openspeech icon indicating copy to clipboard operation
openspeech copied to clipboard
verified publisher icon openspeech-team

Cannot support DDP backend

Open Jxu-Thu opened this issue 4 years ago • 6 comments

Thanks for your great codes.

Environment info

  • Python version: 3.8
  • PyTorch version (GPU?): 1.6.0 Tesla V100
  • Using GPU in script?: Yes
  • Version : 0.3.0

Information

I am using Conformer with transducer ...):

Hope for supporting the ddp backend, which is much efficient than do.

Jxu-Thu avatar Oct 24 '21 13:10 Jxu-Thu

I am also very interested in this feature. Do you guys think it is very hard to deal with?

OleguerCanal avatar Oct 29 '21 15:10 OleguerCanal

@hasangchun We'll take a look over the weekend.

sooftware avatar Oct 30 '21 05:10 sooftware

Can you add the pytorch-lightning ddp function? It won't be hard. @hasangchun

sooftware avatar Oct 30 '21 05:10 sooftware

Thanks!

The issue for me is that supporting ddp for various data samplers, which is not quite easy for me. Directly changing hyperparameter for pytorch-lighting is direct but requires modifications on the data samplers.

How to support smartbatch/bucket sampler in a ddp mode?

I am not familiar with the distributed samplers thus seek for helps :) .

Jxu-Thu avatar Oct 30 '21 06:10 Jxu-Thu

Any updates in this topic? I might try to take a look tomorrow otherwise

OleguerCanal avatar Nov 01 '21 21:11 OleguerCanal

Sorry, I was busy 😭

sooftware avatar Nov 02 '21 01:11 sooftware