Upload 3 files
Browse files- train_fix/bert_genV0.py +53 -0
- train_fix/text/__init__.py +29 -0
- train_fix/train_ms.py +607 -0
train_fix/bert_genV0.py
ADDED
@@ -0,0 +1,53 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from torch.utils.data import DataLoader
|
3 |
+
from multiprocessing import Pool
|
4 |
+
import commons
|
5 |
+
import utils
|
6 |
+
from data_utils import TextAudioSpeakerLoader, TextAudioSpeakerCollate
|
7 |
+
from tqdm import tqdm
|
8 |
+
import warnings
|
9 |
+
|
10 |
+
from text import cleaned_text_to_sequence, get_bert
|
11 |
+
|
12 |
+
config_path = 'configs/base.json'
|
13 |
+
hps = utils.get_hparams_from_file(config_path)
|
14 |
+
|
15 |
+
def process_line(line):
|
16 |
+
_id, spk, language_str, text, phones, tone, word2ph = line.strip().split("|")
|
17 |
+
phone = phones.split(" ")
|
18 |
+
tone = [int(i) for i in tone.split(" ")]
|
19 |
+
word2ph = [int(i) for i in word2ph.split(" ")]
|
20 |
+
w2pho = [i for i in word2ph]
|
21 |
+
word2ph = [i for i in word2ph]
|
22 |
+
phone, tone, language = cleaned_text_to_sequence(phone, tone, language_str)
|
23 |
+
|
24 |
+
if hps.data.add_blank:
|
25 |
+
phone = commons.intersperse(phone, 0)
|
26 |
+
tone = commons.intersperse(tone, 0)
|
27 |
+
language = commons.intersperse(language, 0)
|
28 |
+
for i in range(len(word2ph)):
|
29 |
+
word2ph[i] = word2ph[i] * 2
|
30 |
+
word2ph[0] += 1
|
31 |
+
wav_path = f'{_id}'
|
32 |
+
|
33 |
+
bert_path = wav_path.replace(".wav", ".bert.pt")
|
34 |
+
try:
|
35 |
+
bert = torch.load(bert_path)
|
36 |
+
assert bert.shape[-1] == len(phone)
|
37 |
+
except:
|
38 |
+
bert = get_bert(text, word2ph, language_str)
|
39 |
+
assert bert.shape[-1] == len(phone)
|
40 |
+
torch.save(bert, bert_path)
|
41 |
+
|
42 |
+
|
43 |
+
if __name__ == '__main__':
|
44 |
+
lines = []
|
45 |
+
with open(hps.data.training_files, encoding='utf-8' ) as f:
|
46 |
+
lines.extend(f.readlines())
|
47 |
+
|
48 |
+
# with open(hps.data.validation_files, encoding='utf-8' ) as f:
|
49 |
+
# lines.extend(f.readlines())
|
50 |
+
|
51 |
+
with Pool(processes=2) as pool: #A100 40GB suitable config,if coom,please decrease the processess number.
|
52 |
+
for _ in tqdm(pool.imap_unordered(process_line, lines)):
|
53 |
+
pass
|
train_fix/text/__init__.py
ADDED
@@ -0,0 +1,29 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from text.symbols import *
|
2 |
+
|
3 |
+
|
4 |
+
_symbol_to_id = {s: i for i, s in enumerate(symbols)}
|
5 |
+
|
6 |
+
|
7 |
+
def cleaned_text_to_sequence(cleaned_text, tones, language):
|
8 |
+
"""Converts a string of text to a sequence of IDs corresponding to the symbols in the text.
|
9 |
+
Args:
|
10 |
+
text: string to convert to a sequence
|
11 |
+
Returns:
|
12 |
+
List of integers corresponding to the symbols in the text
|
13 |
+
"""
|
14 |
+
phones = [_symbol_to_id[symbol] for symbol in cleaned_text]
|
15 |
+
tone_start = language_tone_start_map[language]
|
16 |
+
tones = [i + tone_start for i in tones]
|
17 |
+
lang_id = language_id_map[language]
|
18 |
+
lang_ids = [lang_id for i in phones]
|
19 |
+
return phones, tones, lang_ids
|
20 |
+
|
21 |
+
|
22 |
+
def get_bert(norm_text, word2ph, language, device=None):
|
23 |
+
from .chinese_bert import get_bert_feature as zh_bert
|
24 |
+
from .english_bert_mock import get_bert_feature as en_bert
|
25 |
+
from .japanese_bert import get_bert_feature as jp_bert
|
26 |
+
|
27 |
+
lang_bert_func_map = {"ZH": zh_bert, "EN": en_bert, "JP": jp_bert}
|
28 |
+
bert = lang_bert_func_map[language](norm_text, word2ph, device)
|
29 |
+
return bert
|
train_fix/train_ms.py
ADDED
@@ -0,0 +1,607 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# flake8: noqa: E402
|
2 |
+
|
3 |
+
import os
|
4 |
+
import torch
|
5 |
+
from torch.nn import functional as F
|
6 |
+
from torch.utils.data import DataLoader
|
7 |
+
from torch.utils.tensorboard import SummaryWriter
|
8 |
+
import torch.multiprocessing as mp
|
9 |
+
import torch.distributed as dist
|
10 |
+
from torch.nn.parallel import DistributedDataParallel as DDP
|
11 |
+
from torch.cuda.amp import autocast, GradScaler
|
12 |
+
from tqdm import tqdm
|
13 |
+
import logging
|
14 |
+
|
15 |
+
logging.getLogger("numba").setLevel(logging.WARNING)
|
16 |
+
import commons
|
17 |
+
import utils
|
18 |
+
from data_utils import (
|
19 |
+
TextAudioSpeakerLoader,
|
20 |
+
TextAudioSpeakerCollate,
|
21 |
+
DistributedBucketSampler,
|
22 |
+
)
|
23 |
+
from models import (
|
24 |
+
SynthesizerTrn,
|
25 |
+
MultiPeriodDiscriminator,
|
26 |
+
DurationDiscriminator,
|
27 |
+
)
|
28 |
+
from losses import generator_loss, discriminator_loss, feature_loss, kl_loss
|
29 |
+
from mel_processing import mel_spectrogram_torch, spec_to_mel_torch
|
30 |
+
from text.symbols import symbols
|
31 |
+
|
32 |
+
torch.backends.cuda.matmul.allow_tf32 = True
|
33 |
+
torch.backends.cudnn.allow_tf32 = (
|
34 |
+
True # If encontered training problem,please try to disable TF32.
|
35 |
+
)
|
36 |
+
torch.set_float32_matmul_precision("medium")
|
37 |
+
torch.backends.cudnn.benchmark = True
|
38 |
+
torch.backends.cuda.sdp_kernel("flash")
|
39 |
+
torch.backends.cuda.enable_flash_sdp(True)
|
40 |
+
torch.backends.cuda.enable_mem_efficient_sdp(
|
41 |
+
True
|
42 |
+
) # Not available if torch version is lower than 2.0
|
43 |
+
torch.backends.cuda.enable_math_sdp(True)
|
44 |
+
global_step = 0
|
45 |
+
|
46 |
+
|
47 |
+
def main():
|
48 |
+
"""Assume Single Node Multi GPUs Training Only"""
|
49 |
+
assert torch.cuda.is_available(), "CPU training is not allowed."
|
50 |
+
|
51 |
+
n_gpus = torch.cuda.device_count()
|
52 |
+
os.environ['MASTER_ADDR'] = 'localhost'
|
53 |
+
os.environ['MASTER_PORT'] = '65280'
|
54 |
+
|
55 |
+
hps = utils.get_hparams()
|
56 |
+
mp.spawn(run, nprocs=n_gpus, args=(n_gpus, hps,))
|
57 |
+
|
58 |
+
def run(rank, n_gpus, hps):
|
59 |
+
dist.init_process_group(
|
60 |
+
backend="gloo",
|
61 |
+
init_method="env://", # Due to some training problem,we proposed to use gloo instead of nccl.
|
62 |
+
world_size=n_gpus,
|
63 |
+
rank=rank
|
64 |
+
) # Use torchrun instead of mp.spawn
|
65 |
+
# rank = dist.get_rank()
|
66 |
+
# n_gpus = dist.get_world_size()
|
67 |
+
#hps = utils.get_hparams()
|
68 |
+
torch.manual_seed(hps.train.seed)
|
69 |
+
torch.cuda.set_device(rank)
|
70 |
+
global global_step
|
71 |
+
if rank == 0:
|
72 |
+
logger = utils.get_logger(hps.model_dir)
|
73 |
+
logger.info(hps)
|
74 |
+
utils.check_git_hash(hps.model_dir)
|
75 |
+
writer = SummaryWriter(log_dir=hps.model_dir)
|
76 |
+
writer_eval = SummaryWriter(log_dir=os.path.join(hps.model_dir, "eval"))
|
77 |
+
train_dataset = TextAudioSpeakerLoader(hps.data.training_files, hps.data)
|
78 |
+
train_sampler = DistributedBucketSampler(
|
79 |
+
train_dataset,
|
80 |
+
hps.train.batch_size,
|
81 |
+
[32, 300, 400, 500, 600, 700, 800, 900, 1000],
|
82 |
+
num_replicas=n_gpus,
|
83 |
+
rank=rank,
|
84 |
+
shuffle=True,
|
85 |
+
)
|
86 |
+
collate_fn = TextAudioSpeakerCollate()
|
87 |
+
train_loader = DataLoader(
|
88 |
+
train_dataset,
|
89 |
+
num_workers=16,
|
90 |
+
shuffle=False,
|
91 |
+
pin_memory=True,
|
92 |
+
collate_fn=collate_fn,
|
93 |
+
batch_sampler=train_sampler,
|
94 |
+
persistent_workers=True,
|
95 |
+
prefetch_factor=4,
|
96 |
+
) # DataLoader config could be adjusted.
|
97 |
+
if rank == 0:
|
98 |
+
eval_dataset = TextAudioSpeakerLoader(hps.data.validation_files, hps.data)
|
99 |
+
eval_loader = DataLoader(
|
100 |
+
eval_dataset,
|
101 |
+
num_workers=0,
|
102 |
+
shuffle=False,
|
103 |
+
batch_size=1,
|
104 |
+
pin_memory=True,
|
105 |
+
drop_last=False,
|
106 |
+
collate_fn=collate_fn,
|
107 |
+
)
|
108 |
+
if (
|
109 |
+
"use_noise_scaled_mas" in hps.model.keys()
|
110 |
+
and hps.model.use_noise_scaled_mas is True
|
111 |
+
):
|
112 |
+
print("Using noise scaled MAS for VITS2")
|
113 |
+
mas_noise_scale_initial = 0.01
|
114 |
+
noise_scale_delta = 2e-6
|
115 |
+
else:
|
116 |
+
print("Using normal MAS for VITS1")
|
117 |
+
mas_noise_scale_initial = 0.0
|
118 |
+
noise_scale_delta = 0.0
|
119 |
+
if (
|
120 |
+
"use_duration_discriminator" in hps.model.keys()
|
121 |
+
and hps.model.use_duration_discriminator is True
|
122 |
+
):
|
123 |
+
print("Using duration discriminator for VITS2")
|
124 |
+
net_dur_disc = DurationDiscriminator(
|
125 |
+
hps.model.hidden_channels,
|
126 |
+
hps.model.hidden_channels,
|
127 |
+
3,
|
128 |
+
0.1,
|
129 |
+
gin_channels=hps.model.gin_channels if hps.data.n_speakers != 0 else 0,
|
130 |
+
).cuda(rank)
|
131 |
+
if (
|
132 |
+
"use_spk_conditioned_encoder" in hps.model.keys()
|
133 |
+
and hps.model.use_spk_conditioned_encoder is True
|
134 |
+
):
|
135 |
+
if hps.data.n_speakers == 0:
|
136 |
+
raise ValueError(
|
137 |
+
"n_speakers must be > 0 when using spk conditioned encoder to train multi-speaker model"
|
138 |
+
)
|
139 |
+
else:
|
140 |
+
print("Using normal encoder for VITS1")
|
141 |
+
|
142 |
+
net_g = SynthesizerTrn(
|
143 |
+
len(symbols),
|
144 |
+
hps.data.filter_length // 2 + 1,
|
145 |
+
hps.train.segment_size // hps.data.hop_length,
|
146 |
+
n_speakers=hps.data.n_speakers,
|
147 |
+
mas_noise_scale_initial=mas_noise_scale_initial,
|
148 |
+
noise_scale_delta=noise_scale_delta,
|
149 |
+
**hps.model,
|
150 |
+
).cuda(rank)
|
151 |
+
|
152 |
+
net_d = MultiPeriodDiscriminator(hps.model.use_spectral_norm).cuda(rank)
|
153 |
+
optim_g = torch.optim.AdamW(
|
154 |
+
filter(lambda p: p.requires_grad, net_g.parameters()),
|
155 |
+
hps.train.learning_rate,
|
156 |
+
betas=hps.train.betas,
|
157 |
+
eps=hps.train.eps,
|
158 |
+
)
|
159 |
+
optim_d = torch.optim.AdamW(
|
160 |
+
net_d.parameters(),
|
161 |
+
hps.train.learning_rate,
|
162 |
+
betas=hps.train.betas,
|
163 |
+
eps=hps.train.eps,
|
164 |
+
)
|
165 |
+
if net_dur_disc is not None:
|
166 |
+
optim_dur_disc = torch.optim.AdamW(
|
167 |
+
net_dur_disc.parameters(),
|
168 |
+
hps.train.learning_rate,
|
169 |
+
betas=hps.train.betas,
|
170 |
+
eps=hps.train.eps,
|
171 |
+
)
|
172 |
+
else:
|
173 |
+
optim_dur_disc = None
|
174 |
+
net_g = DDP(net_g, device_ids=[rank], find_unused_parameters=True)
|
175 |
+
net_d = DDP(net_d, device_ids=[rank], find_unused_parameters=True)
|
176 |
+
if net_dur_disc is not None:
|
177 |
+
net_dur_disc = DDP(net_dur_disc, device_ids=[rank], find_unused_parameters=True)
|
178 |
+
#dur_resume_lr=0.0003
|
179 |
+
try:
|
180 |
+
if net_dur_disc is not None:
|
181 |
+
_, _, dur_resume_lr, epoch_str = utils.load_checkpoint(
|
182 |
+
utils.latest_checkpoint_path(hps.model_dir, "DUR_*.pth"),
|
183 |
+
net_dur_disc,
|
184 |
+
optim_dur_disc,
|
185 |
+
skip_optimizer=hps.train.skip_optimizer
|
186 |
+
if "skip_optimizer" in hps.train
|
187 |
+
else True,
|
188 |
+
)
|
189 |
+
_, optim_g, g_resume_lr, epoch_str = utils.load_checkpoint(
|
190 |
+
utils.latest_checkpoint_path(hps.model_dir, "G_*.pth"),
|
191 |
+
net_g,
|
192 |
+
optim_g,
|
193 |
+
skip_optimizer=hps.train.skip_optimizer
|
194 |
+
if "skip_optimizer" in hps.train
|
195 |
+
else True,
|
196 |
+
)
|
197 |
+
_, optim_d, d_resume_lr, epoch_str = utils.load_checkpoint(
|
198 |
+
utils.latest_checkpoint_path(hps.model_dir, "D_*.pth"),
|
199 |
+
net_d,
|
200 |
+
optim_d,
|
201 |
+
skip_optimizer=hps.train.skip_optimizer
|
202 |
+
if "skip_optimizer" in hps.train
|
203 |
+
else True,
|
204 |
+
)
|
205 |
+
if not optim_g.param_groups[0].get("initial_lr"):
|
206 |
+
optim_g.param_groups[0]["initial_lr"] = g_resume_lr
|
207 |
+
if not optim_d.param_groups[0].get("initial_lr"):
|
208 |
+
optim_d.param_groups[0]["initial_lr"] = d_resume_lr
|
209 |
+
|
210 |
+
epoch_str = max(epoch_str, 1)
|
211 |
+
global_step = (epoch_str - 1) * len(train_loader)
|
212 |
+
except Exception as e:
|
213 |
+
print(e)
|
214 |
+
epoch_str = 1
|
215 |
+
global_step = 0
|
216 |
+
|
217 |
+
scheduler_g = torch.optim.lr_scheduler.ExponentialLR(
|
218 |
+
optim_g, gamma=hps.train.lr_decay, last_epoch=epoch_str - 2
|
219 |
+
)
|
220 |
+
scheduler_d = torch.optim.lr_scheduler.ExponentialLR(
|
221 |
+
optim_d, gamma=hps.train.lr_decay, last_epoch=epoch_str - 2
|
222 |
+
)
|
223 |
+
if net_dur_disc is not None:
|
224 |
+
if not optim_dur_disc.param_groups[0].get("initial_lr"):
|
225 |
+
optim_dur_disc.param_groups[0]["initial_lr"] = dur_resume_lr
|
226 |
+
scheduler_dur_disc = torch.optim.lr_scheduler.ExponentialLR(
|
227 |
+
optim_dur_disc, gamma=hps.train.lr_decay, last_epoch=epoch_str - 2
|
228 |
+
)
|
229 |
+
else:
|
230 |
+
scheduler_dur_disc = None
|
231 |
+
scaler = GradScaler(enabled=hps.train.fp16_run)
|
232 |
+
|
233 |
+
for epoch in range(epoch_str, hps.train.epochs + 1):
|
234 |
+
if rank == 0:
|
235 |
+
train_and_evaluate(
|
236 |
+
rank,
|
237 |
+
epoch,
|
238 |
+
hps,
|
239 |
+
[net_g, net_d, net_dur_disc],
|
240 |
+
[optim_g, optim_d, optim_dur_disc],
|
241 |
+
[scheduler_g, scheduler_d, scheduler_dur_disc],
|
242 |
+
scaler,
|
243 |
+
[train_loader, eval_loader],
|
244 |
+
logger,
|
245 |
+
[writer, writer_eval],
|
246 |
+
)
|
247 |
+
else:
|
248 |
+
train_and_evaluate(
|
249 |
+
rank,
|
250 |
+
epoch,
|
251 |
+
hps,
|
252 |
+
[net_g, net_d, net_dur_disc],
|
253 |
+
[optim_g, optim_d, optim_dur_disc],
|
254 |
+
[scheduler_g, scheduler_d, scheduler_dur_disc],
|
255 |
+
scaler,
|
256 |
+
[train_loader, None],
|
257 |
+
None,
|
258 |
+
None,
|
259 |
+
)
|
260 |
+
scheduler_g.step()
|
261 |
+
scheduler_d.step()
|
262 |
+
if net_dur_disc is not None:
|
263 |
+
scheduler_dur_disc.step()
|
264 |
+
|
265 |
+
|
266 |
+
def train_and_evaluate(
|
267 |
+
rank, epoch, hps, nets, optims, schedulers, scaler, loaders, logger, writers
|
268 |
+
):
|
269 |
+
net_g, net_d, net_dur_disc = nets
|
270 |
+
optim_g, optim_d, optim_dur_disc = optims
|
271 |
+
scheduler_g, scheduler_d, scheduler_dur_disc = schedulers
|
272 |
+
train_loader, eval_loader = loaders
|
273 |
+
if writers is not None:
|
274 |
+
writer, writer_eval = writers
|
275 |
+
|
276 |
+
train_loader.batch_sampler.set_epoch(epoch)
|
277 |
+
global global_step
|
278 |
+
|
279 |
+
net_g.train()
|
280 |
+
net_d.train()
|
281 |
+
if net_dur_disc is not None:
|
282 |
+
net_dur_disc.train()
|
283 |
+
for batch_idx, (
|
284 |
+
x,
|
285 |
+
x_lengths,
|
286 |
+
spec,
|
287 |
+
spec_lengths,
|
288 |
+
y,
|
289 |
+
y_lengths,
|
290 |
+
speakers,
|
291 |
+
tone,
|
292 |
+
language,
|
293 |
+
bert,
|
294 |
+
ja_bert,
|
295 |
+
) in tqdm(enumerate(train_loader)):
|
296 |
+
if net_g.module.use_noise_scaled_mas:
|
297 |
+
current_mas_noise_scale = (
|
298 |
+
net_g.module.mas_noise_scale_initial
|
299 |
+
- net_g.module.noise_scale_delta * global_step
|
300 |
+
)
|
301 |
+
net_g.module.current_mas_noise_scale = max(current_mas_noise_scale, 0.0)
|
302 |
+
x, x_lengths = x.cuda(rank, non_blocking=True), x_lengths.cuda(
|
303 |
+
rank, non_blocking=True
|
304 |
+
)
|
305 |
+
spec, spec_lengths = spec.cuda(rank, non_blocking=True), spec_lengths.cuda(
|
306 |
+
rank, non_blocking=True
|
307 |
+
)
|
308 |
+
y, y_lengths = y.cuda(rank, non_blocking=True), y_lengths.cuda(
|
309 |
+
rank, non_blocking=True
|
310 |
+
)
|
311 |
+
speakers = speakers.cuda(rank, non_blocking=True)
|
312 |
+
tone = tone.cuda(rank, non_blocking=True)
|
313 |
+
language = language.cuda(rank, non_blocking=True)
|
314 |
+
bert = bert.cuda(rank, non_blocking=True)
|
315 |
+
ja_bert = ja_bert.cuda(rank, non_blocking=True)
|
316 |
+
|
317 |
+
with autocast(enabled=hps.train.fp16_run):
|
318 |
+
(
|
319 |
+
y_hat,
|
320 |
+
l_length,
|
321 |
+
attn,
|
322 |
+
ids_slice,
|
323 |
+
x_mask,
|
324 |
+
z_mask,
|
325 |
+
(z, z_p, m_p, logs_p, m_q, logs_q),
|
326 |
+
(hidden_x, logw, logw_),
|
327 |
+
) = net_g(
|
328 |
+
x,
|
329 |
+
x_lengths,
|
330 |
+
spec,
|
331 |
+
spec_lengths,
|
332 |
+
speakers,
|
333 |
+
tone,
|
334 |
+
language,
|
335 |
+
bert,
|
336 |
+
ja_bert,
|
337 |
+
)
|
338 |
+
mel = spec_to_mel_torch(
|
339 |
+
spec,
|
340 |
+
hps.data.filter_length,
|
341 |
+
hps.data.n_mel_channels,
|
342 |
+
hps.data.sampling_rate,
|
343 |
+
hps.data.mel_fmin,
|
344 |
+
hps.data.mel_fmax,
|
345 |
+
)
|
346 |
+
y_mel = commons.slice_segments(
|
347 |
+
mel, ids_slice, hps.train.segment_size // hps.data.hop_length
|
348 |
+
)
|
349 |
+
y_hat_mel = mel_spectrogram_torch(
|
350 |
+
y_hat.squeeze(1),
|
351 |
+
hps.data.filter_length,
|
352 |
+
hps.data.n_mel_channels,
|
353 |
+
hps.data.sampling_rate,
|
354 |
+
hps.data.hop_length,
|
355 |
+
hps.data.win_length,
|
356 |
+
hps.data.mel_fmin,
|
357 |
+
hps.data.mel_fmax,
|
358 |
+
)
|
359 |
+
|
360 |
+
y = commons.slice_segments(
|
361 |
+
y, ids_slice * hps.data.hop_length, hps.train.segment_size
|
362 |
+
) # slice
|
363 |
+
|
364 |
+
# Discriminator
|
365 |
+
y_d_hat_r, y_d_hat_g, _, _ = net_d(y, y_hat.detach())
|
366 |
+
with autocast(enabled=False):
|
367 |
+
loss_disc, losses_disc_r, losses_disc_g = discriminator_loss(
|
368 |
+
y_d_hat_r, y_d_hat_g
|
369 |
+
)
|
370 |
+
loss_disc_all = loss_disc
|
371 |
+
if net_dur_disc is not None:
|
372 |
+
y_dur_hat_r, y_dur_hat_g = net_dur_disc(
|
373 |
+
hidden_x.detach(), x_mask.detach(), logw.detach(), logw_.detach()
|
374 |
+
)
|
375 |
+
with autocast(enabled=False):
|
376 |
+
# TODO: I think need to mean using the mask, but for now, just mean all
|
377 |
+
(
|
378 |
+
loss_dur_disc,
|
379 |
+
losses_dur_disc_r,
|
380 |
+
losses_dur_disc_g,
|
381 |
+
) = discriminator_loss(y_dur_hat_r, y_dur_hat_g)
|
382 |
+
loss_dur_disc_all = loss_dur_disc
|
383 |
+
optim_dur_disc.zero_grad()
|
384 |
+
scaler.scale(loss_dur_disc_all).backward()
|
385 |
+
scaler.unscale_(optim_dur_disc)
|
386 |
+
commons.clip_grad_value_(net_dur_disc.parameters(), None)
|
387 |
+
scaler.step(optim_dur_disc)
|
388 |
+
|
389 |
+
optim_d.zero_grad()
|
390 |
+
scaler.scale(loss_disc_all).backward()
|
391 |
+
scaler.unscale_(optim_d)
|
392 |
+
grad_norm_d = commons.clip_grad_value_(net_d.parameters(), None)
|
393 |
+
scaler.step(optim_d)
|
394 |
+
|
395 |
+
with autocast(enabled=hps.train.fp16_run):
|
396 |
+
# Generator
|
397 |
+
y_d_hat_r, y_d_hat_g, fmap_r, fmap_g = net_d(y, y_hat)
|
398 |
+
if net_dur_disc is not None:
|
399 |
+
y_dur_hat_r, y_dur_hat_g = net_dur_disc(hidden_x, x_mask, logw, logw_)
|
400 |
+
with autocast(enabled=False):
|
401 |
+
loss_dur = torch.sum(l_length.float())
|
402 |
+
loss_mel = F.l1_loss(y_mel, y_hat_mel) * hps.train.c_mel
|
403 |
+
loss_kl = kl_loss(z_p, logs_q, m_p, logs_p, z_mask) * hps.train.c_kl
|
404 |
+
|
405 |
+
loss_fm = feature_loss(fmap_r, fmap_g)
|
406 |
+
loss_gen, losses_gen = generator_loss(y_d_hat_g)
|
407 |
+
loss_gen_all = loss_gen + loss_fm + loss_mel + loss_dur + loss_kl
|
408 |
+
if net_dur_disc is not None:
|
409 |
+
loss_dur_gen, losses_dur_gen = generator_loss(y_dur_hat_g)
|
410 |
+
loss_gen_all += loss_dur_gen
|
411 |
+
optim_g.zero_grad()
|
412 |
+
scaler.scale(loss_gen_all).backward()
|
413 |
+
scaler.unscale_(optim_g)
|
414 |
+
grad_norm_g = commons.clip_grad_value_(net_g.parameters(), None)
|
415 |
+
scaler.step(optim_g)
|
416 |
+
scaler.update()
|
417 |
+
|
418 |
+
if rank == 0:
|
419 |
+
if global_step % hps.train.log_interval == 0:
|
420 |
+
lr = optim_g.param_groups[0]["lr"]
|
421 |
+
losses = [loss_disc, loss_gen, loss_fm, loss_mel, loss_dur, loss_kl]
|
422 |
+
logger.info(
|
423 |
+
"Train Epoch: {} [{:.0f}%]".format(
|
424 |
+
epoch, 100.0 * batch_idx / len(train_loader)
|
425 |
+
)
|
426 |
+
)
|
427 |
+
logger.info([x.item() for x in losses] + [global_step, lr])
|
428 |
+
|
429 |
+
scalar_dict = {
|
430 |
+
"loss/g/total": loss_gen_all,
|
431 |
+
"loss/d/total": loss_disc_all,
|
432 |
+
"learning_rate": lr,
|
433 |
+
"grad_norm_d": grad_norm_d,
|
434 |
+
"grad_norm_g": grad_norm_g,
|
435 |
+
}
|
436 |
+
scalar_dict.update(
|
437 |
+
{
|
438 |
+
"loss/g/fm": loss_fm,
|
439 |
+
"loss/g/mel": loss_mel,
|
440 |
+
"loss/g/dur": loss_dur,
|
441 |
+
"loss/g/kl": loss_kl,
|
442 |
+
}
|
443 |
+
)
|
444 |
+
scalar_dict.update(
|
445 |
+
{"loss/g/{}".format(i): v for i, v in enumerate(losses_gen)}
|
446 |
+
)
|
447 |
+
scalar_dict.update(
|
448 |
+
{"loss/d_r/{}".format(i): v for i, v in enumerate(losses_disc_r)}
|
449 |
+
)
|
450 |
+
scalar_dict.update(
|
451 |
+
{"loss/d_g/{}".format(i): v for i, v in enumerate(losses_disc_g)}
|
452 |
+
)
|
453 |
+
|
454 |
+
image_dict = {
|
455 |
+
"slice/mel_org": utils.plot_spectrogram_to_numpy(
|
456 |
+
y_mel[0].data.cpu().numpy()
|
457 |
+
),
|
458 |
+
"slice/mel_gen": utils.plot_spectrogram_to_numpy(
|
459 |
+
y_hat_mel[0].data.cpu().numpy()
|
460 |
+
),
|
461 |
+
"all/mel": utils.plot_spectrogram_to_numpy(
|
462 |
+
mel[0].data.cpu().numpy()
|
463 |
+
),
|
464 |
+
"all/attn": utils.plot_alignment_to_numpy(
|
465 |
+
attn[0, 0].data.cpu().numpy()
|
466 |
+
),
|
467 |
+
}
|
468 |
+
utils.summarize(
|
469 |
+
writer=writer,
|
470 |
+
global_step=global_step,
|
471 |
+
images=image_dict,
|
472 |
+
scalars=scalar_dict,
|
473 |
+
)
|
474 |
+
|
475 |
+
if global_step % hps.train.eval_interval == 0:
|
476 |
+
evaluate(hps, net_g, eval_loader, writer_eval)
|
477 |
+
utils.save_checkpoint(
|
478 |
+
net_g,
|
479 |
+
optim_g,
|
480 |
+
hps.train.learning_rate,
|
481 |
+
epoch,
|
482 |
+
os.path.join(hps.model_dir, "G_{}.pth".format(global_step)),
|
483 |
+
)
|
484 |
+
utils.save_checkpoint(
|
485 |
+
net_d,
|
486 |
+
optim_d,
|
487 |
+
hps.train.learning_rate,
|
488 |
+
epoch,
|
489 |
+
os.path.join(hps.model_dir, "D_{}.pth".format(global_step)),
|
490 |
+
)
|
491 |
+
if net_dur_disc is not None:
|
492 |
+
utils.save_checkpoint(
|
493 |
+
net_dur_disc,
|
494 |
+
optim_dur_disc,
|
495 |
+
hps.train.learning_rate,
|
496 |
+
epoch,
|
497 |
+
os.path.join(hps.model_dir, "DUR_{}.pth".format(global_step)),
|
498 |
+
)
|
499 |
+
keep_ckpts = getattr(hps.train, "keep_ckpts", 5)
|
500 |
+
if keep_ckpts > 0:
|
501 |
+
utils.clean_checkpoints(
|
502 |
+
path_to_models=hps.model_dir,
|
503 |
+
n_ckpts_to_keep=keep_ckpts,
|
504 |
+
sort_by_time=True,
|
505 |
+
)
|
506 |
+
|
507 |
+
global_step += 1
|
508 |
+
|
509 |
+
if rank == 0:
|
510 |
+
logger.info("====> Epoch: {}".format(epoch))
|
511 |
+
|
512 |
+
|
513 |
+
def evaluate(hps, generator, eval_loader, writer_eval):
|
514 |
+
generator.eval()
|
515 |
+
image_dict = {}
|
516 |
+
audio_dict = {}
|
517 |
+
print("Evaluating ...")
|
518 |
+
with torch.no_grad():
|
519 |
+
for batch_idx, (
|
520 |
+
x,
|
521 |
+
x_lengths,
|
522 |
+
spec,
|
523 |
+
spec_lengths,
|
524 |
+
y,
|
525 |
+
y_lengths,
|
526 |
+
speakers,
|
527 |
+
tone,
|
528 |
+
language,
|
529 |
+
bert,
|
530 |
+
ja_bert,
|
531 |
+
) in enumerate(eval_loader):
|
532 |
+
x, x_lengths = x.cuda(), x_lengths.cuda()
|
533 |
+
spec, spec_lengths = spec.cuda(), spec_lengths.cuda()
|
534 |
+
y, y_lengths = y.cuda(), y_lengths.cuda()
|
535 |
+
speakers = speakers.cuda()
|
536 |
+
bert = bert.cuda()
|
537 |
+
ja_bert = ja_bert.cuda()
|
538 |
+
tone = tone.cuda()
|
539 |
+
language = language.cuda()
|
540 |
+
for use_sdp in [True, False]:
|
541 |
+
y_hat, attn, mask, *_ = generator.module.infer(
|
542 |
+
x,
|
543 |
+
x_lengths,
|
544 |
+
speakers,
|
545 |
+
tone,
|
546 |
+
language,
|
547 |
+
bert,
|
548 |
+
ja_bert,
|
549 |
+
y=spec,
|
550 |
+
max_len=1000,
|
551 |
+
sdp_ratio=0.0 if not use_sdp else 1.0,
|
552 |
+
)
|
553 |
+
y_hat_lengths = mask.sum([1, 2]).long() * hps.data.hop_length
|
554 |
+
|
555 |
+
mel = spec_to_mel_torch(
|
556 |
+
spec,
|
557 |
+
hps.data.filter_length,
|
558 |
+
hps.data.n_mel_channels,
|
559 |
+
hps.data.sampling_rate,
|
560 |
+
hps.data.mel_fmin,
|
561 |
+
hps.data.mel_fmax,
|
562 |
+
)
|
563 |
+
y_hat_mel = mel_spectrogram_torch(
|
564 |
+
y_hat.squeeze(1).float(),
|
565 |
+
hps.data.filter_length,
|
566 |
+
hps.data.n_mel_channels,
|
567 |
+
hps.data.sampling_rate,
|
568 |
+
hps.data.hop_length,
|
569 |
+
hps.data.win_length,
|
570 |
+
hps.data.mel_fmin,
|
571 |
+
hps.data.mel_fmax,
|
572 |
+
)
|
573 |
+
image_dict.update(
|
574 |
+
{
|
575 |
+
f"gen/mel_{batch_idx}": utils.plot_spectrogram_to_numpy(
|
576 |
+
y_hat_mel[0].cpu().numpy()
|
577 |
+
)
|
578 |
+
}
|
579 |
+
)
|
580 |
+
audio_dict.update(
|
581 |
+
{
|
582 |
+
f"gen/audio_{batch_idx}_{use_sdp}": y_hat[
|
583 |
+
0, :, : y_hat_lengths[0]
|
584 |
+
]
|
585 |
+
}
|
586 |
+
)
|
587 |
+
image_dict.update(
|
588 |
+
{
|
589 |
+
f"gt/mel_{batch_idx}": utils.plot_spectrogram_to_numpy(
|
590 |
+
mel[0].cpu().numpy()
|
591 |
+
)
|
592 |
+
}
|
593 |
+
)
|
594 |
+
audio_dict.update({f"gt/audio_{batch_idx}": y[0, :, : y_lengths[0]]})
|
595 |
+
|
596 |
+
utils.summarize(
|
597 |
+
writer=writer_eval,
|
598 |
+
global_step=global_step,
|
599 |
+
images=image_dict,
|
600 |
+
audios=audio_dict,
|
601 |
+
audio_sampling_rate=hps.data.sampling_rate,
|
602 |
+
)
|
603 |
+
generator.train()
|
604 |
+
|
605 |
+
|
606 |
+
if __name__ == "__main__":
|
607 |
+
main()
|