Stateless decoder for RNN-T
What does this PR do ?
This PR adds stateless decoder support for RNN-T ASR models.
Collection: [Note which collection this PR will affect]
Changelog
- adds a stateless decoder class for RNN-T ASR models.
- adds another set of asr model tests for partial decoding.
- small bus fixes for asr greedy decode during partial decoding
Usage
To launch a RNN-T model with a stateless decoder, change the training script by adding the following lines with recommended settings,
model.decoder._target_=nemo.collections.asr.modules.RNNTStatelessDecoder \
+model.decoder.context_size=2 \
model.decoder.normalization_mode=layer \
~model.decoder.random_state_sampling \
PR Type:
- [ x ] New Feature
- [ ] Bugfix
- [ ] Documentation
This pull request introduces 1 alert when merging 452dac89710902e0fb16b171c75d6e8831b78ee1 into f921ebe0436e55f7547b183ca83a623f6678422d - view on LGTM.com
new alerts:
- 1 for Unreachable code
Needs formatting of code + some docstrings. Other stuff looks quite good.
Updated PR addressing Som's comments.
This pull request introduces 1 alert when merging 707fa760a854c967fefb801b7e3adb4a0ad6b4cb into 3ba267290927b1a2e78d82c7afbc538874c01bcc - view on LGTM.com
new alerts:
- 1 for Unused import
This pull request introduces 1 alert when merging c695911f046afae9f8e4eb59c617166386ba022c into 4a1fcc62971de89db283334c0dd4aa112e473cbf - view on LGTM.com
new alerts:
- 1 for Unused import
Forgot to mention, lets add a directory for stateless transducers in conf, with subdirectory for Conformer. So that people can more easily find and train such models c