tuna2134/sbv2-v2
1
0
Fork 0
Code Issues Pull Requests 1 Actions Packages Projects Releases Wiki Activity

Merge pull request #24 from walledata/fix-resume

Browse Source 
Fix error on resume training
This commit is contained in:
Stardust·减
2023-09-09 00:13:06 +08:00
committed by GitHub
parent 67f4d08520 82409e29ba
commit 1f3ee63a74
1 changed files with 17 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files

38
train_ms.py
View File

@@ -172,24 +172,20 @@ def run():
net_dur_disc = DDP(net_dur_disc, device_ids=[rank], find_unused_parameters=True) net_dur_disc = DDP(net_dur_disc, device_ids=[rank], find_unused_parameters=True)
try: try:
if net_dur_disc is not None: if net_dur_disc is not None:
_, _, _, epoch_str = utils.load_checkpoint( _, _, dur_resume_lr, epoch_str = utils.load_checkpoint(
utils.latest_checkpoint_path(hps.model_dir, "DUR_*.pth"), utils.latest_checkpoint_path(hps.model_dir, "DUR_*.pth"), net_dur_disc, optim_dur_disc,
net_dur_disc, skip_optimizer=True)
optim_dur_disc, _, optim_g, g_resume_lr, epoch_str = utils.load_checkpoint(
skip_optimizer=True, utils.latest_checkpoint_path(hps.model_dir, "G_*.pth"), net_g,
) optim_g, skip_optimizer=True)
_, optim_g, _, epoch_str = utils.load_checkpoint( _, optim_d, d_resume_lr, epoch_str = utils.load_checkpoint(
utils.latest_checkpoint_path(hps.model_dir, "G_*.pth"), utils.latest_checkpoint_path(hps.model_dir, "D_*.pth"), net_d,
net_g, optim_d, skip_optimizer=True)
optim_g, if not optim_g.param_groups[0].get("initial_lr"):
skip_optimizer=True, optim_g.param_groups[0]["initial_lr"] = g_resume_lr
) if not optim_d.param_groups[0].get("initial_lr"):
_, optim_d, _, epoch_str = utils.load_checkpoint( optim_d.param_groups[0]["initial_lr"] = d_resume_lr
utils.latest_checkpoint_path(hps.model_dir, "D_*.pth"),
net_d,
optim_d,
skip_optimizer=True,
)
epoch_str = max(epoch_str, 1) epoch_str = max(epoch_str, 1)
global_step = (epoch_str - 1) * len(train_loader) global_step = (epoch_str - 1) * len(train_loader)
@@ -205,9 +201,9 @@ def run():
optim_d, gamma=hps.train.lr_decay, last_epoch=epoch_str - 2 optim_d, gamma=hps.train.lr_decay, last_epoch=epoch_str - 2
) )
if net_dur_disc is not None: if net_dur_disc is not None:
scheduler_dur_disc = torch.optim.lr_scheduler.ExponentialLR( if not optim_dur_disc.param_groups[0].get("initial_lr"):
optim_dur_disc, gamma=hps.train.lr_decay, last_epoch=epoch_str - 2 optim_dur_disc.param_groups[0]["initial_lr"] = dur_resume_lr
) scheduler_dur_disc = torch.optim.lr_scheduler.ExponentialLR(optim_dur_disc, gamma=hps.train.lr_decay, last_epoch=epoch_str-2)
else: else:
scheduler_dur_disc = None scheduler_dur_disc = None
scaler = GradScaler(enabled=hps.train.fp16_run) scaler = GradScaler(enabled=hps.train.fp16_run)