Refactor Emformer RNNT recipes #2212
Conversation
hwangjeff
force-pushed
the
recipe_refactor
branch
from
871d093 to
bb41509
Compare
|
@hwangjeff has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
Summary: Consolidates LibriSpeech and TED-LIUM Release 3 Emformer RNN-T training recipes in a single directory. Pull Request resolved: pytorch#2212 Differential Revision: D34120104 Pulled By: hwangjeff fbshipit-source-id: 38cf6453d19e74f9851ff0544c6c7f604ec5b630
|
This pull request was exported from Phabricator. Differential Revision: D34120104 |
hwangjeff
force-pushed
the
recipe_refactor
branch
from
bb41509 to
be42df8
Compare
|
|
||
| ## Model Types | ||
|
|
||
| Currently, we have training recipes for the LibriSpeech and TED-LIUM Release 3 datasets. |
There was a problem hiding this comment.
Could you have a brief description on what’s the difference between these models? I know vocab size is one. Is there any other?
There was a problem hiding this comment.
yeah, that's the only difference in the model architecture. the recipes themselves have some differences that are specific to the datasets — this should be apparent in the lightning modules
hwangjeff
force-pushed
the
recipe_refactor
branch
from
be42df8 to
d7564fe
Compare
|
@hwangjeff has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
Summary: Consolidates LibriSpeech and TED-LIUM Release 3 Emformer RNN-T training recipes in a single directory. Pull Request resolved: pytorch#2212 Reviewed By: mthrok Differential Revision: D34120104 Pulled By: hwangjeff fbshipit-source-id: 60deed45731954d5f9b7df73a35d8373184c3c67
|
This pull request was exported from Phabricator. Differential Revision: D34120104 |
hwangjeff
force-pushed
the
recipe_refactor
branch
from
d7564fe to
3fbbf34
Compare
Summary: Consolidates LibriSpeech and TED-LIUM Release 3 Emformer RNN-T training recipes in a single directory. Pull Request resolved: pytorch#2212 Reviewed By: mthrok Differential Revision: D34120104 Pulled By: hwangjeff fbshipit-source-id: 9ed0a5d5b209478a841324f360c5b66268e3b228
hwangjeff
force-pushed
the
recipe_refactor
branch
from
3fbbf34 to
ebda0ec
Compare
|
This pull request was exported from Phabricator. Differential Revision: D34120104 |
|
|
||
| assert len(idx_target_lengths) > 0 | ||
|
|
||
| idx_target_lengths = sorted(idx_target_lengths, key=lambda x: x[1]) | ||
|
|
||
| assert max_token_limit >= idx_target_lengths[-1][1] | ||
|
|
||
| self.batches = _batch_by_token_count(idx_target_lengths, max_token_limit) | ||
| self.batches = batch_by_token_count(idx_target_lengths, max_token_limit)[:100] |
There was a problem hiding this comment.
Why we choose the first 100 batches here?
There was a problem hiding this comment.
good call — this was meant to speed up the training epoch for testing purposes only and should be removed
|
Do we want to keep the |
@nateanl The difference between librispeech model and tedluim model should be clearly outlined in the top level README. It can be as simple as |
| default=pathlib.Path("global_stats.json"), | ||
| type=pathlib.Path, | ||
| help="Path to JSON file containing feature means and stddevs.", | ||
| required=True, |
There was a problem hiding this comment.
As it has a default value, is it really necessary to specify required=True here?
Also, the default value is not correct as that file is inside the recipe dir.
I would suggest removing the default value.
Summary: Consolidates LibriSpeech and TED-LIUM Release 3 Emformer RNN-T training recipes in a single directory. Pull Request resolved: pytorch#2212 Reviewed By: mthrok Differential Revision: D34120104 Pulled By: hwangjeff fbshipit-source-id: 29c6e27195d5998f76d67c35b718110e73529456
Consolidates LibriSpeech and TED-LIUM Release 3 Emformer RNN-T training recipes in a single directory.