Spaces:
Running
Running
Upload train.py
Browse files
train.py
ADDED
@@ -0,0 +1,643 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import sys
|
3 |
+
import logging
|
4 |
+
|
5 |
+
logger = logging.getLogger(__name__)
|
6 |
+
|
7 |
+
now_dir = os.getcwd()
|
8 |
+
sys.path.append(os.path.join(now_dir))
|
9 |
+
|
10 |
+
import datetime
|
11 |
+
|
12 |
+
from infer.lib.train import utils
|
13 |
+
|
14 |
+
hps = utils.get_hparams()
|
15 |
+
os.environ["CUDA_VISIBLE_DEVICES"] = hps.gpus.replace("-", ",")
|
16 |
+
n_gpus = len(hps.gpus.split("-"))
|
17 |
+
from random import randint, shuffle
|
18 |
+
|
19 |
+
import torch
|
20 |
+
|
21 |
+
try:
|
22 |
+
import intel_extension_for_pytorch as ipex # pylint: disable=import-error, unused-import
|
23 |
+
|
24 |
+
if torch.xpu.is_available():
|
25 |
+
from infer.modules.ipex import ipex_init
|
26 |
+
from infer.modules.ipex.gradscaler import gradscaler_init
|
27 |
+
from torch.xpu.amp import autocast
|
28 |
+
|
29 |
+
GradScaler = gradscaler_init()
|
30 |
+
ipex_init()
|
31 |
+
else:
|
32 |
+
from torch.cuda.amp import GradScaler, autocast
|
33 |
+
except Exception:
|
34 |
+
from torch.cuda.amp import GradScaler, autocast
|
35 |
+
|
36 |
+
torch.backends.cudnn.deterministic = False
|
37 |
+
torch.backends.cudnn.benchmark = False
|
38 |
+
from time import sleep
|
39 |
+
from time import time as ttime
|
40 |
+
|
41 |
+
import torch.distributed as dist
|
42 |
+
import torch.multiprocessing as mp
|
43 |
+
from torch.nn import functional as F
|
44 |
+
from torch.nn.parallel import DistributedDataParallel as DDP
|
45 |
+
from torch.utils.data import DataLoader
|
46 |
+
from torch.utils.tensorboard import SummaryWriter
|
47 |
+
|
48 |
+
from infer.lib import commons
|
49 |
+
from infer.lib.data_utils import (
|
50 |
+
DistributedBucketSampler,
|
51 |
+
TextAudioCollate,
|
52 |
+
TextAudioCollateMultiNSFsid,
|
53 |
+
TextAudioLoader,
|
54 |
+
TextAudioLoaderMultiNSFsid,
|
55 |
+
)
|
56 |
+
|
57 |
+
if hps.version == "v1":
|
58 |
+
from infer.lib.infer_pack.models import MultiPeriodDiscriminator
|
59 |
+
from infer.lib.infer_pack.models import SynthesizerTrnMs256NSFsid as RVC_Model_f0
|
60 |
+
from infer.lib.infer_pack.models import (
|
61 |
+
SynthesizerTrnMs256NSFsid_nono as RVC_Model_nof0,
|
62 |
+
)
|
63 |
+
else:
|
64 |
+
from infer.lib.infer_pack.models import (
|
65 |
+
SynthesizerTrnMs768NSFsid as RVC_Model_f0,
|
66 |
+
SynthesizerTrnMs768NSFsid_nono as RVC_Model_nof0,
|
67 |
+
MultiPeriodDiscriminatorV2 as MultiPeriodDiscriminator,
|
68 |
+
)
|
69 |
+
|
70 |
+
from infer.lib.train.losses import (
|
71 |
+
discriminator_loss,
|
72 |
+
feature_loss,
|
73 |
+
generator_loss,
|
74 |
+
kl_loss,
|
75 |
+
)
|
76 |
+
from infer.lib.train.mel_processing import mel_spectrogram_torch, spec_to_mel_torch
|
77 |
+
from infer.lib.train.process_ckpt import savee
|
78 |
+
|
79 |
+
global_step = 0
|
80 |
+
|
81 |
+
|
82 |
+
class EpochRecorder:
|
83 |
+
def __init__(self):
|
84 |
+
self.last_time = ttime()
|
85 |
+
|
86 |
+
def record(self):
|
87 |
+
now_time = ttime()
|
88 |
+
elapsed_time = now_time - self.last_time
|
89 |
+
self.last_time = now_time
|
90 |
+
elapsed_time_str = str(datetime.timedelta(seconds=elapsed_time))
|
91 |
+
current_time = datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S")
|
92 |
+
return f"[{current_time}] | ({elapsed_time_str})"
|
93 |
+
|
94 |
+
|
95 |
+
def main():
|
96 |
+
n_gpus = torch.cuda.device_count()
|
97 |
+
|
98 |
+
if torch.cuda.is_available() == False and torch.backends.mps.is_available() == True:
|
99 |
+
n_gpus = 1
|
100 |
+
if n_gpus < 1:
|
101 |
+
# patch to unblock people without gpus. there is probably a better way.
|
102 |
+
print("NO GPU DETECTED: falling back to CPU - this may take a while")
|
103 |
+
n_gpus = 1
|
104 |
+
os.environ["MASTER_ADDR"] = "localhost"
|
105 |
+
os.environ["MASTER_PORT"] = str(randint(20000, 55555))
|
106 |
+
children = []
|
107 |
+
for i in range(n_gpus):
|
108 |
+
subproc = mp.Process(
|
109 |
+
target=run,
|
110 |
+
args=(i, n_gpus, hps),
|
111 |
+
)
|
112 |
+
children.append(subproc)
|
113 |
+
subproc.start()
|
114 |
+
|
115 |
+
for i in range(n_gpus):
|
116 |
+
children[i].join()
|
117 |
+
|
118 |
+
|
119 |
+
def run(
|
120 |
+
rank,
|
121 |
+
n_gpus,
|
122 |
+
hps,
|
123 |
+
):
|
124 |
+
global global_step
|
125 |
+
if rank == 0:
|
126 |
+
logger = utils.get_logger(hps.model_dir)
|
127 |
+
logger.info(hps)
|
128 |
+
# utils.check_git_hash(hps.model_dir)
|
129 |
+
writer = SummaryWriter(log_dir=hps.model_dir)
|
130 |
+
writer_eval = SummaryWriter(log_dir=os.path.join(hps.model_dir, "eval"))
|
131 |
+
|
132 |
+
dist.init_process_group(
|
133 |
+
backend="gloo", init_method="env://", world_size=n_gpus, rank=rank
|
134 |
+
)
|
135 |
+
torch.manual_seed(hps.train.seed)
|
136 |
+
if torch.cuda.is_available():
|
137 |
+
torch.cuda.set_device(rank)
|
138 |
+
|
139 |
+
if hps.if_f0 == 1:
|
140 |
+
train_dataset = TextAudioLoaderMultiNSFsid(hps.data.training_files, hps.data)
|
141 |
+
else:
|
142 |
+
train_dataset = TextAudioLoader(hps.data.training_files, hps.data)
|
143 |
+
train_sampler = DistributedBucketSampler(
|
144 |
+
train_dataset,
|
145 |
+
hps.train.batch_size * n_gpus,
|
146 |
+
# [100, 200, 300, 400, 500, 600, 700, 800, 900, 1000, 1200,1400], # 16s
|
147 |
+
[100, 200, 300, 400, 500, 600, 700, 800, 900], # 16s
|
148 |
+
num_replicas=n_gpus,
|
149 |
+
rank=rank,
|
150 |
+
shuffle=True,
|
151 |
+
)
|
152 |
+
# It is possible that dataloader's workers are out of shared memory. Please try to raise your shared memory limit.
|
153 |
+
# num_workers=8 -> num_workers=4
|
154 |
+
if hps.if_f0 == 1:
|
155 |
+
collate_fn = TextAudioCollateMultiNSFsid()
|
156 |
+
else:
|
157 |
+
collate_fn = TextAudioCollate()
|
158 |
+
train_loader = DataLoader(
|
159 |
+
train_dataset,
|
160 |
+
num_workers=4,
|
161 |
+
shuffle=False,
|
162 |
+
pin_memory=True,
|
163 |
+
collate_fn=collate_fn,
|
164 |
+
batch_sampler=train_sampler,
|
165 |
+
persistent_workers=True,
|
166 |
+
prefetch_factor=8,
|
167 |
+
)
|
168 |
+
if hps.if_f0 == 1:
|
169 |
+
net_g = RVC_Model_f0(
|
170 |
+
hps.data.filter_length // 2 + 1,
|
171 |
+
hps.train.segment_size // hps.data.hop_length,
|
172 |
+
**hps.model,
|
173 |
+
is_half=hps.train.fp16_run,
|
174 |
+
sr=hps.sample_rate,
|
175 |
+
)
|
176 |
+
else:
|
177 |
+
net_g = RVC_Model_nof0(
|
178 |
+
hps.data.filter_length // 2 + 1,
|
179 |
+
hps.train.segment_size // hps.data.hop_length,
|
180 |
+
**hps.model,
|
181 |
+
is_half=hps.train.fp16_run,
|
182 |
+
)
|
183 |
+
if torch.cuda.is_available():
|
184 |
+
net_g = net_g.cuda(rank)
|
185 |
+
net_d = MultiPeriodDiscriminator(hps.model.use_spectral_norm)
|
186 |
+
if torch.cuda.is_available():
|
187 |
+
net_d = net_d.cuda(rank)
|
188 |
+
optim_g = torch.optim.AdamW(
|
189 |
+
net_g.parameters(),
|
190 |
+
hps.train.learning_rate,
|
191 |
+
betas=hps.train.betas,
|
192 |
+
eps=hps.train.eps,
|
193 |
+
)
|
194 |
+
optim_d = torch.optim.AdamW(
|
195 |
+
net_d.parameters(),
|
196 |
+
hps.train.learning_rate,
|
197 |
+
betas=hps.train.betas,
|
198 |
+
eps=hps.train.eps,
|
199 |
+
)
|
200 |
+
# net_g = DDP(net_g, device_ids=[rank], find_unused_parameters=True)
|
201 |
+
# net_d = DDP(net_d, device_ids=[rank], find_unused_parameters=True)
|
202 |
+
if hasattr(torch, "xpu") and torch.xpu.is_available():
|
203 |
+
pass
|
204 |
+
elif torch.cuda.is_available():
|
205 |
+
net_g = DDP(net_g, device_ids=[rank])
|
206 |
+
net_d = DDP(net_d, device_ids=[rank])
|
207 |
+
else:
|
208 |
+
net_g = DDP(net_g)
|
209 |
+
net_d = DDP(net_d)
|
210 |
+
|
211 |
+
try: # 如果能加载自动resume
|
212 |
+
_, _, _, epoch_str = utils.load_checkpoint(
|
213 |
+
utils.latest_checkpoint_path(hps.model_dir, "D_*.pth"), net_d, optim_d
|
214 |
+
) # D多半加载没事
|
215 |
+
if rank == 0:
|
216 |
+
logger.info("loaded D")
|
217 |
+
# _, _, _, epoch_str = utils.load_checkpoint(utils.latest_checkpoint_path(hps.model_dir, "G_*.pth"), net_g, optim_g,load_opt=0)
|
218 |
+
_, _, _, epoch_str = utils.load_checkpoint(
|
219 |
+
utils.latest_checkpoint_path(hps.model_dir, "G_*.pth"), net_g, optim_g
|
220 |
+
)
|
221 |
+
global_step = (epoch_str - 1) * len(train_loader)
|
222 |
+
# epoch_str = 1
|
223 |
+
# global_step = 0
|
224 |
+
except: # 如果首次不能加载,加载pretrain
|
225 |
+
# traceback.print_exc()
|
226 |
+
epoch_str = 1
|
227 |
+
global_step = 0
|
228 |
+
if hps.pretrainG != "":
|
229 |
+
if rank == 0:
|
230 |
+
logger.info("loaded pretrained %s" % (hps.pretrainG))
|
231 |
+
if hasattr(net_g, "module"):
|
232 |
+
logger.info(
|
233 |
+
net_g.module.load_state_dict(
|
234 |
+
torch.load(hps.pretrainG, map_location="cpu")["model"]
|
235 |
+
)
|
236 |
+
) ##测试不加载优化器
|
237 |
+
else:
|
238 |
+
logger.info(
|
239 |
+
net_g.load_state_dict(
|
240 |
+
torch.load(hps.pretrainG, map_location="cpu")["model"]
|
241 |
+
)
|
242 |
+
) ##测试不加载优化器
|
243 |
+
if hps.pretrainD != "":
|
244 |
+
if rank == 0:
|
245 |
+
logger.info("loaded pretrained %s" % (hps.pretrainD))
|
246 |
+
if hasattr(net_d, "module"):
|
247 |
+
logger.info(
|
248 |
+
net_d.module.load_state_dict(
|
249 |
+
torch.load(hps.pretrainD, map_location="cpu")["model"]
|
250 |
+
)
|
251 |
+
)
|
252 |
+
else:
|
253 |
+
logger.info(
|
254 |
+
net_d.load_state_dict(
|
255 |
+
torch.load(hps.pretrainD, map_location="cpu")["model"]
|
256 |
+
)
|
257 |
+
)
|
258 |
+
|
259 |
+
scheduler_g = torch.optim.lr_scheduler.ExponentialLR(
|
260 |
+
optim_g, gamma=hps.train.lr_decay, last_epoch=epoch_str - 2
|
261 |
+
)
|
262 |
+
scheduler_d = torch.optim.lr_scheduler.ExponentialLR(
|
263 |
+
optim_d, gamma=hps.train.lr_decay, last_epoch=epoch_str - 2
|
264 |
+
)
|
265 |
+
|
266 |
+
scaler = GradScaler(enabled=hps.train.fp16_run)
|
267 |
+
|
268 |
+
cache = []
|
269 |
+
for epoch in range(epoch_str, hps.train.epochs + 1):
|
270 |
+
if rank == 0:
|
271 |
+
train_and_evaluate(
|
272 |
+
rank,
|
273 |
+
epoch,
|
274 |
+
hps,
|
275 |
+
[net_g, net_d],
|
276 |
+
[optim_g, optim_d],
|
277 |
+
[scheduler_g, scheduler_d],
|
278 |
+
scaler,
|
279 |
+
[train_loader, None],
|
280 |
+
logger,
|
281 |
+
[writer, writer_eval],
|
282 |
+
cache,
|
283 |
+
)
|
284 |
+
else:
|
285 |
+
train_and_evaluate(
|
286 |
+
rank,
|
287 |
+
epoch,
|
288 |
+
hps,
|
289 |
+
[net_g, net_d],
|
290 |
+
[optim_g, optim_d],
|
291 |
+
[scheduler_g, scheduler_d],
|
292 |
+
scaler,
|
293 |
+
[train_loader, None],
|
294 |
+
None,
|
295 |
+
None,
|
296 |
+
cache,
|
297 |
+
)
|
298 |
+
scheduler_g.step()
|
299 |
+
scheduler_d.step()
|
300 |
+
|
301 |
+
|
302 |
+
def train_and_evaluate(
|
303 |
+
rank, epoch, hps, nets, optims, schedulers, scaler, loaders, logger, writers, cache
|
304 |
+
):
|
305 |
+
net_g, net_d = nets
|
306 |
+
optim_g, optim_d = optims
|
307 |
+
train_loader, eval_loader = loaders
|
308 |
+
if writers is not None:
|
309 |
+
writer, writer_eval = writers
|
310 |
+
|
311 |
+
train_loader.batch_sampler.set_epoch(epoch)
|
312 |
+
global global_step
|
313 |
+
|
314 |
+
net_g.train()
|
315 |
+
net_d.train()
|
316 |
+
|
317 |
+
# Prepare data iterator
|
318 |
+
if hps.if_cache_data_in_gpu == True:
|
319 |
+
# Use Cache
|
320 |
+
data_iterator = cache
|
321 |
+
if cache == []:
|
322 |
+
# Make new cache
|
323 |
+
for batch_idx, info in enumerate(train_loader):
|
324 |
+
# Unpack
|
325 |
+
if hps.if_f0 == 1:
|
326 |
+
(
|
327 |
+
phone,
|
328 |
+
phone_lengths,
|
329 |
+
pitch,
|
330 |
+
pitchf,
|
331 |
+
spec,
|
332 |
+
spec_lengths,
|
333 |
+
wave,
|
334 |
+
wave_lengths,
|
335 |
+
sid,
|
336 |
+
) = info
|
337 |
+
else:
|
338 |
+
(
|
339 |
+
phone,
|
340 |
+
phone_lengths,
|
341 |
+
spec,
|
342 |
+
spec_lengths,
|
343 |
+
wave,
|
344 |
+
wave_lengths,
|
345 |
+
sid,
|
346 |
+
) = info
|
347 |
+
# Load on CUDA
|
348 |
+
if torch.cuda.is_available():
|
349 |
+
phone = phone.cuda(rank, non_blocking=True)
|
350 |
+
phone_lengths = phone_lengths.cuda(rank, non_blocking=True)
|
351 |
+
if hps.if_f0 == 1:
|
352 |
+
pitch = pitch.cuda(rank, non_blocking=True)
|
353 |
+
pitchf = pitchf.cuda(rank, non_blocking=True)
|
354 |
+
sid = sid.cuda(rank, non_blocking=True)
|
355 |
+
spec = spec.cuda(rank, non_blocking=True)
|
356 |
+
spec_lengths = spec_lengths.cuda(rank, non_blocking=True)
|
357 |
+
wave = wave.cuda(rank, non_blocking=True)
|
358 |
+
wave_lengths = wave_lengths.cuda(rank, non_blocking=True)
|
359 |
+
# Cache on list
|
360 |
+
if hps.if_f0 == 1:
|
361 |
+
cache.append(
|
362 |
+
(
|
363 |
+
batch_idx,
|
364 |
+
(
|
365 |
+
phone,
|
366 |
+
phone_lengths,
|
367 |
+
pitch,
|
368 |
+
pitchf,
|
369 |
+
spec,
|
370 |
+
spec_lengths,
|
371 |
+
wave,
|
372 |
+
wave_lengths,
|
373 |
+
sid,
|
374 |
+
),
|
375 |
+
)
|
376 |
+
)
|
377 |
+
else:
|
378 |
+
cache.append(
|
379 |
+
(
|
380 |
+
batch_idx,
|
381 |
+
(
|
382 |
+
phone,
|
383 |
+
phone_lengths,
|
384 |
+
spec,
|
385 |
+
spec_lengths,
|
386 |
+
wave,
|
387 |
+
wave_lengths,
|
388 |
+
sid,
|
389 |
+
),
|
390 |
+
)
|
391 |
+
)
|
392 |
+
else:
|
393 |
+
# Load shuffled cache
|
394 |
+
shuffle(cache)
|
395 |
+
else:
|
396 |
+
# Loader
|
397 |
+
data_iterator = enumerate(train_loader)
|
398 |
+
|
399 |
+
# Run steps
|
400 |
+
epoch_recorder = EpochRecorder()
|
401 |
+
for batch_idx, info in data_iterator:
|
402 |
+
# Data
|
403 |
+
## Unpack
|
404 |
+
if hps.if_f0 == 1:
|
405 |
+
(
|
406 |
+
phone,
|
407 |
+
phone_lengths,
|
408 |
+
pitch,
|
409 |
+
pitchf,
|
410 |
+
spec,
|
411 |
+
spec_lengths,
|
412 |
+
wave,
|
413 |
+
wave_lengths,
|
414 |
+
sid,
|
415 |
+
) = info
|
416 |
+
else:
|
417 |
+
phone, phone_lengths, spec, spec_lengths, wave, wave_lengths, sid = info
|
418 |
+
## Load on CUDA
|
419 |
+
if (hps.if_cache_data_in_gpu == False) and torch.cuda.is_available():
|
420 |
+
phone = phone.cuda(rank, non_blocking=True)
|
421 |
+
phone_lengths = phone_lengths.cuda(rank, non_blocking=True)
|
422 |
+
if hps.if_f0 == 1:
|
423 |
+
pitch = pitch.cuda(rank, non_blocking=True)
|
424 |
+
pitchf = pitchf.cuda(rank, non_blocking=True)
|
425 |
+
sid = sid.cuda(rank, non_blocking=True)
|
426 |
+
spec = spec.cuda(rank, non_blocking=True)
|
427 |
+
spec_lengths = spec_lengths.cuda(rank, non_blocking=True)
|
428 |
+
wave = wave.cuda(rank, non_blocking=True)
|
429 |
+
# wave_lengths = wave_lengths.cuda(rank, non_blocking=True)
|
430 |
+
|
431 |
+
# Calculate
|
432 |
+
with autocast(enabled=hps.train.fp16_run):
|
433 |
+
if hps.if_f0 == 1:
|
434 |
+
(
|
435 |
+
y_hat,
|
436 |
+
ids_slice,
|
437 |
+
x_mask,
|
438 |
+
z_mask,
|
439 |
+
(z, z_p, m_p, logs_p, m_q, logs_q),
|
440 |
+
) = net_g(phone, phone_lengths, pitch, pitchf, spec, spec_lengths, sid)
|
441 |
+
else:
|
442 |
+
(
|
443 |
+
y_hat,
|
444 |
+
ids_slice,
|
445 |
+
x_mask,
|
446 |
+
z_mask,
|
447 |
+
(z, z_p, m_p, logs_p, m_q, logs_q),
|
448 |
+
) = net_g(phone, phone_lengths, spec, spec_lengths, sid)
|
449 |
+
mel = spec_to_mel_torch(
|
450 |
+
spec,
|
451 |
+
hps.data.filter_length,
|
452 |
+
hps.data.n_mel_channels,
|
453 |
+
hps.data.sampling_rate,
|
454 |
+
hps.data.mel_fmin,
|
455 |
+
hps.data.mel_fmax,
|
456 |
+
)
|
457 |
+
y_mel = commons.slice_segments(
|
458 |
+
mel, ids_slice, hps.train.segment_size // hps.data.hop_length
|
459 |
+
)
|
460 |
+
with autocast(enabled=False):
|
461 |
+
y_hat_mel = mel_spectrogram_torch(
|
462 |
+
y_hat.float().squeeze(1),
|
463 |
+
hps.data.filter_length,
|
464 |
+
hps.data.n_mel_channels,
|
465 |
+
hps.data.sampling_rate,
|
466 |
+
hps.data.hop_length,
|
467 |
+
hps.data.win_length,
|
468 |
+
hps.data.mel_fmin,
|
469 |
+
hps.data.mel_fmax,
|
470 |
+
)
|
471 |
+
if hps.train.fp16_run == True:
|
472 |
+
y_hat_mel = y_hat_mel.half()
|
473 |
+
wave = commons.slice_segments(
|
474 |
+
wave, ids_slice * hps.data.hop_length, hps.train.segment_size
|
475 |
+
) # slice
|
476 |
+
|
477 |
+
# Discriminator
|
478 |
+
y_d_hat_r, y_d_hat_g, _, _ = net_d(wave, y_hat.detach())
|
479 |
+
with autocast(enabled=False):
|
480 |
+
loss_disc, losses_disc_r, losses_disc_g = discriminator_loss(
|
481 |
+
y_d_hat_r, y_d_hat_g
|
482 |
+
)
|
483 |
+
optim_d.zero_grad()
|
484 |
+
scaler.scale(loss_disc).backward()
|
485 |
+
scaler.unscale_(optim_d)
|
486 |
+
grad_norm_d = commons.clip_grad_value_(net_d.parameters(), None)
|
487 |
+
scaler.step(optim_d)
|
488 |
+
|
489 |
+
with autocast(enabled=hps.train.fp16_run):
|
490 |
+
# Generator
|
491 |
+
y_d_hat_r, y_d_hat_g, fmap_r, fmap_g = net_d(wave, y_hat)
|
492 |
+
with autocast(enabled=False):
|
493 |
+
loss_mel = F.l1_loss(y_mel, y_hat_mel) * hps.train.c_mel
|
494 |
+
loss_kl = kl_loss(z_p, logs_q, m_p, logs_p, z_mask) * hps.train.c_kl
|
495 |
+
loss_fm = feature_loss(fmap_r, fmap_g)
|
496 |
+
loss_gen, losses_gen = generator_loss(y_d_hat_g)
|
497 |
+
loss_gen_all = loss_gen + loss_fm + loss_mel + loss_kl
|
498 |
+
optim_g.zero_grad()
|
499 |
+
scaler.scale(loss_gen_all).backward()
|
500 |
+
scaler.unscale_(optim_g)
|
501 |
+
grad_norm_g = commons.clip_grad_value_(net_g.parameters(), None)
|
502 |
+
scaler.step(optim_g)
|
503 |
+
scaler.update()
|
504 |
+
|
505 |
+
if rank == 0:
|
506 |
+
if global_step % hps.train.log_interval == 0:
|
507 |
+
lr = optim_g.param_groups[0]["lr"]
|
508 |
+
logger.info(
|
509 |
+
"Train Epoch: {} [{:.0f}%]".format(
|
510 |
+
epoch, 100.0 * batch_idx / len(train_loader)
|
511 |
+
)
|
512 |
+
)
|
513 |
+
# Amor For Tensorboard display
|
514 |
+
if loss_mel > 75:
|
515 |
+
loss_mel = 75
|
516 |
+
if loss_kl > 9:
|
517 |
+
loss_kl = 9
|
518 |
+
|
519 |
+
logger.info([global_step, lr])
|
520 |
+
logger.info(
|
521 |
+
f"loss_disc={loss_disc:.3f}, loss_gen={loss_gen:.3f}, loss_fm={loss_fm:.3f},loss_mel={loss_mel:.3f}, loss_kl={loss_kl:.3f}"
|
522 |
+
)
|
523 |
+
scalar_dict = {
|
524 |
+
"loss/g/total": loss_gen_all,
|
525 |
+
"loss/d/total": loss_disc,
|
526 |
+
"learning_rate": lr,
|
527 |
+
"grad_norm_d": grad_norm_d,
|
528 |
+
"grad_norm_g": grad_norm_g,
|
529 |
+
}
|
530 |
+
scalar_dict.update(
|
531 |
+
{
|
532 |
+
"loss/g/fm": loss_fm,
|
533 |
+
"loss/g/mel": loss_mel,
|
534 |
+
"loss/g/kl": loss_kl,
|
535 |
+
}
|
536 |
+
)
|
537 |
+
|
538 |
+
scalar_dict.update(
|
539 |
+
{"loss/g/{}".format(i): v for i, v in enumerate(losses_gen)}
|
540 |
+
)
|
541 |
+
scalar_dict.update(
|
542 |
+
{"loss/d_r/{}".format(i): v for i, v in enumerate(losses_disc_r)}
|
543 |
+
)
|
544 |
+
scalar_dict.update(
|
545 |
+
{"loss/d_g/{}".format(i): v for i, v in enumerate(losses_disc_g)}
|
546 |
+
)
|
547 |
+
image_dict = {
|
548 |
+
"slice/mel_org": utils.plot_spectrogram_to_numpy(
|
549 |
+
y_mel[0].data.cpu().numpy()
|
550 |
+
),
|
551 |
+
"slice/mel_gen": utils.plot_spectrogram_to_numpy(
|
552 |
+
y_hat_mel[0].data.cpu().numpy()
|
553 |
+
),
|
554 |
+
"all/mel": utils.plot_spectrogram_to_numpy(
|
555 |
+
mel[0].data.cpu().numpy()
|
556 |
+
),
|
557 |
+
}
|
558 |
+
utils.summarize(
|
559 |
+
writer=writer,
|
560 |
+
global_step=global_step,
|
561 |
+
images=image_dict,
|
562 |
+
scalars=scalar_dict,
|
563 |
+
)
|
564 |
+
global_step += 1
|
565 |
+
# /Run steps
|
566 |
+
|
567 |
+
if epoch % hps.save_every_epoch == 0 and rank == 0:
|
568 |
+
if hps.if_latest == 0:
|
569 |
+
utils.save_checkpoint(
|
570 |
+
net_g,
|
571 |
+
optim_g,
|
572 |
+
hps.train.learning_rate,
|
573 |
+
epoch,
|
574 |
+
os.path.join(hps.model_dir, "G_{}.pth".format(global_step)),
|
575 |
+
)
|
576 |
+
utils.save_checkpoint(
|
577 |
+
net_d,
|
578 |
+
optim_d,
|
579 |
+
hps.train.learning_rate,
|
580 |
+
epoch,
|
581 |
+
os.path.join(hps.model_dir, "D_{}.pth".format(global_step)),
|
582 |
+
)
|
583 |
+
else:
|
584 |
+
utils.save_checkpoint(
|
585 |
+
net_g,
|
586 |
+
optim_g,
|
587 |
+
hps.train.learning_rate,
|
588 |
+
epoch,
|
589 |
+
os.path.join(hps.model_dir, "G_{}.pth".format(2333333)),
|
590 |
+
)
|
591 |
+
utils.save_checkpoint(
|
592 |
+
net_d,
|
593 |
+
optim_d,
|
594 |
+
hps.train.learning_rate,
|
595 |
+
epoch,
|
596 |
+
os.path.join(hps.model_dir, "D_{}.pth".format(2333333)),
|
597 |
+
)
|
598 |
+
if rank == 0 and hps.save_every_weights == "1":
|
599 |
+
if hasattr(net_g, "module"):
|
600 |
+
ckpt = net_g.module.state_dict()
|
601 |
+
else:
|
602 |
+
ckpt = net_g.state_dict()
|
603 |
+
logger.info(
|
604 |
+
"saving ckpt %s_e%s:%s"
|
605 |
+
% (
|
606 |
+
hps.name,
|
607 |
+
epoch,
|
608 |
+
savee(
|
609 |
+
ckpt,
|
610 |
+
hps.sample_rate,
|
611 |
+
hps.if_f0,
|
612 |
+
hps.name + "_e%s_s%s" % (epoch, global_step),
|
613 |
+
epoch,
|
614 |
+
hps.version,
|
615 |
+
hps,
|
616 |
+
),
|
617 |
+
)
|
618 |
+
)
|
619 |
+
|
620 |
+
if rank == 0:
|
621 |
+
logger.info("====> Epoch: {} {}".format(epoch, epoch_recorder.record()))
|
622 |
+
if epoch >= hps.total_epoch and rank == 0:
|
623 |
+
logger.info("Training is done. The program is closed.")
|
624 |
+
|
625 |
+
if hasattr(net_g, "module"):
|
626 |
+
ckpt = net_g.module.state_dict()
|
627 |
+
else:
|
628 |
+
ckpt = net_g.state_dict()
|
629 |
+
logger.info(
|
630 |
+
"saving final ckpt:%s"
|
631 |
+
% (
|
632 |
+
savee(
|
633 |
+
ckpt, hps.sample_rate, hps.if_f0, hps.name, epoch, hps.version, hps
|
634 |
+
)
|
635 |
+
)
|
636 |
+
)
|
637 |
+
sleep(1)
|
638 |
+
os._exit(2333333)
|
639 |
+
|
640 |
+
|
641 |
+
if __name__ == "__main__":
|
642 |
+
torch.multiprocessing.set_start_method("spawn")
|
643 |
+
main()
|