Getting error when num_workers > 0

[amilamad]

Hi,
I have tried to train the lj speech model with latest mater and it gives me error like this, with

num_workers = 2

It looks like _frontend for worker processes didn`t got assigned. I tried injecting _frontend object to the TextDataSource. But It failed. Is there a fix for this ?

When I set the num_workers = 0 , it is training ok.
After quick google search it tells me that when num_workers = 0 it will do all the work in main thread.
My question is, will it slow down my training process significantly ?

[r9y9]

Seems dup of #37.

[r9y9]

This is not really tested on Windows, so I'd recommend Linux instead if possible.

_frontend should be deepvoice3_pytorch.frontend.en for English datasets, not TextDataSource. I'm not sure wha's happening in the num_workers=0 case.

[amilamad]

Oh sorry again for duplicate issue.
_frontend is initialized correctly with the deepvoice3_pytorch.frontend.en in the main python thread. But when it is accessed by worker threads _frontend is None .I think this issue is due to the nature of the python worker threads on windows.
Ill place a issue on the https://github.com/peterjc123/pytorch-scripts

Now my training is running ok with num_workers=0. Can I know will it significantly slow down compared to num_workers=2 case ?

[engiecat]

@amilamad
I am currently working with VCTK set on Win 10 and num_workers=1 does work well. (About 10% loss of performance is observed.)

As seen in the screenshot, THAllocator error does occur(approx. once per day for me) though, but according to pytorch/pytorch#4239, it can be alleviated by setting lower num_workers.
(for me, careful monitoring and restarting the training from checkpoint worked)