File size: 10,947 Bytes
d1ceaed
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
287
288
289
290
291
292
293
294
295
import os
import json
import argparse
import itertools
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 torch.nn.parallel import DistributedDataParallel as DDP
from torch.cuda.amp import autocast, GradScaler

import commons
import utils
from data_utils import (
  TextAudioSpeakerLoader,
  TextAudioSpeakerCollate,
  DistributedBucketSampler
)
from models import (
  SynthesizerTrn,
  MultiPeriodDiscriminator,
)
from losses import (
  generator_loss,
  discriminator_loss,
  feature_loss,
  kl_loss
)
from mel_processing import mel_spectrogram_torch, spec_to_mel_torch
from text.symbols import symbols


torch.backends.cudnn.benchmark = True
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(run, nprocs=n_gpus, args=(n_gpus, hps,))


def run(rank, n_gpus, hps):
  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 = TextAudioSpeakerLoader(hps.data.training_files, hps.data)
  train_sampler = DistributedBucketSampler(
      train_dataset,
      hps.train.batch_size,
      [32,300,400,500,600,700,800,900,1000],
      num_replicas=n_gpus,
      rank=rank,
      shuffle=True)
  collate_fn = TextAudioSpeakerCollate()
  train_loader = DataLoader(train_dataset, num_workers=8, shuffle=False, pin_memory=True,
      collate_fn=collate_fn, batch_sampler=train_sampler)
  if rank == 0:
    eval_dataset = TextAudioSpeakerLoader(hps.data.validation_files, hps.data)
    eval_loader = DataLoader(eval_dataset, num_workers=8, shuffle=False,
        batch_size=hps.train.batch_size, pin_memory=True,
        drop_last=False, collate_fn=collate_fn)

  net_g = SynthesizerTrn(
      len(symbols),
      hps.data.filter_length // 2 + 1,
      hps.train.segment_size // hps.data.hop_length,
      n_speakers=hps.data.n_speakers,
      **hps.model).cuda(rank)
  net_d = MultiPeriodDiscriminator(hps.model.use_spectral_norm).cuda(rank)
  optim_g = torch.optim.AdamW(
      net_g.parameters(), 
      hps.train.learning_rate, 
      betas=hps.train.betas, 
      eps=hps.train.eps)
  optim_d = torch.optim.AdamW(
      net_d.parameters(),
      hps.train.learning_rate, 
      betas=hps.train.betas, 
      eps=hps.train.eps)
  net_g = DDP(net_g, device_ids=[rank])
  net_d = DDP(net_d, device_ids=[rank])

  try:
    _, _, _, epoch_str = utils.load_checkpoint(utils.latest_checkpoint_path(hps.model_dir, "G_*.pth"), net_g, optim_g)
    _, _, _, epoch_str = utils.load_checkpoint(utils.latest_checkpoint_path(hps.model_dir, "D_*.pth"), net_d, optim_d)
    global_step = (epoch_str - 1) * len(train_loader)
  except:
    epoch_str = 1
    global_step = 0

  scheduler_g = torch.optim.lr_scheduler.ExponentialLR(optim_g, gamma=hps.train.lr_decay, last_epoch=epoch_str-2)
  scheduler_d = torch.optim.lr_scheduler.ExponentialLR(optim_d, gamma=hps.train.lr_decay, last_epoch=epoch_str-2)

  scaler = GradScaler(enabled=hps.train.fp16_run)

  for epoch in range(epoch_str, hps.train.epochs + 1):
    if rank==0:
      train_and_evaluate(rank, epoch, hps, [net_g, net_d], [optim_g, optim_d], [scheduler_g, scheduler_d], scaler, [train_loader, eval_loader], logger, [writer, writer_eval])
    else:
      train_and_evaluate(rank, epoch, hps, [net_g, net_d], [optim_g, optim_d], [scheduler_g, scheduler_d], scaler, [train_loader, None], None, None)
    scheduler_g.step()
    scheduler_d.step()


def train_and_evaluate(rank, epoch, hps, nets, optims, schedulers, scaler, loaders, logger, writers):
  net_g, net_d = nets
  optim_g, optim_d = optims
  scheduler_g, scheduler_d = schedulers
  train_loader, eval_loader = loaders
  if writers is not None:
    writer, writer_eval = writers

  train_loader.batch_sampler.set_epoch(epoch)
  global global_step

  net_g.train()
  net_d.train()
  for batch_idx, (x, x_lengths, spec, spec_lengths, y, y_lengths, speakers) in enumerate(train_loader):
    x, x_lengths = x.cuda(rank, non_blocking=True), x_lengths.cuda(rank, non_blocking=True)
    spec, spec_lengths = spec.cuda(rank, non_blocking=True), spec_lengths.cuda(rank, non_blocking=True)
    y, y_lengths = y.cuda(rank, non_blocking=True), y_lengths.cuda(rank, non_blocking=True)
    speakers = speakers.cuda(rank, non_blocking=True)

    with autocast(enabled=hps.train.fp16_run):
      y_hat, l_length, attn, ids_slice, x_mask, z_mask,\
      (z, z_p, m_p, logs_p, m_q, logs_q) = net_g(x, x_lengths, spec, spec_lengths, speakers)

      mel = spec_to_mel_torch(
          spec, 
          hps.data.filter_length, 
          hps.data.n_mel_channels, 
          hps.data.sampling_rate,
          hps.data.mel_fmin, 
          hps.data.mel_fmax)
      y_mel = commons.slice_segments(mel, ids_slice, hps.train.segment_size // hps.data.hop_length)
      y_hat_mel = mel_spectrogram_torch(
          y_hat.squeeze(1), 
          hps.data.filter_length, 
          hps.data.n_mel_channels, 
          hps.data.sampling_rate, 
          hps.data.hop_length, 
          hps.data.win_length, 
          hps.data.mel_fmin, 
          hps.data.mel_fmax
      )

      y = commons.slice_segments(y, ids_slice * hps.data.hop_length, hps.train.segment_size) # slice 

      # Discriminator
      y_d_hat_r, y_d_hat_g, _, _ = net_d(y, y_hat.detach())
      with autocast(enabled=False):
        loss_disc, losses_disc_r, losses_disc_g = discriminator_loss(y_d_hat_r, y_d_hat_g)
        loss_disc_all = loss_disc
    optim_d.zero_grad()
    scaler.scale(loss_disc_all).backward()
    scaler.unscale_(optim_d)
    grad_norm_d = commons.clip_grad_value_(net_d.parameters(), None)
    scaler.step(optim_d)

    with autocast(enabled=hps.train.fp16_run):
      # Generator
      y_d_hat_r, y_d_hat_g, fmap_r, fmap_g = net_d(y, y_hat)
      with autocast(enabled=False):
        loss_dur = torch.sum(l_length.float())
        loss_mel = F.l1_loss(y_mel, y_hat_mel) * hps.train.c_mel
        loss_kl = kl_loss(z_p, logs_q, m_p, logs_p, z_mask) * hps.train.c_kl

        loss_fm = feature_loss(fmap_r, fmap_g)
        loss_gen, losses_gen = generator_loss(y_d_hat_g)
        loss_gen_all = loss_gen + loss_fm + loss_mel + loss_dur + loss_kl
    optim_g.zero_grad()
    scaler.scale(loss_gen_all).backward()
    scaler.unscale_(optim_g)
    grad_norm_g = commons.clip_grad_value_(net_g.parameters(), None)
    scaler.step(optim_g)
    scaler.update()

    if rank==0:
      if global_step % hps.train.log_interval == 0:
        lr = optim_g.param_groups[0]['lr']
        losses = [loss_disc, loss_gen, loss_fm, loss_mel, loss_dur, loss_kl]
        logger.info('Train Epoch: {} [{:.0f}%]'.format(
          epoch,
          100. * batch_idx / len(train_loader)))
        logger.info([x.item() for x in losses] + [global_step, lr])
        
        scalar_dict = {"loss/g/total": loss_gen_all, "loss/d/total": loss_disc_all, "learning_rate": lr, "grad_norm_d": grad_norm_d, "grad_norm_g": grad_norm_g}
        scalar_dict.update({"loss/g/fm": loss_fm, "loss/g/mel": loss_mel, "loss/g/dur": loss_dur, "loss/g/kl": loss_kl})

        scalar_dict.update({"loss/g/{}".format(i): v for i, v in enumerate(losses_gen)})
        scalar_dict.update({"loss/d_r/{}".format(i): v for i, v in enumerate(losses_disc_r)})
        scalar_dict.update({"loss/d_g/{}".format(i): v for i, v in enumerate(losses_disc_g)})
        image_dict = { 
            "slice/mel_org": utils.plot_spectrogram_to_numpy(y_mel[0].data.cpu().numpy()),
            "slice/mel_gen": utils.plot_spectrogram_to_numpy(y_hat_mel[0].data.cpu().numpy()), 
            "all/mel": utils.plot_spectrogram_to_numpy(mel[0].data.cpu().numpy()),
            "all/attn": utils.plot_alignment_to_numpy(attn[0,0].data.cpu().numpy())
        }
        utils.summarize(
          writer=writer,
          global_step=global_step, 
          images=image_dict,
          scalars=scalar_dict)

      if global_step % hps.train.eval_interval == 0:
        evaluate(hps, net_g, eval_loader, writer_eval)
        utils.save_checkpoint(net_g, optim_g, hps.train.learning_rate, epoch, os.path.join(hps.model_dir, "G_{}.pth".format(global_step)))
        utils.save_checkpoint(net_d, optim_d, hps.train.learning_rate, epoch, os.path.join(hps.model_dir, "D_{}.pth".format(global_step)))
    global_step += 1
  
  if rank == 0:
    logger.info('====> Epoch: {}'.format(epoch))

 
def evaluate(hps, generator, eval_loader, writer_eval):
    generator.eval()
    with torch.no_grad():
      for batch_idx, (x, x_lengths, spec, spec_lengths, y, y_lengths, speakers) in enumerate(eval_loader):
        x, x_lengths = x.cuda(0), x_lengths.cuda(0)
        spec, spec_lengths = spec.cuda(0), spec_lengths.cuda(0)
        y, y_lengths = y.cuda(0), y_lengths.cuda(0)
        speakers = speakers.cuda(0)

        # remove else
        x = x[:1]
        x_lengths = x_lengths[:1]
        spec = spec[:1]
        spec_lengths = spec_lengths[:1]
        y = y[:1]
        y_lengths = y_lengths[:1]
        speakers = speakers[:1]
        break
      y_hat, attn, mask, *_ = generator.module.infer(x, x_lengths, speakers, max_len=1000)
      y_hat_lengths = mask.sum([1,2]).long() * hps.data.hop_length

      mel = spec_to_mel_torch(
        spec, 
        hps.data.filter_length, 
        hps.data.n_mel_channels, 
        hps.data.sampling_rate,
        hps.data.mel_fmin, 
        hps.data.mel_fmax)
      y_hat_mel = mel_spectrogram_torch(
        y_hat.squeeze(1).float(),
        hps.data.filter_length,
        hps.data.n_mel_channels,
        hps.data.sampling_rate,
        hps.data.hop_length,
        hps.data.win_length,
        hps.data.mel_fmin,
        hps.data.mel_fmax
      )
    image_dict = {
      "gen/mel": utils.plot_spectrogram_to_numpy(y_hat_mel[0].cpu().numpy())
    }
    audio_dict = {
      "gen/audio": y_hat[0,:,:y_hat_lengths[0]]
    }
    if global_step == 0:
      image_dict.update({"gt/mel": utils.plot_spectrogram_to_numpy(mel[0].cpu().numpy())})
      audio_dict.update({"gt/audio": y[0,:,:y_lengths[0]]})

    utils.summarize(
      writer=writer_eval,
      global_step=global_step, 
      images=image_dict,
      audios=audio_dict,
      audio_sampling_rate=hps.data.sampling_rate
    )
    generator.train()

                           
if __name__ == "__main__":
  main()