Refactor LibriSpeech Conformer RNN-T recipe #2366
Conversation
| trainer.fit(model) | ||
| model = ConformerRNNTModule(str(args.sp_model_path)) | ||
| data_module = get_data_module(str(args.librispeech_path), str(args.global_stats_path), str(args.sp_model_path)) | ||
| trainer.fit(model, data_module) |
There was a problem hiding this comment.
It'll be great if we add a --resume-checkpoint argument with None as default value, and change this line to
trainer.fit(model, data_module, ckpt_path=args.resume_checkpoint)
There was a problem hiding this comment.
--resume-from-checkpoint may sound better
hwangjeff
force-pushed
the
librispeech_conformer_rnnt_refactor
branch
from
0ec7ce4 to
0078c68
Compare
| from pytorch_lightning import LightningDataModule, seed_everything | ||
|
|
||
|
|
||
| seed_everything(1) |
There was a problem hiding this comment.
Is this the right place to seed? I would imagine this happen once (and only once) at the very beginning of the CLI entry point.
| seed_everything(1) | ||
|
|
||
|
|
||
| def _batch_by_token_count(idx_target_lengths, token_limit, sample_limit=None): |
There was a problem hiding this comment.
Can you change sample_limit to batch_size and token_limit to max_tokens per our previous discussion? It's ok if you want to do that in another PR.
| return dataloader | ||
|
|
||
| def test_dataloader(self): | ||
| dataset = torchaudio.datasets.LIBRISPEECH(self.librispeech_path, url="test-clean") |
There was a problem hiding this comment.
better make the eval split customizable as we discussed. again I can do that in a later PR as well.
| train_transform, | ||
| val_transform, | ||
| test_transform, | ||
| max_token_limit=700, |
There was a problem hiding this comment.
max_tokens should be fine. "max" duplicates with "limit"
| val_transform, | ||
| test_transform, | ||
| max_token_limit=700, | ||
| sample_limit=2, |
|
@hwangjeff has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
|
Hey @hwangjeff. |
Modifies the example LibriSpeech Conformer RNN-T recipe as follows:
partialequivalents (resolves pickling issues in certain runtime environments).