Add Conformer RNN-T LibriSpeech training recipe #2329
Conversation
hwangjeff
force-pushed
the
librispeech_conformer_rnnt
branch
from
745e36d to
dda2ce3
Compare
hwangjeff
requested review from
mthrok,
xiaohui-zhang,
nateanl and
carolineechen
|
@hwangjeff has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
|
@hwangjeff has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
| ): | ||
| super().__init__() | ||
|
|
||
| self.model = conformer_rnnt_base() |
There was a problem hiding this comment.
I just realized num_symbols is hardcoded as 1024 inside conformer_rnnt_base. IMO conformer_rnnt_base should only be used in test cases. Here we should initiate the model explicitly via conformer_rnnt_model, and pass self.sp_model.get_piece_size() to "num_symbols", and consider exposing more inputs to the input of LightningModule later on.
hwangjeff
force-pushed
the
librispeech_conformer_rnnt
branch
from
6fd974d to
854ece9
Compare
|
@hwangjeff has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
1 similar comment
|
@hwangjeff has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
hwangjeff
force-pushed
the
librispeech_conformer_rnnt
branch
from
da751eb to
3437eef
Compare
|
@hwangjeff has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
|
Hey @hwangjeff. |
Summary: Adds Conformer RNN-T LibriSpeech training recipe to examples directory. Produces 30M-parameter model that achieves the following WER: | | WER | |:-------------------:|-------------:| | test-clean | 0.0310 | | test-other | 0.0805 | | dev-clean | 0.0314 | | dev-other | 0.0827 | Pull Request resolved: pytorch#2329 Reviewed By: xiaohui-zhang Differential Revision: D35578727 Pulled By: hwangjeff fbshipit-source-id: afa9146c5b647727b8605d104d928110a1d3976d
Adds Conformer RNN-T LibriSpeech training recipe to examples directory.
Produces 30M-parameter model that achieves the following WER: