deepgeek/CosyVoice
2
0
Fork 0
mirror of https://github.com/FunAudioLLM/CosyVoice.git synced 2026-02-04 17:39:25 +08:00
Code Issues Packages Projects Releases Wiki Activity

remove unnecessary f0 loss in discrimnator

Browse Source
This commit is contained in:
lyuxiang.lx
2024-10-18 15:58:19 +08:00
parent a4db3db8ed
commit 5bd5dfecab
3 changed files with 27 additions and 27 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
cosyvoice/bin/train.py
View File

@@ -95,6 +95,8 @@ def main():
override_dict.pop('hift') override_dict.pop('hift')
with open(args.config, 'r') as f: with open(args.config, 'r') as f:
configs = load_hyperpyyaml(f, overrides=override_dict) configs = load_hyperpyyaml(f, overrides=override_dict)
if gan is True:
configs['train_conf'] = configs['train_conf_gan']
configs['train_conf'].update(vars(args)) configs['train_conf'].update(vars(args))
# Init env for ddp # Init env for ddp

5
cosyvoice/hifigan/hifigan.py
View File

@@ -64,6 +64,5 @@ class HiFiGan(nn.Module):
loss_tpr = tpr_loss(y_d_rs, y_d_gs, self.tpr_loss_tau) loss_tpr = tpr_loss(y_d_rs, y_d_gs, self.tpr_loss_tau)
else: else:
loss_tpr = torch.zeros(1).to(device) loss_tpr = torch.zeros(1).to(device)
loss_f0 = F.l1_loss(generated_f0, pitch_feat) loss = loss_disc + self.tpr_loss_weight * loss_tpr
loss = loss_disc + self.tpr_loss_weight * loss_tpr + loss_f0 return {'loss': loss, 'loss_disc': loss_disc, 'loss_tpr': loss_tpr}
return {'loss': loss, 'loss_disc': loss_disc, 'loss_tpr': loss_tpr, 'loss_f0': loss_f0}

47
cosyvoice/utils/train_utils.py
View File

@@ -110,30 +110,29 @@ def wrap_cuda_model(args, model):
def init_optimizer_and_scheduler(args, configs, model, gan): def init_optimizer_and_scheduler(args, configs, model, gan):
key = 'train_conf_gan' if gan is True else 'train_conf' if configs['train_conf']['optim'] == 'adam':
if configs[key]['optim'] == 'adam': optimizer = optim.Adam(model.parameters(), **configs['train_conf']['optim_conf'])
optimizer = optim.Adam(model.parameters(), **configs[key]['optim_conf']) elif configs['train_conf']['optim'] == 'adamw':
elif configs[key]['optim'] == 'adamw': optimizer = optim.AdamW(model.parameters(), **configs['train_conf']['optim_conf'])
optimizer = optim.AdamW(model.parameters(), **configs[key]['optim_conf'])
else: else:
raise ValueError("unknown optimizer: " + configs[key]) raise ValueError("unknown optimizer: " + configs['train_conf'])
if configs[key]['scheduler'] == 'warmuplr': if configs['train_conf']['scheduler'] == 'warmuplr':
scheduler_type = WarmupLR scheduler_type = WarmupLR
scheduler = WarmupLR(optimizer, **configs[key]['scheduler_conf']) scheduler = WarmupLR(optimizer, **configs['train_conf']['scheduler_conf'])
elif configs[key]['scheduler'] == 'NoamHoldAnnealing': elif configs['train_conf']['scheduler'] == 'NoamHoldAnnealing':
scheduler_type = NoamHoldAnnealing scheduler_type = NoamHoldAnnealing
scheduler = NoamHoldAnnealing(optimizer, **configs[key]['scheduler_conf']) scheduler = NoamHoldAnnealing(optimizer, **configs['train_conf']['scheduler_conf'])
elif configs[key]['scheduler'] == 'constantlr': elif configs['train_conf']['scheduler'] == 'constantlr':
scheduler_type = ConstantLR scheduler_type = ConstantLR
scheduler = ConstantLR(optimizer) scheduler = ConstantLR(optimizer)
else: else:
raise ValueError("unknown scheduler: " + configs[key]) raise ValueError("unknown scheduler: " + configs['train_conf'])
# use deepspeed optimizer for speedup # use deepspeed optimizer for speedup
if args.train_engine == "deepspeed": if args.train_engine == "deepspeed":
def scheduler(opt): def scheduler(opt):
return scheduler_type(opt, **configs[key]['scheduler_conf']) return scheduler_type(opt, **configs['train_conf']['scheduler_conf'])
model, optimizer, _, scheduler = deepspeed.initialize( model, optimizer, _, scheduler = deepspeed.initialize(
args=args, args=args,
model=model, model=model,
@@ -143,24 +142,24 @@ def init_optimizer_and_scheduler(args, configs, model, gan):
# currently we wrap generator and discriminator in one model, so we cannot use deepspeed # currently we wrap generator and discriminator in one model, so we cannot use deepspeed
if gan is True: if gan is True:
if configs[key]['optim_d'] == 'adam': if configs['train_conf']['optim_d'] == 'adam':
optimizer_d = optim.Adam(model.module.discriminator.parameters(), **configs[key]['optim_conf']) optimizer_d = optim.Adam(model.module.discriminator.parameters(), **configs['train_conf']['optim_conf'])
elif configs[key]['optim_d'] == 'adamw': elif configs['train_conf']['optim_d'] == 'adamw':
optimizer_d = optim.AdamW(model.module.discriminator.parameters(), **configs[key]['optim_conf']) optimizer_d = optim.AdamW(model.module.discriminator.parameters(), **configs['train_conf']['optim_conf'])
else: else:
raise ValueError("unknown optimizer: " + configs[key]) raise ValueError("unknown optimizer: " + configs['train_conf'])
if configs[key]['scheduler_d'] == 'warmuplr': if configs['train_conf']['scheduler_d'] == 'warmuplr':
scheduler_type = WarmupLR scheduler_type = WarmupLR
scheduler_d = WarmupLR(optimizer_d, **configs[key]['scheduler_conf']) scheduler_d = WarmupLR(optimizer_d, **configs['train_conf']['scheduler_conf'])
elif configs[key]['scheduler_d'] == 'NoamHoldAnnealing': elif configs['train_conf']['scheduler_d'] == 'NoamHoldAnnealing':
scheduler_type = NoamHoldAnnealing scheduler_type = NoamHoldAnnealing
scheduler_d = NoamHoldAnnealing(optimizer_d, **configs[key]['scheduler_conf']) scheduler_d = NoamHoldAnnealing(optimizer_d, **configs['train_conf']['scheduler_conf'])
elif configs[key]['scheduler'] == 'constantlr': elif configs['train_conf']['scheduler'] == 'constantlr':
scheduler_type = ConstantLR scheduler_type = ConstantLR
scheduler_d = ConstantLR(optimizer_d) scheduler_d = ConstantLR(optimizer_d)
else: else:
raise ValueError("unknown scheduler: " + configs[key]) raise ValueError("unknown scheduler: " + configs['train_conf'])
else: else:
optimizer_d, scheduler_d = None, None optimizer_d, scheduler_d = None, None
return model, optimizer, scheduler, optimizer_d, scheduler_d return model, optimizer, scheduler, optimizer_d, scheduler_d