parser.add_argument('-d', '--data_size', type=int, default=-1,
help="data size. -1 means full data")
parser.add_argument('-n', '--num_sample', type=int, default=64,
help='the num of samples Generator creates')
args, unknown = parser.parse_known_args()
config = {attr: getattr(args, attr) for attr in dir(args) if attr[0]!='_'}
pprint(config)
main(config)