-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathtrain.py
196 lines (160 loc) · 8.73 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
import os
import time
import datetime
from tqdm import tqdm
from absl import app
from absl import flags
import tensorflow as tf
from tensorflow.keras import mixed_precision
from data_loader import *
from utils import *
from model import build_model
FLAGS = flags.FLAGS
flags.DEFINE_string("attr_path",
"data/celeba/list_attr_celeba.txt",
"path to the attribute label")
flags.DEFINE_list("selected_attrs",
"Black_Hair, Blond_Hair, Brown_Hair, Male, Young",
"attributes for training")
flags.DEFINE_integer("c_dim", 5, "dimension of domain labels")
flags.DEFINE_integer("batch_size", 16, "mini-batch size")
flags.DEFINE_string("ckpt_dir", "ckpts/train/", "path to the checkpoint dir")
flags.DEFINE_string("tfrecord_dir", "data/celeba/tfrecords/", "path to the tfrecord dir")
flags.DEFINE_string("test_result_dir", "test_results/", "path to the test result dir")
flags.DEFINE_string("logdir", "logs/", "path to the log dir")
flags.DEFINE_integer("num_epochs", 20, "number of epopchs to train")
flags.DEFINE_integer("num_epochs_decay", 10, "number of epochs to start lr decay")
flags.DEFINE_integer("num_iters", 200000, "number of total iterations for training D")
flags.DEFINE_integer("num_iters_decay", 100000, "number of iterations for decaying lr")
flags.DEFINE_float("lambda_cls", 1.0, "weight for domain classification loss")
flags.DEFINE_float("lambda_rec", 10.0, "weight for reconstruction loss")
flags.DEFINE_float("lambda_gp", 10.0, "weight for gradient penalty loss")
flags.DEFINE_integer("model_save_epoch", 1, "to save model every specified epochs")
flags.DEFINE_integer("num_critic_updates", 5, "number of a Discriminator updates "
"every time a generator updates")
flags.DEFINE_float("g_lr", 0.0001, "learning rate for the generator")
flags.DEFINE_float("d_lr", 0.0001, "learning rate for the discriminator")
flags.DEFINE_float("beta1", 0.5, "beta1 for Adam optimizer")
flags.DEFINE_float("beta2", 0.999, "beta2 for Adam optimizer")
flags.DEFINE_integer("num_test", 10, "number of test examples")
#flags.DEFINE_bool("use_mp", True, "whether to use mixed precision for training")
#tf.config.experimental.enable_tensor_float_32_execution(enabled=False)
def main(argv):
#if FLAGS.use_mp:
# mixed_precision.set_global_policy('mixed_float16')
gpus = tf.config.experimental.list_physical_devices('GPU')
for gpu in gpus:
tf.config.experimental.set_memory_growth(gpu, True)
AUTOTUNE = tf.data.experimental.AUTOTUNE
os.makedirs(FLAGS.ckpt_dir, exist_ok=True)
os.makedirs(FLAGS.logdir, exist_ok=True)
os.makedirs(FLAGS.test_result_dir, exist_ok=True)
_, _, test_imgs, test_lbls = get_data(FLAGS.attr_path,
FLAGS.selected_attrs)
# Prepare the dataset for training and testing
train_dir = os.path.join(FLAGS.tfrecord_dir, "train")
test_dir = os.path.join(FLAGS.tfrecord_dir, "test")
# For training
train_dataset = tf.data.Dataset.list_files(os.path.join(train_dir, "*.tfrecord"))
train_dataset = train_dataset.interleave(tf.data.TFRecordDataset,
num_parallel_calls=AUTOTUNE,
deterministic=False)
train_dataset = train_dataset.map(parse_tfrecords)
train_dataset = train_dataset.map(preprocess_for_training,
num_parallel_calls=AUTOTUNE)
train_dataset = train_dataset.batch(batch_size=FLAGS.batch_size, drop_remainder=True)
train_dataset = train_dataset.prefetch(buffer_size=AUTOTUNE)
# Get fixed inputs for testing and debugging.
c_fixed_trg_list = create_labels(test_lbls[:FLAGS.num_test],
FLAGS.c_dim,
FLAGS.selected_attrs)
# Build the generator and discriminator
#gen, disc = build_model(FLAGS.c_dim, FLAGS.use_mp)
gen, disc = build_model(FLAGS.c_dim, False)
# Define the optimizers for the generator and the discriminator
gen_opt = tf.keras.optimizers.Adam(FLAGS.g_lr, FLAGS.beta1, FLAGS.beta2)
disc_opt = tf.keras.optimizers.Adam(FLAGS.d_lr, FLAGS.beta1, FLAGS.beta2)
#if FLAGS.use_mp:
# gen_opt = mixed_precision.LossScaleOptimizer(gen_opt)
# disc_opt = mixed_precision.LossScaleOptimizer(disc_opt)
# Set the checkpoint and the checkpoint manager.
ckpt = tf.train.Checkpoint(epoch=tf.Variable(0, dtype=tf.int64),
gen=gen,
disc=disc,
gen_opt=gen_opt,
disc_opt=disc_opt)
ckpt_manager = tf.train.CheckpointManager(ckpt,
FLAGS.ckpt_dir,
max_to_keep=5)
# If a checkpoint exists, restore the latest checkpoint.
if ckpt_manager.latest_checkpoint:
ckpt.restore(ckpt_manager.latest_checkpoint)
print("Latest checkpoint is restored!")
# Create a summary writer to track the losses
summary_writer = tf.summary.create_file_writer(
os.path.join(FLAGS.logdir,
datetime.datetime.now().strftime("%Y%m%d-%H%M%S"))
)
d_loss_list, g_loss_list = initialize_loss_trackers()
#train_d, train_g = define_train_loop(FLAGS.use_mp)
train_d, train_g = define_train_loop(False)
iters_per_epoch = FLAGS.num_iters // FLAGS.num_epochs
diff_iter = FLAGS.num_iters - FLAGS.num_iters_decay
# Train the discriminator and the generator
while ckpt.epoch < FLAGS.num_epochs:
ckpt.epoch.assign_add(1)
step = tf.constant(0)
reset_loss_trackers(d_loss_list)
reset_loss_trackers(g_loss_list)
#if ckpt.epoch > FLAGS.num_epochs_decay:
# update_lr(gen_opt, disc_opt, FLAGS.num_epochs, ckpt.epoch, FLAGS.g_lr, FLAGS.d_lr)
start = time.time()
for x_real, label_org, label_trg in tqdm(train_dataset):
step += 1
if step.numpy() > FLAGS.num_iters_decay:
update_lr_by_iter(gen_opt, disc_opt, step, diff_iter, FLAGS.g_lr, FLAGS.d_lr)
d_losses = train_d(disc,
gen,
x_real,
label_org,
label_trg,
FLAGS.lambda_cls,
FLAGS.lambda_gp,
disc_opt)
update_loss_trackers(d_loss_list, d_losses)
if step.numpy() % FLAGS.num_critic_updates == 0:
g_losses = train_g(disc,
gen,
x_real,
label_org,
label_trg,
FLAGS.lambda_cls,
FLAGS.lambda_rec,
gen_opt)
update_loss_trackers(g_loss_list, g_losses)
if step.numpy() == iters_per_epoch:
break
#if step.numpy() % 100 == 0:
# fpath = os.path.join(FLAGS.test_result_dir, "{}-images.jpg".format(step.numpy()))
# save_test_results(gen, test_imgs[:FLAGS.num_test], c_fixed_trg_list, fpath)
end = time.time()
print_log(ckpt.epoch.numpy(), start, end, d_losses, g_losses)
# keep the log for the losses
with summary_writer.as_default():
tf.summary.scalar("d_loss_real", d_loss_list[0].result(), step=ckpt.epoch)
tf.summary.scalar("d_loss_fake", d_loss_list[1].result(), step=ckpt.epoch)
tf.summary.scalar("d_loss_gp", d_loss_list[2].result(), step=ckpt.epoch)
tf.summary.scalar("d_loss_cls", d_loss_list[3].result(), step=ckpt.epoch)
tf.summary.scalar("d_loss", d_loss_list[4].result(), step=ckpt.epoch)
tf.summary.scalar("g_loss_fake", g_loss_list[0].result(), step=ckpt.epoch)
tf.summary.scalar("g_loss_rec", g_loss_list[1].result(), step=ckpt.epoch)
tf.summary.scalar("g_loss_cls", g_loss_list[2].result(), step=ckpt.epoch)
tf.summary.scalar("g_loss", g_loss_list[3].result(), step=ckpt.epoch)
# test the generator model and save the results for each epoch
fpath = os.path.join(FLAGS.test_result_dir, "{}-images.jpg".format(ckpt.epoch.numpy()))
save_test_results(gen, test_imgs[:FLAGS.num_test], c_fixed_trg_list, fpath)
if (ckpt.epoch) % FLAGS.model_save_epoch == 0:
ckpt_save_path = ckpt_manager.save()
print("Saving a checkpoint for epoch {} at {}".format(ckpt.epoch.numpy(), ckpt_save_path))
if __name__ == "__main__":
app.run(main)