mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-08-10 10:32:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

Reduce layer-drop prob after warmup to 1 in 100

Browse Source
This commit is contained in:
Daniel Povey 2022-03-27 00:25:32 +08:00
parent b43468bb67
commit 953aecf5e3
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_stateless2/conformer.py
View File

@ -231,7 +231,7 @@ class ConformerEncoderLayer(nn.Module):
# compensate for the small scale by just producing larger output. # compensate for the small scale by just producing larger output.
warmup = max(warmup, 0.1) warmup = max(warmup, 0.1)
if self.training: if self.training:
warmup = min(warmup, 0.98) # effectively, layer-drop with 1-in-50 prob. warmup = min(warmup, 0.99) # effectively, layer-drop with 1-in-100 prob.
alpha = 1.0 if torch.rand(()).item() <= warmup else 0.1 alpha = 1.0 if torch.rand(()).item() <= warmup else 0.1
# macaron style feed forward module # macaron style feed forward module