Spaces:
Build error
Build error
File size: 11,975 Bytes
4817bcc |
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 |
import warnings
warnings.simplefilter(action='ignore', category=FutureWarning)
import itertools
import os
import time
import argparse
import json
import torch
import torch.nn.functional as F
from torch.utils.tensorboard import SummaryWriter
from torch.utils.data import DistributedSampler, DataLoader
import torch.multiprocessing as mp
from torch.distributed import init_process_group
from torch.nn.parallel import DistributedDataParallel
from vocoder.hifigan.meldataset import MelDataset, mel_spectrogram, get_dataset_filelist
from vocoder.hifigan.models import Generator, MultiPeriodDiscriminator, MultiScaleDiscriminator, feature_loss, generator_loss,\
discriminator_loss
from vocoder.hifigan.utils import plot_spectrogram, scan_checkpoint, load_checkpoint, save_checkpoint
torch.backends.cudnn.benchmark = True
def train(rank, a, h):
a.checkpoint_path = a.models_dir.joinpath(a.run_id+'_hifigan')
a.checkpoint_path.mkdir(exist_ok=True)
a.training_epochs = 3100
a.stdout_interval = 5
a.checkpoint_interval = a.backup_every
a.summary_interval = 5000
a.validation_interval = 1000
a.fine_tuning = True
a.input_wavs_dir = a.syn_dir.joinpath("audio")
a.input_mels_dir = a.syn_dir.joinpath("mels")
if h.num_gpus > 1:
init_process_group(backend=h.dist_config['dist_backend'], init_method=h.dist_config['dist_url'],
world_size=h.dist_config['world_size'] * h.num_gpus, rank=rank)
torch.cuda.manual_seed(h.seed)
device = torch.device('cuda:{:d}'.format(rank))
generator = Generator(h).to(device)
mpd = MultiPeriodDiscriminator().to(device)
msd = MultiScaleDiscriminator().to(device)
if rank == 0:
print(generator)
os.makedirs(a.checkpoint_path, exist_ok=True)
print("checkpoints directory : ", a.checkpoint_path)
if os.path.isdir(a.checkpoint_path):
cp_g = scan_checkpoint(a.checkpoint_path, 'g_hifigan_')
cp_do = scan_checkpoint(a.checkpoint_path, 'do_hifigan_')
steps = 0
if cp_g is None or cp_do is None:
state_dict_do = None
last_epoch = -1
else:
state_dict_g = load_checkpoint(cp_g, device)
state_dict_do = load_checkpoint(cp_do, device)
generator.load_state_dict(state_dict_g['generator'])
mpd.load_state_dict(state_dict_do['mpd'])
msd.load_state_dict(state_dict_do['msd'])
steps = state_dict_do['steps'] + 1
last_epoch = state_dict_do['epoch']
if h.num_gpus > 1:
generator = DistributedDataParallel(generator, device_ids=[rank]).to(device)
mpd = DistributedDataParallel(mpd, device_ids=[rank]).to(device)
msd = DistributedDataParallel(msd, device_ids=[rank]).to(device)
optim_g = torch.optim.AdamW(generator.parameters(), h.learning_rate, betas=[h.adam_b1, h.adam_b2])
optim_d = torch.optim.AdamW(itertools.chain(msd.parameters(), mpd.parameters()),
h.learning_rate, betas=[h.adam_b1, h.adam_b2])
if state_dict_do is not None:
optim_g.load_state_dict(state_dict_do['optim_g'])
optim_d.load_state_dict(state_dict_do['optim_d'])
scheduler_g = torch.optim.lr_scheduler.ExponentialLR(optim_g, gamma=h.lr_decay, last_epoch=last_epoch)
scheduler_d = torch.optim.lr_scheduler.ExponentialLR(optim_d, gamma=h.lr_decay, last_epoch=last_epoch)
training_filelist, validation_filelist = get_dataset_filelist(a)
# print(training_filelist)
# exit()
trainset = MelDataset(training_filelist, h.segment_size, h.n_fft, h.num_mels,
h.hop_size, h.win_size, h.sampling_rate, h.fmin, h.fmax, n_cache_reuse=0,
shuffle=False if h.num_gpus > 1 else True, fmax_loss=h.fmax_for_loss, device=device,
fine_tuning=a.fine_tuning, base_mels_path=a.input_mels_dir)
train_sampler = DistributedSampler(trainset) if h.num_gpus > 1 else None
train_loader = DataLoader(trainset, num_workers=h.num_workers, shuffle=False,
sampler=train_sampler,
batch_size=h.batch_size,
pin_memory=True,
drop_last=True)
if rank == 0:
validset = MelDataset(validation_filelist, h.segment_size, h.n_fft, h.num_mels,
h.hop_size, h.win_size, h.sampling_rate, h.fmin, h.fmax, False, False, n_cache_reuse=0,
fmax_loss=h.fmax_for_loss, device=device, fine_tuning=a.fine_tuning,
base_mels_path=a.input_mels_dir)
validation_loader = DataLoader(validset, num_workers=1, shuffle=False,
sampler=None,
batch_size=1,
pin_memory=True,
drop_last=True)
sw = SummaryWriter(os.path.join(a.checkpoint_path, 'logs'))
generator.train()
mpd.train()
msd.train()
for epoch in range(max(0, last_epoch), a.training_epochs):
if rank == 0:
start = time.time()
print("Epoch: {}".format(epoch+1))
if h.num_gpus > 1:
train_sampler.set_epoch(epoch)
for i, batch in enumerate(train_loader):
if rank == 0:
start_b = time.time()
x, y, _, y_mel = batch
x = torch.autograd.Variable(x.to(device, non_blocking=True))
y = torch.autograd.Variable(y.to(device, non_blocking=True))
y_mel = torch.autograd.Variable(y_mel.to(device, non_blocking=True))
y = y.unsqueeze(1)
y_g_hat = generator(x)
y_g_hat_mel = mel_spectrogram(y_g_hat.squeeze(1), h.n_fft, h.num_mels, h.sampling_rate, h.hop_size, h.win_size,
h.fmin, h.fmax_for_loss)
if steps > h.disc_start_step:
optim_d.zero_grad()
# MPD
y_df_hat_r, y_df_hat_g, _, _ = mpd(y, y_g_hat.detach())
loss_disc_f, losses_disc_f_r, losses_disc_f_g = discriminator_loss(y_df_hat_r, y_df_hat_g)
# MSD
y_ds_hat_r, y_ds_hat_g, _, _ = msd(y, y_g_hat.detach())
loss_disc_s, losses_disc_s_r, losses_disc_s_g = discriminator_loss(y_ds_hat_r, y_ds_hat_g)
loss_disc_all = loss_disc_s + loss_disc_f
loss_disc_all.backward()
optim_d.step()
# Generator
optim_g.zero_grad()
# L1 Mel-Spectrogram Loss
loss_mel = F.l1_loss(y_mel, y_g_hat_mel) * 45
if steps > h.disc_start_step:
y_df_hat_r, y_df_hat_g, fmap_f_r, fmap_f_g = mpd(y, y_g_hat)
y_ds_hat_r, y_ds_hat_g, fmap_s_r, fmap_s_g = msd(y, y_g_hat)
loss_fm_f = feature_loss(fmap_f_r, fmap_f_g)
loss_fm_s = feature_loss(fmap_s_r, fmap_s_g)
loss_gen_f, losses_gen_f = generator_loss(y_df_hat_g)
loss_gen_s, losses_gen_s = generator_loss(y_ds_hat_g)
loss_gen_all = loss_gen_s + loss_gen_f + loss_fm_s + loss_fm_f + loss_mel
else:
loss_gen_all = loss_mel
loss_gen_all.backward()
optim_g.step()
if rank == 0:
# STDOUT logging
if steps % a.stdout_interval == 0:
with torch.no_grad():
mel_error = F.l1_loss(y_mel, y_g_hat_mel).item()
print('Steps : {:d}, Gen Loss Total : {:4.3f}, Mel-Spec. Error : {:4.3f}, s/b : {:4.3f}'.
format(steps, loss_gen_all, mel_error, time.time() - start_b))
# checkpointing
if steps % a.checkpoint_interval == 0 and steps != 0:
checkpoint_path = "{}/g_hifigan_{:08d}.pt".format(a.checkpoint_path, steps)
save_checkpoint(checkpoint_path,
{'generator': (generator.module if h.num_gpus > 1 else generator).state_dict()})
checkpoint_path = "{}/do_hifigan_{:08d}.pt".format(a.checkpoint_path, steps)
save_checkpoint(checkpoint_path,
{'mpd': (mpd.module if h.num_gpus > 1 else mpd).state_dict(),
'msd': (msd.module if h.num_gpus > 1 else msd).state_dict(),
'optim_g': optim_g.state_dict(), 'optim_d': optim_d.state_dict(), 'steps': steps,
'epoch': epoch})
# Tensorboard summary logging
if steps % a.summary_interval == 0:
sw.add_scalar("training/gen_loss_total", loss_gen_all, steps)
sw.add_scalar("training/mel_spec_error", mel_error, steps)
# save temperate hifigan model
if steps % a.save_every == 0:
checkpoint_path = "{}/g_hifigan.pt".format(a.checkpoint_path)
save_checkpoint(checkpoint_path,
{'generator': (generator.module if h.num_gpus > 1 else generator).state_dict()})
checkpoint_path = "{}/do_hifigan.pt".format(a.checkpoint_path)
save_checkpoint(checkpoint_path,
{'mpd': (mpd.module if h.num_gpus > 1 else mpd).state_dict(),
'msd': (msd.module if h.num_gpus > 1 else msd).state_dict(),
'optim_g': optim_g.state_dict(), 'optim_d': optim_d.state_dict(), 'steps': steps,
'epoch': epoch})
# Validation
if steps % a.validation_interval == 0: # and steps != 0:
generator.eval()
torch.cuda.empty_cache()
val_err_tot = 0
with torch.no_grad():
for j, batch in enumerate(validation_loader):
x, y, _, y_mel = batch
y_g_hat = generator(x.to(device))
y_mel = torch.autograd.Variable(y_mel.to(device, non_blocking=True))
y_g_hat_mel = mel_spectrogram(y_g_hat.squeeze(1), h.n_fft, h.num_mels, h.sampling_rate,
h.hop_size, h.win_size,
h.fmin, h.fmax_for_loss)
# val_err_tot += F.l1_loss(y_mel, y_g_hat_mel).item()
if j <= 4:
if steps == 0:
sw.add_audio('gt/y_{}'.format(j), y[0], steps, h.sampling_rate)
sw.add_figure('gt/y_spec_{}'.format(j), plot_spectrogram(x[0]), steps)
sw.add_audio('generated/y_hat_{}'.format(j), y_g_hat[0], steps, h.sampling_rate)
y_hat_spec = mel_spectrogram(y_g_hat.squeeze(1), h.n_fft, h.num_mels,
h.sampling_rate, h.hop_size, h.win_size,
h.fmin, h.fmax)
sw.add_figure('generated/y_hat_spec_{}'.format(j),
plot_spectrogram(y_hat_spec.squeeze(0).cpu().numpy()), steps)
val_err = val_err_tot / (j+1)
sw.add_scalar("validation/mel_spec_error", val_err, steps)
generator.train()
steps += 1
scheduler_g.step()
scheduler_d.step()
if rank == 0:
print('Time taken for epoch {} is {} sec\n'.format(epoch + 1, int(time.time() - start)))
|