Giters

etzinis / sudo_rm_rf

Code for SuDoRm-Rf networks for efficient audio source separation. SuDoRm-Rf stands for SUccessive DOwnsampling and Resampling of Multi-Resolution Features which enables a more efficient way of separating sources from mixtures.

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Different batchsize settings lead to different results

guggugg opened this issue · comments

commented

Hello, thank you for sharing. I trained this code on the LibriMix dataset. When the batchsize=1, the loss can drop normally, but when the batchsize=4, the loss is always positive. What could be the problem, can you give me some help?

commented

Thanks for reaching out. What does it mean drops normally? With batch size of 1 the results should be worse since with batch size of 1 there is no augmentation but I do not understand the always positive part.

commented

Thanks for replying to me. I use the trainer [https://github.com/JusperLee/Dual-Path-RNN-Pytorch], and you see, here is the result of batchsize=1
image
and here is the batchsize=4,
image
I don't know why...

commented

If you use other trainers from other repos I can't possibly know what goes wrong.