File size: 14,210 Bytes
94d1117
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
import logging
import multiprocessing
import os
import time

import torch
import torch.distributed as dist
import torch.multiprocessing as mp
from torch.cuda.amp import GradScaler, autocast
from torch.nn import functional as F
from torch.nn.parallel import DistributedDataParallel as DDP
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter

import modules.commons as commons
import utils
from data_utils import TextAudioCollate, TextAudioSpeakerLoader
from models import (
    MultiPeriodDiscriminator,
    SynthesizerTrn,
)
from modules.losses import discriminator_loss, feature_loss, generator_loss, kl_loss
from modules.mel_processing import mel_spectrogram_torch, spec_to_mel_torch

logging.getLogger('matplotlib').setLevel(logging.WARNING)
logging.getLogger('numba').setLevel(logging.WARNING)

torch.backends.cudnn.benchmark = True
global_step = 0
start_time = time.time()

# os.environ['TORCH_DISTRIBUTED_DEBUG'] = 'INFO'


def main():
    """Assume Single Node Multi GPUs Training Only"""
    assert torch.cuda.is_available(), "CPU training is not allowed."
    hps = utils.get_hparams()

    n_gpus = torch.cuda.device_count()
    os.environ['MASTER_ADDR'] = 'localhost'
    os.environ['MASTER_PORT'] = hps.train.port

    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"))
    
    # for pytorch on win, backend use gloo    
    dist.init_process_group(backend=  'gloo' if os.name == 'nt' else 'nccl', init_method='env://', world_size=n_gpus, rank=rank)
    torch.manual_seed(hps.train.seed)
    torch.cuda.set_device(rank)
    collate_fn = TextAudioCollate()
    all_in_mem = hps.train.all_in_mem   # If you have enough memory, turn on this option to avoid disk IO and speed up training.
    train_dataset = TextAudioSpeakerLoader(hps.data.training_files, hps, all_in_mem=all_in_mem)
    num_workers = 5 if multiprocessing.cpu_count() > 4 else multiprocessing.cpu_count()
    if all_in_mem:
        num_workers = 0
    train_loader = DataLoader(train_dataset, num_workers=num_workers, shuffle=False, pin_memory=True,
                              batch_size=hps.train.batch_size, collate_fn=collate_fn)
    if rank == 0:
        eval_dataset = TextAudioSpeakerLoader(hps.data.validation_files, hps, all_in_mem=all_in_mem,vol_aug = False)
        eval_loader = DataLoader(eval_dataset, num_workers=1, shuffle=False,
                                 batch_size=1, pin_memory=False,
                                 drop_last=False, collate_fn=collate_fn)

    net_g = SynthesizerTrn(
        hps.data.filter_length // 2 + 1,
        hps.train.segment_size // hps.data.hop_length,
        **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])  # , find_unused_parameters=True)
    net_d = DDP(net_d, device_ids=[rank])

    skip_optimizer = False
    try:
        _, _, _, epoch_str = utils.load_checkpoint(utils.latest_checkpoint_path(hps.model_dir, "G_*.pth"), net_g,
                                                   optim_g, skip_optimizer)
        _, _, _, epoch_str = utils.load_checkpoint(utils.latest_checkpoint_path(hps.model_dir, "D_*.pth"), net_d,
                                                   optim_d, skip_optimizer)
        epoch_str = max(epoch_str, 1)
        name=utils.latest_checkpoint_path(hps.model_dir, "D_*.pth")
        global_step=int(name[name.rfind("_")+1:name.rfind(".")])+1
        #global_step = (epoch_str - 1) * len(train_loader)
    except Exception:
        print("load old checkpoint failed...")
        epoch_str = 1
        global_step = 0
    if skip_optimizer:
        epoch_str = 1
        global_step = 0

    warmup_epoch = hps.train.warmup_epochs
    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):
        # set up warm-up learning rate
        if epoch <= warmup_epoch:
            for param_group in optim_g.param_groups:
                param_group['lr'] = hps.train.learning_rate / warmup_epoch * epoch
            for param_group in optim_d.param_groups:
                param_group['lr'] = hps.train.learning_rate / warmup_epoch * epoch
        # training
        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)
        # update learning rate
        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
    
    half_type = torch.bfloat16 if hps.train.half_type=="bf16" else torch.float16

    # train_loader.batch_sampler.set_epoch(epoch)
    global global_step

    net_g.train()
    net_d.train()
    for batch_idx, items in enumerate(train_loader):
        c, f0, spec, y, spk, lengths, uv,volume = items
        g = spk.cuda(rank, non_blocking=True)
        spec, y = spec.cuda(rank, non_blocking=True), y.cuda(rank, non_blocking=True)
        c = c.cuda(rank, non_blocking=True)
        f0 = f0.cuda(rank, non_blocking=True)
        uv = uv.cuda(rank, non_blocking=True)
        lengths = lengths.cuda(rank, non_blocking=True)
        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)
        
        with autocast(enabled=hps.train.fp16_run, dtype=half_type):
            y_hat, ids_slice, z_mask, \
            (z, z_p, m_p, logs_p, m_q, logs_q), pred_lf0, norm_lf0, lf0 = net_g(c, f0, uv, spec, g=g, c_lengths=lengths,
                                                                                spec_lengths=lengths,vol = volume)

            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, dtype=half_type):
                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, dtype=half_type):
            # Generator
            y_d_hat_r, y_d_hat_g, fmap_r, fmap_g = net_d(y, y_hat)
            with autocast(enabled=False, dtype=half_type):
                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_lf0 = F.mse_loss(pred_lf0, lf0) if net_g.module.use_automatic_f0_prediction else 0
                loss_gen_all = loss_gen + loss_fm + loss_mel + loss_kl + loss_lf0
        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_kl]
                reference_loss=0
                for i in losses:
                    reference_loss += i
                logger.info('Train Epoch: {} [{:.0f}%]'.format(
                    epoch,
                    100. * batch_idx / len(train_loader)))
                logger.info(f"Losses: {[x.item() for x in losses]}, step: {global_step}, lr: {lr}, reference_loss: {reference_loss}")

                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/kl": loss_kl,
                                    "loss/g/lf0": loss_lf0})

                # 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())
                }

                if net_g.module.use_automatic_f0_prediction:
                    image_dict.update({
                        "all/lf0": utils.plot_data_to_numpy(lf0[0, 0, :].cpu().numpy(),
                                                              pred_lf0[0, 0, :].detach().cpu().numpy()),
                        "all/norm_lf0": utils.plot_data_to_numpy(lf0[0, 0, :].cpu().numpy(),
                                                                   norm_lf0[0, 0, :].detach().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)))
                keep_ckpts = getattr(hps.train, 'keep_ckpts', 0)
                if keep_ckpts > 0:
                    utils.clean_checkpoints(path_to_models=hps.model_dir, n_ckpts_to_keep=keep_ckpts, sort_by_time=True)

        global_step += 1

    if rank == 0:
        global start_time
        now = time.time()
        durtaion = format(now - start_time, '.2f')
        logger.info(f'====> Epoch: {epoch}, cost {durtaion} s')
        start_time = now


def evaluate(hps, generator, eval_loader, writer_eval):
    generator.eval()
    image_dict = {}
    audio_dict = {}
    with torch.no_grad():
        for batch_idx, items in enumerate(eval_loader):
            c, f0, spec, y, spk, _, uv,volume = items
            g = spk[:1].cuda(0)
            spec, y = spec[:1].cuda(0), y[:1].cuda(0)
            c = c[:1].cuda(0)
            f0 = f0[:1].cuda(0)
            uv= uv[:1].cuda(0)
            if volume is not None:
                volume = volume[:1].cuda(0)
            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,_ = generator.module.infer(c, f0, uv, g=g,vol = volume)

            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
            )

            audio_dict.update({
                f"gen/audio_{batch_idx}": y_hat[0],
                f"gt/audio_{batch_idx}": y[0]
            })
        image_dict.update({
            "gen/mel": utils.plot_spectrogram_to_numpy(y_hat_mel[0].cpu().numpy()),
            "gt/mel": utils.plot_spectrogram_to_numpy(mel[0].cpu().numpy())
        })
    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()