mirrors/icefall
Archived
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Decrease feature_mask_dropout_prob back from 0.2 to 0.15, i.e. revert the 43->48 change.

Browse Source
This commit is contained in:
Daniel Povey 2022-10-02 14:00:36 +08:00
parent cf5f7e5dfd
commit 1be455438a
1 changed files with 1 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
egs/librispeech/ASR/pruned_transducer_stateless7/conformer.py
View File

@ -356,7 +356,7 @@ class ConformerEncoder(nn.Module):
if self.training: if self.training:
# feature mask. # feature mask.
# on 0.25 of the frames, drop out the extra features [force a bottleneck.] # on 0.25 of the frames, drop out the extra features [force a bottleneck.]
feature_mask_dropout_prob = 0.2 feature_mask_dropout_prob = 0.15
feature_unmasked_dim = 256 # hardcode dim for now, 1st 256 are non-masked. feature_unmasked_dim = 256 # hardcode dim for now, 1st 256 are non-masked.
feature_mask = torch.ones_like(src) # S, N, E feature_mask = torch.ones_like(src) # S, N, E