mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-08-09 10:02:22 +00:00
Code Issues Packages Projects Releases Wiki Activity

remove cosyvoice scheduler

Browse Source
This commit is contained in:
root 2025-01-20 08:25:30 +00:00
parent 3f8e64b9dc
commit 29de94ee2a
3 changed files with 6 additions and 25 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
egs/wenetspeech4tts/TTS/f5-tts/model/__init__.py
View File

@ -1,7 +0,0 @@
# from f5_tts.model.backbones.dit import DiT
# from f5_tts.model.backbones.mmdit import MMDiT
# from f5_tts.model.backbones.unett import UNetT
# from f5_tts.model.cfm import CFM
# from f5_tts.model.trainer import Trainer
# __all__ = ["CFM", "UNetT", "DiT", "MMDiT", "Trainer"]

2
egs/wenetspeech4tts/TTS/f5-tts/model/cfm.py
View File

@ -290,8 +290,6 @@ class CFM(nn.Module):
# time step
time = torch.rand((batch,), dtype=dtype, device=self.device)
# add cosyvoice cosine scheduler
time = 1 - torch.cos(time * 0.5 * torch.pi)
# TODO. noise_scheduler
# sample xt (φ_t(x) in the paper)

22
egs/wenetspeech4tts/TTS/f5-tts/train.py
View File

@ -998,24 +998,14 @@ def run(rank, world_size, args):
weight_decay=1e-2,
eps=1e-8,
)
elif params.optimizer_name == "Adam":
optimizer = torch.optim.Adam(
model_parameters,
lr=params.base_lr,
betas=(0.9, 0.95),
eps=1e-8,
)
else:
raise NotImplementedError()
if params.decay_steps:
warmup_scheduler = LinearLR(optimizer, start_factor=1e-8, end_factor=1.0, total_iters=params.warmup_steps)
decay_scheduler = LinearLR(optimizer, start_factor=1.0, end_factor=1e-8, total_iters=params.decay_steps)
scheduler = SequentialLR(
optimizer, schedulers=[warmup_scheduler, decay_scheduler], milestones=[params.warmup_steps]
)
assert 1==2
else:
scheduler = Eden(optimizer, 50000, 10, warmup_batches=params.warmup_steps)
warmup_scheduler = LinearLR(optimizer, start_factor=1e-8, end_factor=1.0, total_iters=params.warmup_steps)
decay_scheduler = LinearLR(optimizer, start_factor=1.0, end_factor=1e-8, total_iters=params.decay_steps)
scheduler = SequentialLR(
optimizer, schedulers=[warmup_scheduler, decay_scheduler], milestones=[params.warmup_steps]
)
optimizer.zero_grad()