-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtrain.py
286 lines (222 loc) · 10.7 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
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
# -*- utf-8 -*-
"""
For uint16
"""
import os
import shutil
import sys
import time
from xml.dom.minidom import parse
import numpy as np
import tensorflow as tf
from PIL import Image
from model import CustomEfficientNetB0
from utils.XMLUtil import readXML
from utils.Dataloader import image2tfrecord, load_dataset
input_height, input_width = 512, 512
batch_size = 1
buffer_size = batch_size * 4
checkpoint_path = 'checkpoints/train/'
LAMBDA = 100
loss_obj = tf.keras.losses.BinaryCrossentropy(from_logits=True)
def gradient_penalty(discriminator, real_image, fake_image):
batch_size = real_image.shape[0]
# [b, h, w, c]
t = tf.random.uniform([batch_size, 1, 1, 1])
# [b, 1, 1, 1] => [b, h, w, c]
t = tf.broadcast_to(t, real_image.shape)
interpolate = t * real_image + (1 - t) * fake_image
with tf.GradientTape() as tape:
tape.watch([interpolate])
d_interpolate_logits = discriminator(interpolate, training=True)
grads = tape.gradient(d_interpolate_logits, interpolate)
# grads: [b, h, w, c] => [b, -1]
grads = tf.reshape(grads, [grads.shape[0], -1])
gp = tf.norm(grads, axis=1)
gp = tf.reduce_mean((gp - 1) ** 2)
return gp
def generator_loss(generator, disc_generated_output, gen_output, target):
gan_loss = -tf.reduce_mean(disc_generated_output)
l1_loss = tf.reduce_mean(tf.abs(target - gen_output))
# Parameter regularization
loss_regularizations = []
for p in generator.trainable_variables:
loss_regularizations.append(tf.nn.l2_loss(p))
loss_regularization = tf.reduce_sum(tf.stack(loss_regularizations))
total_gen_loss = gan_loss + \
(LAMBDA * l1_loss) + .001 * loss_regularization
return total_gen_loss, gan_loss, l1_loss, loss_regularization
def discriminator_loss(discriminator, real_image, fake_image, disc_real_output, disc_generated_output):
d_loss_real = -tf.reduce_mean(disc_real_output)
d_loss_fake = tf.reduce_mean(disc_generated_output)
gp = gradient_penalty(discriminator, real_image, fake_image)
loss = d_loss_real + d_loss_fake + 10. * gp
return loss, gp
def train_loop(args, output_path):
# Load train and val sets
if not args['from_tfrecord']:
tfrecord_path = os.path.join(output_path, 'tfrecord')
else:
tfrecord_path = os.path.join(input_path, 'tfrecord')
train_names = os.listdir(tfrecord_path)
train_names = [os.path.join(tfrecord_path, name) for name in train_names]
train_set = load_dataset(train_names, batch_size=batch_size)
# Log
logdir = os.path.join(output_path, 'logs/')
file_writer = tf.summary.create_file_writer(logdir + 'metrics')
file_writer.set_as_default()
if args['visualization']:
import matplotlib.pyplot as plt
vis_path = os.path.join(output_path, 'visualizations')
if not os.path.exists(vis_path):
os.makedirs(vis_path)
# Generator: hazy --> dehazy
generator = CustomEfficientNetB0.create_generator(norm_type='instancenorm')
# Discriminator: real or fake
discriminator = CustomEfficientNetB0.create_discriminator(
input_channels=args['input_channels'], norm_type='instancenorm', target=False)
# Optimizers
lr_schedule = tf.keras.optimizers.schedules.ExponentialDecay(
args['initial_learning_rate'],
decay_steps=args['decay_steps'],
decay_rate=args['decay_rate']
)
generator_optimizer = tf.keras.optimizers.Adam(lr_schedule, beta_1=0.5)
discriminator_optimizer = tf.keras.optimizers.Adam(lr_schedule, beta_1=0.5)
# Checkpoints and manager
ckpt = tf.train.Checkpoint(
generator=generator,
discriminator=discriminator,
generator_optimizer=generator_optimizer,
discriminator_optimizer=discriminator_optimizer,
)
ckpt_manager = tf.train.CheckpointManager(
ckpt, os.path.join(output_path, args['checkpoint_path']), max_to_keep=5)
# if restoration is enabled and a checkpoint exists, restore the latest checkpoint.
if args['restore'] and ckpt_manager.latest_checkpoint:
ckpt.restore(ckpt_manager.latest_checkpoint)
print('Latest checkpoint restored, from {}'.format(
ckpt_manager.latest_checkpoint))
def _random_crop(record):
hazy, gt = record['image'], record['label']
images = tf.concat([hazy, gt], axis=-1)
images = tf.image.resize_with_pad(images, input_height + input_height // 2,
input_width + input_width // 2, method=tf.image.ResizeMethod.NEAREST_NEIGHBOR)
images = tf.image.random_crop(
images, [batch_size, input_height, input_width, args['input_channels'] * 2])
images = tf.image.random_flip_left_right(images)
images = tf.image.random_flip_up_down(images)
hazy, gt = images[..., :args['input_channels']
], images[..., args['input_channels']:]
record['image'], record['label'] = hazy, gt
return record
def _est_ale(hazy):
hazy_shape = tf.shape(hazy)
hsv = tf.image.rgb_to_hsv(hazy)
hsv_shape = tf.shape(hsv)
hsv = tf.reshape(hsv[..., 2], [-1])
idx = tf.argmax(hsv)
ale = tf.reshape(hazy, [-1, hazy_shape[-1]])[idx]
ale = ale[tf.newaxis, tf.newaxis, tf.newaxis, ...]
return ale
@tf.function
def _train_step(hazy, gt, ale):
# hazy: 0 ... 1, gt: 0 mean, 1 std
with tf.GradientTape(persistent=True) as tape:
dehazy, rtme, dehazy0, tme = generator(
[hazy * 255., ale * 255.], training=True) # 0 ... 255 --> -1.5 ... 1.5
disc_real_output = discriminator(gt, training=True)
disc_generated_output = discriminator(dehazy, training=True)
gen_total_loss, gen_gan_loss, gen_l1_loss, l_reg = generator_loss(
generator, disc_generated_output, dehazy, gt)
disc_loss, gp = discriminator_loss(
discriminator, gt, dehazy, disc_real_output, disc_generated_output)
generator_gradients = tape.gradient(
gen_total_loss, generator.trainable_variables)
discriminator_gradients = tape.gradient(
disc_loss, discriminator.trainable_variables)
generator_optimizer.apply_gradients(
zip(generator_gradients, generator.trainable_variables))
discriminator_optimizer.apply_gradients(
zip(discriminator_gradients, discriminator.trainable_variables))
return gen_total_loss, gen_gan_loss, gen_l1_loss, l_reg, disc_loss, gp
def standardize(tensor):
# assert tf.less_equal(tf.shape(tensor)[0], 1) # batch_size should be 1.
mean, std = tf.math.reduce_mean(tensor), tf.math.reduce_std(tensor)
return (tensor - mean) / std
def normalize(im):
return (im - im.min()) / (im.max() - im.min())
with file_writer.as_default():
step = 0
for epoch in range(args['num_epochs']):
start = time.time()
train_set_ = train_set.map(_random_crop).shuffle(
buffer_size=buffer_size)
for record in train_set_:
hazy, gt = record['image'], record['label']
ale = _est_ale(hazy)
gen_total_loss, gen_gan_loss, gen_l1_loss, l_reg, disc_loss, gp = _train_step(
hazy, standardize(gt), ale)
step += 1
tf.summary.scalar('gen_total_loss', gen_total_loss, step=step)
tf.summary.scalar('gen_gan_loss', gen_gan_loss, step=step)
tf.summary.scalar('gen_l1_loss', gen_l1_loss, step=step)
tf.summary.scalar('l_reg', l_reg, step=step)
tf.summary.scalar('disc_loss', disc_loss, step=step)
tf.summary.scalar('gp', gp, step=step)
print('Step {}, gen_total_loss: {}, gen_gan_loss: {}, gen_l1_loss: {}, l_reg: {}, disc_loss: {}, gp: {}'.format(
step, gen_total_loss.numpy(), gen_gan_loss.numpy(), gen_l1_loss.numpy(), l_reg.numpy(), disc_loss.numpy(), gp.numpy()))
print('Time taken for epoch {} is {} sec'.format(
epoch + 1, time.time() - start))
ckpt_save_path = ckpt_manager.save()
print('Checkpoint for epoch {} saved at {}'.format(
epoch + 1, ckpt_save_path))
# Train visualization
if args['visualization']:
for record in train_set_.take(1):
hazy, gt = record['image'], record['label']
ale = _est_ale(hazy)
dehazy, rtme, dehazy0, tme = generator(
[hazy * 255., ale * 255.], training=False)
hazy, gt, ale = hazy.numpy()[0], gt.numpy()[
0], ale.numpy()[0]
dehazy, rtme, dehazy0, tme = dehazy.numpy()[0], rtme.numpy()[
0], dehazy0.numpy()[0], tme.numpy()[0]
np.savez(os.path.join(
vis_path, 'epoch_{}.npz'.format(epoch)),
hazy=hazy,
gt=gt,
ale=ale,
dehazy=dehazy,
rtme=rtme,
dehazy0=dehazy0,
tme=tme)
dehazy, dehazy0 = normalize(dehazy), normalize(dehazy0)
ale = np.broadcast_to(ale, hazy.shape)
rtme, tme = np.broadcast_to(
rtme, hazy.shape), np.broadcast_to(tme, hazy.shape)
rtme = (rtme - rtme.min()) / (rtme.max() - rtme.min())
tme = (tme - tme.min()) / (tme.max() - tme.min())
imgs = [hazy, gt, ale, np.ones_like(hazy),
dehazy, rtme, dehazy0, tme]
titles = ['Hazy', 'GT', 'ALE', '',
'Dehazy', 'RTME', 'Hazy', 'TME']
row, col = 2, 4
plt.figure(figsize=(4 * col, 4 * row))
for i in range(len(imgs)):
plt.subplot(row, col, i + 1)
plt.title(titles[i])
plt.imshow(imgs[i])
plt.savefig(os.path.join(
vis_path, 'epoch_{}.png'.format(epoch)))
plt.close()
tf.saved_model.save(generator, os.path.join(output_path, 'generator_g/1/'))
if __name__ == "__main__":
input_path = sys.argv[1]
output_path = sys.argv[2]
train_args = readXML(os.path.join(input_path, 'train.xml'))
if not train_args['from_tfrecord']:
image_path = os.path.join(input_path, 'cloud')
label_path = os.path.join(input_path, 'label')
image2tfrecord(image_path, label_path, output_path)
train_loop(train_args, output_path)