mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-08-09 01:52:41 +00:00
Code Issues Packages Projects Releases Wiki Activity

Scale down pruned loss in warmup mode

Browse Source
This commit is contained in:
Daniel Povey 2022-03-17 16:09:35 +08:00
parent 13db33ffa2
commit acc0eda5b0
1 changed files with 3 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
egs/librispeech/ASR/pruned_transducer_stateless2/train.py
View File

@ -450,7 +450,9 @@ def compute_loss(
lm_scale=params.lm_scale, lm_scale=params.lm_scale,
warmup_mode=warmup_mode, warmup_mode=warmup_mode,
) )
loss = params.simple_loss_scale * simple_loss + pruned_loss loss = params.simple_loss_scale * simple_loss
if not warmup_mode:
loss = loss + pruned_loss * (0.1 if warmup_mode else 1.0)
assert loss.requires_grad == is_training assert loss.requires_grad == is_training