[Error with rnnt_loss while training subword_conformer] ValueError: Input tensor 'rnnt_loss/add_4:0' enters the loop with shape (1, None), but has shape (None, None) after one iteration

TensorSpeech / TensorFlowASR

:zap: TensorFlowASR: Almost State-of-the-art Automatic Speech Recognition in Tensorflow 2. Supported languages that can use characters or subwords

https://huylenguyen.com/asr

Apache License 2.0

938 stars 245 forks source link

[Error with rnnt_loss while training subword_conformer] ValueError: Input tensor 'rnnt_loss/add_4:0' enters the loop with shape (1, None), but has shape (None, None) after one iteration #144

Closed mukurgupta closed 3 years ago

mukurgupta commented 3 years ago

Hello I am trying to train subword_conformer with rnnt loss in TF without installing warp-transducer as instructed in the readme but getting the following error. Any help would be appreciated.

Thanks issue0

issue

nglehuy commented 3 years ago

The issue is reproduced. It happens for batch_size=1

nglehuy commented 3 years ago

@mukurgupta Fixed in the commit

mukurgupta commented 3 years ago

Yes, it's working fine now. Thanks a lot, @usimarit for your prompt response and help.