Giorgos Paraskevopoulos

Hi Scott, Great to hear that there is a plan to support tf 2.0 and pytorch.

hi @lhl2017 unfortunately no. I ended up using other alternatives like the reformer. You can check out a recent implementation of block sparse in pytorch available though https://github.com/ptillet/torch-blocksparse

Better make it 13 as a sane default

Yes that's by design. Imagine the case where actually the user doesn't have a backup of the model and the script goes on to delete it. Its better be annoying...

TODO: Add a flag to remove old models during model building

What is the max workers value you pass? (https://github.com/georgepar/kaldi-grpc-server/blob/main/Makefile#L24) 10-15gb ram sounds a bit excessive

TODO: Specify memory requirements as a function of num-workers and model size

You can check the ivector_extractor folder, produced by kaldi during DNN training. If you use that and the mfcc_hires.conf it should work

Hi, I could try to help you out a bit, but I would need more info. For example 1. I assume you have deployed a DNN-HMM model. Does the wav...

Though the error you posted indicates that the client fails to connect to the server.