forked from jaywalnut310/glow-tts
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain_custom_multi.py
202 lines (166 loc) · 8.72 KB
/
train_custom_multi.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
import os
import json
import argparse
import math
import torch
from torch import nn, optim
from torch.nn import functional as F
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter
import torch.multiprocessing as mp
import torch.distributed as dist
from apex.parallel import DistributedDataParallel as DDP
from apex import amp
from data_utils_custom import TextMelSpeakerLoader, TextMelSpeakerCollate
import models
import commons
import utils
from korean_text.symbols import kor_symbols
from setproctitle import setproctitle
global_step = 0
def main():
"""Assume Single Node Multi GPUs Training Only"""
assert torch.cuda.is_available(), "CPU training is not allowed."
n_gpus = torch.cuda.device_count()
os.environ['MASTER_ADDR'] = 'localhost'
os.environ['MASTER_PORT'] = '80000'
hps = utils.get_hparams()
mp.spawn(train_and_eval, nprocs=n_gpus, args=(n_gpus, hps,))
def train_and_eval(rank, n_gpus, hps):
setproctitle("Glow-kor-multi-b{}".format(hps.train.batch_size))
global global_step
if rank == 0:
logger = utils.get_logger(hps.model_dir)
logger.info(hps)
utils.check_git_hash(hps.model_dir)
writer = SummaryWriter(log_dir=hps.model_dir)
writer_eval = SummaryWriter(log_dir=os.path.join(hps.model_dir, "eval"))
dist.init_process_group(backend='nccl', init_method='env://', world_size=n_gpus, rank=rank)
torch.manual_seed(hps.train.seed)
torch.cuda.set_device(rank)
train_dataset = TextMelSpeakerLoader(hps.data.training_files, hps.data)
train_sampler = torch.utils.data.distributed.DistributedSampler(
train_dataset,
num_replicas=n_gpus,
rank=rank,
shuffle=True)
collate_fn = TextMelSpeakerCollate(1)
train_loader = DataLoader(train_dataset, num_workers=8, shuffle=False,
batch_size=hps.train.batch_size, pin_memory=True,
drop_last=True, collate_fn=collate_fn, sampler=train_sampler)
if rank == 0:
val_dataset = TextMelSpeakerLoader(hps.data.validation_files, hps.data)
val_loader = DataLoader(val_dataset, num_workers=8, shuffle=False,
batch_size=hps.train.batch_size, pin_memory=True,
drop_last=True, collate_fn=collate_fn)
generator = models.FlowGenerator(
n_vocab=len(kor_symbols) + getattr(hps.data, "add_blank", False),
out_channels=hps.data.n_mel_channels,
**hps.model).cuda(rank)
# n_speakers=hps.model.n_speakers, gin_channels=hps.model.gin_channels, # TypeError: type object got multiple values for keyword argument 'n_speakers'
optimizer_g = commons.Adam(generator.parameters(), scheduler=hps.train.scheduler, dim_model=hps.model.hidden_channels, warmup_steps=hps.train.warmup_steps, lr=hps.train.learning_rate, betas=hps.train.betas, eps=hps.train.eps)
if hps.train.fp16_run:
generator, optimizer_g._optim = amp.initialize(generator, optimizer_g._optim, opt_level="O1")
generator = DDP(generator)
epoch_str = 1
global_step = 0
try:
_, _, _, epoch_str = utils.load_checkpoint(utils.latest_checkpoint_path(hps.model_dir, "G_*.pth"), generator, optimizer_g)
epoch_str += 1
optimizer_g.step_num = (epoch_str - 1) * len(train_loader)
optimizer_g._update_learning_rate()
global_step = (epoch_str - 1) * len(train_loader)
except:
if hps.train.ddi and os.path.isfile(os.path.join(hps.model_dir, "ddi_G.pth")):
_ = utils.load_checkpoint(os.path.join(hps.model_dir, "ddi_G.pth"), generator, optimizer_g)
for epoch in range(epoch_str, hps.train.epochs + 1):
if rank == 0:
train(rank, epoch, hps, generator, optimizer_g, train_loader, logger, writer)
evaluate(rank, epoch, hps, generator, optimizer_g, val_loader, logger, writer_eval)
if epoch % 10 == 0: # nayoun edited
utils.save_checkpoint(generator, optimizer_g, hps.train.learning_rate, epoch, os.path.join(hps.model_dir, "G_{}.pth".format(epoch)))
else:
train(rank, epoch, hps, generator, optimizer_g, train_loader, None, None)
def train(rank, epoch, hps, generator, optimizer_g, train_loader, logger, writer):
train_loader.sampler.set_epoch(epoch)
global global_step
generator.train()
for batch_idx, (x, x_lengths, y, y_lengths, g) in enumerate(train_loader):
x, x_lengths = x.cuda(rank, non_blocking=True), x_lengths.cuda(rank, non_blocking=True)
y, y_lengths = y.cuda(rank, non_blocking=True), y_lengths.cuda(rank, non_blocking=True)
g = g.cuda(rank, non_blocking=True)
# Train Generator
optimizer_g.zero_grad()
(z, z_m, z_logs, logdet, z_mask), (x_m, x_logs, x_mask), (attn, logw, logw_) = \
generator(x = x, x_lengths = x_lengths, y = y, y_lengths = y_lengths, g = g, gen=False)
l_mle = commons.mle_loss(z, z_m, z_logs, logdet, z_mask)
l_length = commons.duration_loss(logw, logw_, x_lengths)
loss_gs = [l_mle, l_length]
loss_g = sum(loss_gs)
if hps.train.fp16_run:
with amp.scale_loss(loss_g, optimizer_g._optim) as scaled_loss:
scaled_loss.backward()
grad_norm = commons.clip_grad_value_(amp.master_params(optimizer_g._optim), 5)
else:
loss_g.backward()
grad_norm = commons.clip_grad_value_(generator.parameters(), 5)
optimizer_g.step()
if rank==0:
if batch_idx % hps.train.log_interval == 0:
(y_gen, *_), *_ = generator.module(
x = x[:1], x_lengths = x_lengths[:1], g = g[:1], gen=True)
logger.info('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
epoch, batch_idx * len(x), len(train_loader.dataset),
100. * batch_idx / len(train_loader),
loss_g.item()))
logger.info([x.item() for x in loss_gs] + [global_step, optimizer_g.get_lr()])
scalar_dict = {"loss/g/total": loss_g, "learning_rate": optimizer_g.get_lr(), "grad_norm": grad_norm}
scalar_dict.update({"loss/g/{}".format(i): v for i, v in enumerate(loss_gs)})
utils.summarize(
writer=writer,
global_step=global_step,
images={"y_org": utils.plot_spectrogram_to_numpy(y[0].data.cpu().numpy()),
"y_gen": utils.plot_spectrogram_to_numpy(y_gen[0].data.cpu().numpy()),
"attn": utils.plot_alignment_to_numpy(attn[0,0].data.cpu().numpy()),
},
scalars=scalar_dict)
global_step += 1
if rank == 0:
logger.info('====> Epoch: {}'.format(epoch))
def evaluate(rank, epoch, hps, generator, optimizer_g, val_loader, logger, writer_eval):
if rank == 0:
global global_step
generator.eval()
losses_tot = []
with torch.no_grad():
for batch_idx, (x, x_lengths, y, y_lengths, g) in enumerate(val_loader):
x, x_lengths = x.cuda(rank, non_blocking=True), x_lengths.cuda(rank, non_blocking=True)
y, y_lengths = y.cuda(rank, non_blocking=True), y_lengths.cuda(rank, non_blocking=True)
g = g.cuda(rank, non_blocking=True)
(z, z_m, z_logs, logdet, z_mask), (x_m, x_logs, x_mask), (attn, logw, logw_) = \
generator(x = x, x_lengths = x_lengths, y = y, y_lengths = y_lengths, g = g, gen=False)
l_mle = commons.mle_loss(z, z_m, z_logs, logdet, z_mask)
l_length = commons.duration_loss(logw, logw_, x_lengths)
loss_gs = [l_mle, l_length]
loss_g = sum(loss_gs)
if batch_idx == 0:
losses_tot = loss_gs
else:
losses_tot = [x + y for (x, y) in zip(losses_tot, loss_gs)]
if batch_idx % hps.train.log_interval == 0:
logger.info('Eval Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
epoch, batch_idx * len(x), len(val_loader.dataset),
100. * batch_idx / len(val_loader),
loss_g.item()))
logger.info([x.item() for x in loss_gs])
losses_tot = [x / len(val_loader) for x in losses_tot]
loss_tot = sum(losses_tot)
scalar_dict = {"loss/g/total": loss_tot}
scalar_dict.update({"loss/g/{}".format(i): v for i, v in enumerate(losses_tot)})
utils.summarize(
writer=writer_eval,
global_step=global_step,
scalars=scalar_dict)
logger.info('====> Epoch: {}'.format(epoch))
if __name__ == "__main__":
main()