mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-12-11 06:55:27 +00:00
Code Issues Packages Projects Releases Wiki Activity

Set all aux-loss probs to zero.

Browse Source
This commit is contained in:
Daniel Povey 2022-12-13 19:25:08 +08:00
parent 52d18e405e
commit 57040e382a
1 changed files with 5 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py
View File

@ -361,9 +361,9 @@ def _whitening_schedule(x: float, ratio: float = 2.0) -> ScheduledFloat:
def _aux_grad_scale() -> float: def _aux_grad_scale() -> float:
return 0.2 return 0.2
def _aux_grad_prob_out() -> ScheduledFloat: def _aux_grad_prob_out() -> ScheduledFloat:
return ScheduledFloat((0.0, 0.25), (1000.0, 0.05), (8000.0, 0.0125)) return 0.0 # ScheduledFloat((0.0, 0.25), (1000.0, 0.05), (8000.0, 0.0125))
def _aux_grad_prob_in() -> ScheduledFloat: def _aux_grad_prob_in() -> ScheduledFloat:
return ScheduledFloat((0.0, 0.25), (1000.0, 0.0)) return 0.0 # ScheduledFloat((0.0, 0.25), (1000.0, 0.0))
#return ScheduledFloat((0.0, 0.25), (1000.0, 0.05), (8000.0, 0.0125)) #return ScheduledFloat((0.0, 0.25), (1000.0, 0.05), (8000.0, 0.0125))
@ -1356,7 +1356,9 @@ class AttentionSqueeze(nn.Module):
aux_grad_scale=_aux_grad_scale(), prob=_aux_grad_prob_in()) aux_grad_scale=_aux_grad_scale(), prob=_aux_grad_prob_in())
self.to_bottleneck_proj = LinearWithAuxLoss(embed_dim, self.to_bottleneck_proj = LinearWithAuxLoss(embed_dim,
bottleneck_dim) bottleneck_dim,
aux_grad_scale=_aux_grad_scale(),
prob=_aux_grad_prob_in())
# bottleneck_balancer is before the actiation. Mostly, for well-trained # bottleneck_balancer is before the actiation. Mostly, for well-trained