mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-12-09 14:05:33 +00:00
Code Issues Packages Projects Releases Wiki Activity

add tot_score inf mask to make training stable

Browse Source 
I find there are some inf in tot_score, it makes model cannot converge, add inf mask can make training more stable.
This commit is contained in:
qweasdzxcvde 2025-09-08 14:29:57 +08:00 committed by GitHub
parent 34fc1fdf0d
commit 1e3d264b46
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 1 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
icefall/mmi.py
View File

@ -124,6 +124,7 @@ def _compute_mmi_loss_exact_non_optimized(
den_tot_scores = den_lats.get_tot_scores(log_semiring=True, use_double_scores=True) den_tot_scores = den_lats.get_tot_scores(log_semiring=True, use_double_scores=True)
tot_scores = num_tot_scores - den_scale * den_tot_scores tot_scores = num_tot_scores - den_scale * den_tot_scores
tot_scores = tot_scores.masked_fill(torch.isinf(tot_scores), 0.0)
loss = -1 * tot_scores.sum() loss = -1 * tot_scores.sum()
return loss return loss