-
Notifications
You must be signed in to change notification settings - Fork 49
/
train.py
243 lines (188 loc) · 8.64 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
import tensorflow as tf
import tensorboard
import numpy as np
from tqdm import tqdm
import time
import datetime
import os
import subprocess
from utils import Utils_functions
from models import Models_functions
from losses import *
class Train_functions:
def __init__(self, args):
self.args = args
self.U = Utils_functions(args)
self.M = Models_functions(args)
def gradient_penalty(self, x, net):
x_hat = x
with tf.GradientTape() as t:
t.watch(x_hat)
d_hat, _ = net(x_hat, training=True)
gradients = t.gradient(d_hat, x_hat)
ddx = tf.sqrt(1e-6 + tf.reduce_sum(gradients**2, axis=[1, 2, 3]))
d_regularizer = tf.reduce_mean((ddx - 1.0) ** 2)
return d_regularizer
def train_all(self, a, ema, g_train=True, disc_train=True, models_ls=None):
critic, gen, enc, dec, enc2, dec2, gen_ema, [opt_dec, opt_disc], switch = models_ls
a = tf.expand_dims(a, -3)
a = self.U.rand_channel_swap(a)
noiseg = tf.random.normal([self.args.bs, self.args.coorddepth], dtype=tf.float32)
noisel = tf.concat([tf.random.normal([self.args.bs, self.args.coorddepth], dtype=tf.float32), noiseg], -1)
noisec = tf.concat([tf.random.normal([self.args.bs, self.args.coorddepth], dtype=tf.float32), noiseg], -1)
noiser = tf.concat([tf.random.normal([self.args.bs, self.args.coorddepth], dtype=tf.float32), noiseg], -1)
rl = tf.linspace(noisel, noisec, self.args.coordlen + 1, axis=-2)[:, :-1, :]
rr = tf.linspace(noisec, noiser, self.args.coordlen + 1, axis=-2)
noisetot = tf.concat([rl, rr], -2)
noisetot = self.U.center_coordinate(noisetot)
noise = self.U.crop_coordinate(noisetot)
with tf.GradientTape() as tape_gen, tf.GradientTape() as tape_disc, tf.GradientTape() as tape_gp:
if not disc_train:
tape_disc.stop_recording()
if not g_train:
tape_gen.stop_recording()
tape_gp.watch(a)
ab = gen(noise, training=True)
loss_dtr = 0.0
loss_dtf = 0.0
loss_gt = 0.0
loss_did = 0.0
loss_gp = 0.0
if disc_train or g_train:
ca = critic(a, training=True)
cab = critic(ab, training=True)
switch.assign(self.U.update_switch(switch, ca, cab))
grad_gp = tape_gp.gradient(tf.reduce_sum(ca), [a])[0]
loss_gp = tf.reduce_mean(tf.reduce_sum(tf.reshape(grad_gp**2, [tf.shape(grad_gp)[0], -1]), -1))
if disc_train:
loss_dtr = d_loss_r(ca)
loss_dtf = d_loss_f(cab)
loss_dt = (loss_dtr + loss_dtf) / 2.0
loss_d = loss_dt + self.args.gp_max_weight * (-switch) * loss_gp
if self.args.mixed_precision:
loss_d = opt_disc.get_scaled_loss(loss_d)
if g_train:
loss_gt = g_loss_f(cab)
loss_gen = loss_gt
if self.args.mixed_precision:
loss_gen = opt_dec.get_scaled_loss(loss_gen)
if disc_train:
grad_disc = tape_disc.gradient(loss_d, critic.trainable_weights)
if self.args.mixed_precision:
grad_disc = opt_disc.get_unscaled_gradients(grad_disc)
opt_disc.apply_gradients(zip(grad_disc, critic.trainable_weights))
if g_train:
grad_dec = tape_gen.gradient(loss_gen, gen.trainable_variables)
if self.args.mixed_precision:
grad_dec = opt_dec.get_unscaled_gradients(grad_dec)
opt_dec.apply_gradients(zip(grad_dec, gen.trainable_variables))
ema.apply(gen.trainable_variables)
return loss_dtr, loss_dtf, loss_gp, loss_gt
# @tf.function(jit_compile=True)
# def train_tot(self, a, ema, models_ls=None):
# return self.train_all(a, ema, g_train=True, disc_train=True, models_ls=models_ls)
def update_lr(self, lr, opts=None):
opt_dec, opt_disc = opts
opt_dec.learning_rate = lr
opt_disc.learning_rate = lr * 1.0
def train(self, ds, models_ls=None):
@tf.function(jit_compile=self.args.xla)
def train_tot(a, ema, models_ls=None):
return self.train_all(a, ema, g_train=True, disc_train=True, models_ls=models_ls)
current_time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
train_log_dir = (
f"{self.args.log_path}/MUSIKA_latlen_{self.args.latlen}_latdepth_{self.args.latdepth}_sr_{self.args.sr}/"
+ current_time
+ "/train"
)
train_summary_writer = tf.summary.create_file_writer(train_log_dir)
exp_path = f"{self.args.save_path}/MUSIKA_latlen_{self.args.latlen}_latdepth_{self.args.latdepth}_sr_{self.args.sr}_time_{current_time}"
os.makedirs(exp_path, exist_ok=True)
print("--------------------------------")
print("--------------------------------")
print("--------------------------------")
print("--------------------------------")
print("--------------------------------")
_ = subprocess.Popen(
[
"tensorboard",
"--logdir",
f"{self.args.log_path}/MUSIKA_latlen_{self.args.latlen}_latdepth_{self.args.latdepth}_sr_{self.args.sr}",
"--port",
"6006",
]
)
print("CLICK ON LINK BELOW TO OPEN TENSORBOARD INTERFACE")
print("http://localhost:6006/")
print("--------------------------------")
print("--------------------------------")
print("--------------------------------")
print("--------------------------------")
print("--------------------------------")
ema = tf.train.ExponentialMovingAverage(decay=0.999)
critic, gen, enc, dec, enc2, dec2, gen_ema, [opt_dec, opt_disc], switch = models_ls
ema.apply(gen_ema.trainable_variables)
self.update_lr(self.args.lr, [opt_dec, opt_disc])
c = 0
g = 0
m = 0
idloss = 0.0
print("Preparing for Training (this can take one or two minutes)...")
for epoch in range(self.args.epochs):
bef = time.time()
bef_loop = time.time()
dtr_list = []
dtf_list = []
did_list = []
gt_list = []
id_list = []
pbar = tqdm(
ds,
desc=f"Epoch {epoch}/{self.args.epochs}",
position=0,
leave=True,
total=self.args.totsamples // self.args.bs,
)
for batchi, (wv) in enumerate(pbar):
a = wv
dloss_tr, dloss_tf, dloss_id, gloss_t = train_tot(a, ema, models_ls=models_ls)
with train_summary_writer.as_default():
tf.summary.scalar("disc_loss_r", dloss_tr, step=m)
tf.summary.scalar("disc_loss_f", dloss_tf, step=m)
tf.summary.scalar("gen_loss", gloss_t, step=m)
tf.summary.scalar("gradient_penalty", dloss_id, step=m)
tf.summary.scalar("gp_weight", -switch.value() * self.args.gp_max_weight, step=m)
tf.summary.scalar("lr", self.args.lr, step=m)
dtr_list.append(dloss_tr)
dtf_list.append(dloss_tf)
did_list.append(dloss_id)
gt_list.append(gloss_t)
c += 1
g += 1
m += 1
if batchi % 20 == 0:
pbar.set_postfix(
{
"DR": np.mean(dtr_list[-g:], axis=0),
"DF": np.mean(dtf_list[-g:], axis=0),
"G": np.mean(gt_list[-g:], axis=0),
"GP": np.mean(did_list[-g:], axis=0),
"LR": self.args.lr,
"TIME": (time.time() - bef_loop) / 20,
}
)
bef_loop = time.time()
nbatch = batchi
for var, var_ema in zip(gen.trainable_variables, gen_ema.trainable_variables):
var_ema.assign(ema.average(var))
self.U.save_end(
epoch,
np.mean(gt_list[-self.args.save_every * c :], axis=0),
np.mean(dtr_list[-self.args.save_every * c :], axis=0),
np.mean(dtf_list[-self.args.save_every * c :], axis=0),
n_save=self.args.save_every,
models_ls=models_ls,
save_path=exp_path,
)
c = 0
g = 0