Loren Lugosch

The reason I added `num_sequences` was that some recipes were running out of memory because we were using too many sequences to train the tokenizer. (Maybe the AISHELL recipe?) I...

(poster of the original joke about the acronym here): Will it really require constant clarification and disambiguation, though? I think in any context involving processing variable-length sequences of raw text...

Good call---I'll work on adding that. Also, when you restart training, it loads the previous state but starts at epoch 1, so that needs to be fixed too.

@vineelpratap can you merge this quick change? I put the correct arch cpp link in the notebook

Thanks @patrickvonplaten and @cwkeam for working on porting M-CTC-T! This script by @tlikhomanenko for loading Flashlight tensors into PyTorch might help: https://github.com/flashlight/wav2letter/blob/main/recipes/utilities/convlm_serializer/save_pytorch_model.py I'll look into this as well when I...

Oh, actually that might be specifically for a convolutional model

I ran into this issue too. Deleting the temporary directory worked for me as well.

Thanks!

Hm, unfortunately I'm not sure what your problem might be. What's the exact command you ran?

Oh, I think the problem is you used `L=0.5`. That enables the syndrome loss. I found that the syndrome loss can help FER but hurt BER, so maybe that's the...