yhhhli / BRECQ

Pytorch implementation of BRECQ, ICLR 2021
MIT License
247 stars 56 forks source link

How to deal with data parallel and distributed data parallel? #13

Closed jang0977 closed 3 years ago

jang0977 commented 3 years ago

On my eyes, your code is just running with single gpu while I need to test this code with multi-gpu for other implementations. I just want to check that you have ran your code using data parallel and distributed data parallel.

yhhhli commented 3 years ago

Good question, you cannot warp the model with data parallel functions To support multi-GPU reconstruction, you should initialize distributed environments and allreduce the gradients manually. I will add the dist mode in a few days.

xiayizhan2017 commented 2 years ago

@yhhhli hi,multi-GPU reconstruction does not reduce the time consumption. Is there any way to speed up reconstruction.