-
Notifications
You must be signed in to change notification settings - Fork 7
/
train.py
187 lines (152 loc) · 8.68 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
""" USAGE
python ./train.py --train_src_dir ./datasets/horse2zebra/trainA --train_tar_dir ./datasets/horse2zebra/trainB --test_src_dir ./datasets/horse2zebra/testA --test_tar_dir ./datasets/horse2zebra/testB
"""
import os
import argparse
import datetime
import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt
from modules.cut_model import CUT_model
from utils import create_dir, load_image
def ArgParse():
# Parse command line arguments
parser = argparse.ArgumentParser(description='CUT training usage.')
# Training
parser.add_argument('--mode', help="Model's mode be one of: 'cut', 'fastcut'", type=str, default='cut', choices=['cut', 'fastcut'])
parser.add_argument('--epochs', help='Number of training epochs', type=int, default=400)
parser.add_argument('--batch_size', help='Training batch size', type=int, default=1)
parser.add_argument('--beta_1', help='First Momentum term of adam', type=float, default=0.5)
parser.add_argument('--beta_2', help='Second Momentum term of adam', type=float, default=0.999)
parser.add_argument('--lr', help='Initial learning rate for adam', type=float, default=0.0002)
parser.add_argument('--lr_decay_rate', help='lr_decay_rate', type=float, default=0.9)
parser.add_argument('--lr_decay_step', help='lr_decay_step', type=int, default=100000)
# Define data
parser.add_argument('--out_dir', help='Outputs folder', type=str, default='./output')
parser.add_argument('--train_src_dir', help='Train-source dataset folder', type=str, default='./datasets/horse2zebra/trainA')
parser.add_argument('--train_tar_dir', help='Train-target dataset folder', type=str, default='./datasets/horse2zebra/trainB')
parser.add_argument('--test_src_dir', help='Test-source dataset folder', type=str, default='./datasets/horse2zebra/testA')
parser.add_argument('--test_tar_dir', help='Test-target dataset folder', type=str, default='./datasets/horse2zebra/testB')
# Misc
parser.add_argument('--ckpt', help='Resume training from checkpoint', type=str)
parser.add_argument('--save_n_epoch', help='Every n epochs to save checkpoints', type=int, default=5)
parser.add_argument('--impl', help="(Faster)Custom op use:'cuda'; (Slower)Tensorflow op use:'ref'", type=str, default='ref', choices=['ref', 'cuda'])
args = parser.parse_args()
# Check arguments
assert args.lr > 0
assert args.epochs > 0
assert args.batch_size > 0
assert args.save_n_epoch > 0
assert os.path.exists(args.train_src_dir), 'Error: Train source dataset does not exist.'
assert os.path.exists(args.train_tar_dir), 'Error: Train target dataset does not exist.'
assert os.path.exists(args.test_src_dir), 'Error: Test source dataset does not exist.'
assert os.path.exists(args.test_tar_dir), 'Error: Test target dataset does not exist.'
return args
def main(args):
# Create datasets
train_dataset, test_dataset = create_dataset(args.train_src_dir,
args.train_tar_dir,
args.test_src_dir,
args.test_tar_dir,
args.batch_size)
# Get image shape
source_image, target_image = next(iter(train_dataset))
source_shape = source_image.shape[1:]
target_shape = target_image.shape[1:]
# Create model
cut = CUT_model(source_shape, target_shape, cut_mode=args.mode, impl=args.impl)
# Define learning rate schedule
lr_schedule = tf.keras.optimizers.schedules.ExponentialDecay(initial_learning_rate=args.lr,
decay_steps=args.lr_decay_step,
decay_rate=args.lr_decay_rate,
staircase=True)
# Compile model
cut.compile(G_optimizer = tf.keras.optimizers.Adam(learning_rate=lr_schedule, beta_1=args.beta_1, beta_2=args.beta_2),
F_optimizer = tf.keras.optimizers.Adam(learning_rate=lr_schedule, beta_1=args.beta_1, beta_2=args.beta_2),
D_optimizer = tf.keras.optimizers.Adam(learning_rate=lr_schedule, beta_1=args.beta_1, beta_2=args.beta_2),)
# Restored from previous checkpoints, or initialize checkpoints from scratch
if args.ckpt is not None:
latest_ckpt = tf.train.latest_checkpoint(args.ckpt)
cut.load_weights(latest_ckpt)
initial_epoch = int(latest_ckpt[-3:])
print(f"Restored from {latest_ckpt}.")
else:
initial_epoch = 0
print("Initializing from scratch...")
# Create folders to store the output information
result_dir = f'{args.out_dir}/images'
checkpoint_dir = f'{args.out_dir}/checkpoints'
log_dir = f'{args.out_dir}/logs/{datetime.datetime.now().strftime("%Y%m%d-%H%M%S")}'
# Create validating callback to generate output image every epoch
plotter_callback = GANMonitor(cut.netG, test_dataset, result_dir)
# Create checkpoint callback to save model's checkpoints every n epoch (default 5)
# "period" to save every n epochs, "save_freq" to save every n batches
checkpoint_callback = tf.keras.callbacks.ModelCheckpoint(filepath=checkpoint_dir+'/{epoch:03d}', period=args.save_n_epoch, verbose=1)
# Create tensorboard callback to log losses every epoch
tensorboard_callback = tf.keras.callbacks.TensorBoard(log_dir=log_dir)
# Train cut model
cut.fit(train_dataset,
epochs=args.epochs,
initial_epoch=initial_epoch,
callbacks=[plotter_callback, checkpoint_callback, tensorboard_callback],
verbose=1)
def create_dataset(train_src_folder,
train_tar_folder,
test_src_folder,
test_tar_folder,
batch_size):
""" Create tf.data.Dataset.
"""
# Create train dataset
train_src_dataset = tf.data.Dataset.list_files([train_src_folder+'/*.jpg', train_src_folder+'/*.png'], shuffle=True)
train_src_dataset = (
train_src_dataset.map(load_image, num_parallel_calls=tf.data.experimental.AUTOTUNE)
.batch(batch_size, drop_remainder=True)
.prefetch(tf.data.experimental.AUTOTUNE)
)
train_tar_dataset = tf.data.Dataset.list_files([train_tar_folder+'/*.jpg', train_tar_folder+'/*.png'], shuffle=True)
train_tar_dataset = (
train_tar_dataset.map(load_image, num_parallel_calls=tf.data.experimental.AUTOTUNE)
.batch(batch_size, drop_remainder=True)
.prefetch(tf.data.experimental.AUTOTUNE)
)
train_dataset = tf.data.Dataset.zip((train_src_dataset, train_tar_dataset))
# Create test dataset
test_src_dataset = tf.data.Dataset.list_files([test_src_folder+'/*.jpg', test_src_folder+'/*.png'])
test_src_dataset = (
test_src_dataset.map(load_image, num_parallel_calls=tf.data.experimental.AUTOTUNE)
.batch(batch_size, drop_remainder=True)
.prefetch(tf.data.experimental.AUTOTUNE)
)
test_tar_dataset = tf.data.Dataset.list_files([test_tar_folder+'/*.jpg', test_tar_folder+'/*.png'])
test_tar_dataset = (
test_tar_dataset.map(load_image, num_parallel_calls=tf.data.experimental.AUTOTUNE)
.batch(batch_size, drop_remainder=True)
.prefetch(tf.data.experimental.AUTOTUNE)
)
test_dataset = tf.data.Dataset.zip((test_src_dataset, test_tar_dataset))
return train_dataset, test_dataset
class GANMonitor(tf.keras.callbacks.Callback):
""" A callback to generate and save images after each epoch
"""
def __init__(self, generator, test_dataset, out_dir, num_img=2):
self.num_img = num_img
self.generator = generator
self.test_dataset = test_dataset
self.out_dir = create_dir(out_dir)
def on_epoch_end(self, epoch, logs=None):
_, ax = plt.subplots(self.num_img, 4, figsize=(20, 10))
[ax[0, i].set_title(title) for i, title in enumerate(['Source', "Translated", "Target", "Identity"])]
for i, (source, target) in enumerate(self.test_dataset.take(self.num_img)):
translated = self.generator(source)[0].numpy()
translated = (translated * 127.5 + 127.5).astype(np.uint8)
source = (source[0] * 127.5 + 127.5).numpy().astype(np.uint8)
idt = self.generator(target)[0].numpy()
idt = (idt * 127.5 + 127.5).astype(np.uint8)
target = (target[0] * 127.5 + 127.5).numpy().astype(np.uint8)
[ax[i, j].imshow(img) for j, img in enumerate([source, translated, target, idt])]
[ax[i, j].axis("off") for j in range(4)]
plt.savefig(f'{self.out_dir}/epoch={epoch + 1}.png')
plt.close()
if __name__ == '__main__':
main(ArgParse())