Giters

santi-pdp / pase

Problem Agnostic Speech Encoder

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Running on multiple GPUs

emirdemirel opened this issue · comments

commented

Thanks for this repository.

I had an issue when trying to run train.py on multiple GPUs. The code gave the following error:

AttributeError: 'DataParallel' object has no attribute 'loss_weight'

It appears that when running on multiple GPU, the model will be wrapped up in DataParallel() object. I did the following modification on the code (pase/models/core.py - line 404), wherever it raised the same issue:
minion.loss_weight --> minion.module.loss_weight

This seems to work fine after this modification.

commented

Can you share the changes you made to make use of multiple GPUs?

commented

i wanna know why i use 2 gpu, and the log show using 2 devices, but the "nvidia-smi" command only show 1 gpu is working, can anybody else use multi-gpu training?