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

Revert optim schedule

Browse Source
This commit is contained in:
Daniel Povey 2023-05-18 15:35:23 +08:00
parent 299482d02d
commit cdfa388ac0
2 changed files with 2 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
egs/libriheavy/LM/zipformer1/subformer.py
View File

@ -875,16 +875,7 @@ class LearnedDownsamplingModule(nn.Module):
if random.random() < 0.01 or __name__ == '__main__': if random.random() < 0.01 or __name__ == '__main__':
logging.info(f"mean weight={weights.mean()}, mean-abs-scores={scores.abs().mean()} positive-scores={(scores>0).to(torch.float32).mean()}, discarded-weights={weights_discarded.mean()}, seq_len={seq_len}, seq_len_reduced={seq_len_reduced}") logging.info(f"mean weight={weights.mean()}, mean-abs-scores={scores.abs().mean()} positive-scores={(scores>0).to(torch.float32).mean()}, discarded-weights={weights_discarded.mean()}, seq_len={seq_len}, seq_len_reduced={seq_len_reduced}")
weights_discarded = weights_discarded.flip(dims=1)
# randomly rotate `weights_discarded` on the sequence axis; this is
# intended to ensure that it doesn't assign the highest scores to
# not-so-important elements to avoid the randomness of these
# discarded weights.
r = random.randint(0, seq_len_reduced - 1)
weights_discarded = torch.cat((weights_discarded[:, r:],
weights_discarded[:, :r]),
dim=1)
weights = (weights[:, :seq_len_reduced] - weights_discarded) weights = (weights[:, :seq_len_reduced] - weights_discarded)
else: else:

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

@ -882,7 +882,7 @@ class Eden(LRScheduler):
warmup_factor = ( warmup_factor = (
1.0 1.0
if self.batch >= self.warmup_batches if self.batch >= self.warmup_batches
else 0.1 + 0.9 * (self.batch / self.warmup_batches) else 0.5 + 0.5 * (self.batch / self.warmup_batches)
) )
return [x * factor * warmup_factor for x in self.base_lrs] return [x * factor * warmup_factor for x in self.base_lrs]