Piotr Żelasko comments

Results 523 comments of


                                            Piotr Żelasko

Performance degradation when training with jsonl manifests (i.e. DynamicBucketingSampler)

Thanks, this is very interesting. I'm surprised that you are seeing a degradation with `global shuffled` vs `locally `shuffled` version (test-other WER on modified_beam_search 6.43 vs 6.27), it's small but...

Very minor change in alimeeting recipe (revised)

The flac issue could be related to torchaudio. If you set torchaudio.set_audio_backend(“soundfile”) then the issue should be gone; you can also change flac to sth else which would also resolve...

Very minor change in alimeeting recipe (revised)

Hmmm can you show the full stack trace?

Very minor change in alimeeting recipe (revised)

What's your version of lhotse, torch, torchaudio, and webdataset?

Very minor change in alimeeting recipe (revised)

Torchaudio 0.7.2 has an old backend “sox” as default and it doesn’t support the “format” keyword arg. You can fix that by setting eg “torchaudio.set_audio_backend(“soundfile”)”

Very minor change in alimeeting recipe (revised)

Thanks! Would you be willing to make a PR that fixes it?

`cut_id` does not match original `utt_id`

Good point, maybe we should change trim to supervisions behavior in Lhotse to adopt the supervision ID instead. Could you make a PR?

`StopIteration` while trying to resume training from a checkpoint

I’m aware of the issue, it’s on my list as soon as I get a bit more time. Please watch https://github.com/lhotse-speech/lhotse/issues/785

`StopIteration` while trying to resume training from a checkpoint

Unfortunately no, I can't find any time to work on this (but still remember about it). If somebody could help in debugging it that would be great.

`StopIteration` while trying to resume training from a checkpoint

I’ll try to revisit the issue today or tomorrow. In the meantime, if it’s giving you trouble, the easiest workaround for the bug is to not resume the sampler checkpoint....