Finetuning SepFormer : CUDA out of Memory Error

I was trying to finetune SepFormer using the WSJ0Mix recipe and the default config file, but got stuck with RuntimeError: CUDA out of memory.

I have tried 1080 Ti (8G) and Tesla T4 (15 GB), setting batch size to 1 and sample rate to 8000 and audio length to 1 second.

Does finetuneing SepFormer really need that much memory ? Or something else cause this problem?

@mravanelli any ideas on this problem ?

@jerrymark hi~ i seem to meet the same problems when I try to train libriMix dataset with the default yaml file, and i used a V100(32G) GPU. did you solve this problem? any helps will be thanksful~~

Hi guys,

I have the same issue with an A100 with the librimix recipe.