mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-09-07 08:04:18 +00:00
Code Issues Packages Projects Releases Wiki Activity

make biases learn faster in a different way.

Browse Source
This commit is contained in:
Daniel Povey 2022-05-21 17:47:30 +08:00
parent 992f922f81
commit dcab1aee4e
2 changed files with 2 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
egs/librispeech/ASR/pruned_transducer_stateless2/scaling.py
View File

@ -179,7 +179,7 @@ class ScaledLinear(nn.Linear):
with torch.no_grad(): with torch.no_grad():
self.weight[:] *= initial_scale self.weight[:] *= initial_scale
if self.bias is not None: if self.bias is not None:
self.bias[:] *= initial_scale self.bias[:] *= initial_scale * 4.0
def get_weight(self): # not needed any more but kept for back compatibility def get_weight(self): # not needed any more but kept for back compatibility
return self.weight return self.weight

6
egs/librispeech/ASR/pruned_transducer_stateless4/train.py
View File

@ -871,11 +871,7 @@ def run(rank, world_size, args):
logging.info("Using DDP") logging.info("Using DDP")
model = DDP(model, device_ids=[rank]) model = DDP(model, device_ids=[rank])
optimizer = Cain(model.parameters(), lr=params.initial_lr)
params_to_pass = [ {'params': [ p for (name,p) in model.named_parameters() if 'bias' not in name] },
{'params': [ p for (name,p) in model.named_parameters() if 'bias' in name ], 'lr': params.initial_lr*2.0 } ]
optimizer = Cain(params_to_pass, lr=params.initial_lr)
scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)