-
Notifications
You must be signed in to change notification settings - Fork 0
/
ae_trainer.py
220 lines (191 loc) · 8.75 KB
/
ae_trainer.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
'''
Author: Emilio Morales (mil.mor.mor@gmail.com)
Jun 2023
'''
import os
import tensorflow as tf
import matplotlib.pyplot as plt
from utils import *
class AutoencoderKL():
def __init__(self, augmenter, ae, discriminator,
ae_opt, d_opt, config):
self.augmenter = augmenter
self.ae = ae
self.discriminator = discriminator
self.ae_opt = ae_opt
self.d_opt = d_opt
self.batch_size = config['batch_size']
self.rec_weight = config['rec_weight']
self.adv_weight = config['adv_weight']
self.kl_weight = config['kl_weight']
self.d_start = config['d_start']
self.n_images = tf.Variable(0, dtype=tf.int64)
# metrics
self.train_metrics = {}
self.test_metrics = {}
self._build_metrics()
# loss
self.loss = tf.keras.losses.MeanAbsoluteError()
def _build_metrics(self):
metric_names = [
'rec_loss',
'kl_loss',
'd_loss',
'ae_total_loss',
'g_loss',
]
for metric_name in metric_names:
self.train_metrics[metric_name] = tf.keras.metrics.Mean()
self.test_metrics[metric_name] = tf.keras.metrics.Mean()
def denormalize(self, images):
# convert the pixel values back to 0-1 range
return tf.clip_by_value(images, 0.0, 1.0)
def discriminator_loss(self, real_img, fake_img):
real_loss = tf.reduce_mean(tf.nn.relu(1.0 - real_img))
fake_loss = tf.reduce_mean(tf.nn.relu(1.0 + fake_img))
return 0.5 * (fake_loss + real_loss)
def generator_loss(self, fake_img):
return -tf.reduce_mean(fake_img)
@tf.function
def train_step(self, real_img):
real_img = self.augmenter(real_img, training=True)
disc_factor = 0.0
if self.n_images > self.d_start:
disc_factor = 1.0
with tf.GradientTape() as ae_tape, tf.GradientTape() as disc_tape:
rec_img, z_mean, z_log_var = self.ae(real_img, training=True)
kl_loss = (-0.5 * (1 + z_log_var - tf.square(z_mean) - tf.exp(z_log_var)))
kl_loss = tf.reduce_mean(tf.reduce_sum(kl_loss, axis=[1, 2, 3])) * self.kl_weight
fake_logits = self.discriminator(rec_img, training=True)[0]
real_logits = self.discriminator(real_img, training=True)[0]
d_loss = self.discriminator_loss(real_logits, fake_logits) * disc_factor
g_loss = self.generator_loss(fake_logits) * self.adv_weight * disc_factor
rec_loss = self.loss(rec_img, real_img) * self.rec_weight
ae_total_loss = rec_loss + kl_loss + g_loss
ae_grad = ae_tape.gradient(ae_total_loss, self.ae.trainable_weights)
self.ae_opt.apply_gradients(zip(ae_grad, self.ae.trainable_weights))
d_grad = disc_tape.gradient(d_loss, self.discriminator.trainable_variables)
self.d_opt.apply_gradients(zip(d_grad, self.discriminator.trainable_variables))
update_metrics(
self.train_metrics,
rec_loss=rec_loss,
kl_loss=kl_loss,
d_loss=d_loss,
ae_total_loss=ae_total_loss,
g_loss=g_loss,
)
@tf.function
def test_step(self, real_img):
disc_factor = 0.0
if self.n_images > self.d_start:
disc_factor = 1.0
rec_img, z_mean, z_log_var = self.ae(real_img, training=False)
kl_loss = (-0.5 * (1 + z_log_var - tf.square(z_mean) - tf.exp(z_log_var)))
kl_loss = tf.reduce_mean(tf.reduce_sum(kl_loss, axis=[1, 2, 3])) * self.kl_weight
fake_logits = self.discriminator(rec_img, training=False)[0]
real_logits = self.discriminator(real_img, training=False)[0]
d_loss = self.discriminator_loss(real_logits, fake_logits) * disc_factor
g_loss = self.generator_loss(fake_logits) * self.adv_weight * disc_factor
rec_loss = self.loss(rec_img, real_img) * self.rec_weight
ae_total_loss = rec_loss + kl_loss + g_loss
update_metrics(
self.test_metrics,
rec_loss=rec_loss,
kl_loss=kl_loss,
d_loss=d_loss,
ae_total_loss=ae_total_loss,
g_loss=g_loss,
)
def plot_images(self, img, step=None, num_rows=5, num_cols=5, plot_image_size=128,
is_jupyter=False):
img_dir = os.path.join(self.model_dir, 'ae-log-img')
os.makedirs(img_dir, exist_ok=True)
generated_images, _, _ = self.ae(img[:num_rows*num_cols], training=False)
generated_images = self.denormalize(generated_images)
# organize generated images into a grid
generated_images = tf.image.resize(
generated_images, (plot_image_size, plot_image_size), method="nearest"
)
generated_images = tf.reshape(
generated_images,
(num_rows, num_cols, plot_image_size, plot_image_size, 3),
)
generated_images = tf.transpose(generated_images, (0, 2, 1, 3, 4))
generated_images = tf.reshape(
generated_images,
(num_rows * plot_image_size, num_cols * plot_image_size, 3),
)
if is_jupyter:
plt.figure(figsize=(num_cols * 1.5, num_rows * 1.5))
plt.imshow(generated_images.numpy())
plt.axis("off")
plt.tight_layout()
plt.show()
plt.close()
else:
plt.imsave(os.path.join(
img_dir, f'{step}.png'), generated_images.numpy()
)
def create_ckpt(self, model_dir, max_ckpt_to_keep, restore_best=True):
# log dir
self.model_dir = model_dir
ae_log_dir = os.path.join(self.model_dir, 'ae-log-dir')
self.writer = tf.summary.create_file_writer(ae_log_dir)
# checkpoint dir
checkpoint_dir = os.path.join(model_dir, 'ae-ckpt')
best_checkpoint_dir = os.path.join(model_dir, 'ae-best-ckpt')
self.ckpt = tf.train.Checkpoint(
ae=self.ae, ae_opt=self.ae_opt, discriminator=self.discriminator,
d_opt=self.d_opt, n_images=self.n_images,
best_loss=tf.Variable(10000.0) # initialize with big value
)
self.ckpt_manager = tf.train.CheckpointManager(
self.ckpt, directory=checkpoint_dir, max_to_keep=max_ckpt_to_keep
)
self.best_ckpt_manager = tf.train.CheckpointManager(
self.ckpt, directory=best_checkpoint_dir, max_to_keep=max_ckpt_to_keep
)
if restore_best == True and self.best_ckpt_manager.latest_checkpoint:
last_ckpt = self.best_ckpt_manager.latest_checkpoint
self.ckpt.restore(last_ckpt)
print(f'Best checkpoint restored from {last_ckpt}')
elif restore_best == False and self.ckpt_manager.latest_checkpoint:
last_ckpt = self.ckpt_manager.latest_checkpoint
self.ckpt.restore(last_ckpt)
print(f'Checkpoint restored from {last_ckpt}')
else:
print(f'Checkpoint created at {self.model_dir} dir')
def save_ckpt(self, n_images, verbose=1, reset_states=True):
# tensorboard
with self.writer.as_default():
print('Train metrics: ')
for name, metric in self.train_metrics.items():
print(f'train_{name}: {metric.result():.4f}', end=" - ")
tf.summary.scalar(f'train_{name}', metric.result(), step=n_images)
print('\nVal metrics: ')
for name, metric in self.test_metrics.items():
print(f'val_{name}: {metric.result():.4f}', end=" - ")
tf.summary.scalar(f'val_{name}', metric.result(), step=n_images)
print(f'n_images: {n_images}')
self.n_images.assign(n_images)
# checkpoint
if self.test_metrics['rec_loss'].result() < self.ckpt.best_loss:
self.ckpt.best_loss.assign(self.test_metrics['rec_loss'].result())
self.ckpt_manager.save(n_images)
self.best_ckpt_manager.save(n_images)
print(f'Best checkpoint saved at {n_images} images')
else:
self.ckpt_manager.save(n_images)
print(f'Checkpoint saved at {n_images} images')
# reset metrics
reset_metrics(self.train_metrics)
reset_metrics(self.test_metrics)
def restore_ae(self, model_dir, max_ckpt_to_keep=1):
best_checkpoint_dir = os.path.join(model_dir, 'ae-best-ckpt')
ckpt = tf.train.Checkpoint(ae=self.ae)
ckpt_manager = tf.train.CheckpointManager(
ckpt, directory=best_checkpoint_dir, max_to_keep=max_ckpt_to_keep
)
last_ckpt = ckpt_manager.latest_checkpoint
ckpt.restore(last_ckpt).expect_partial()
print(f'AutoencoderKL checkpoint restored from {last_ckpt}')