TorchSemiSeg icon indicating copy to clipboard operation
TorchSemiSeg copied to clipboard

Published 20 hours ago • verified publisher icon charlesCXK

training with one GPU

Open wo0423 opened this issue 2 years ago • 4 comments

excuse me,i want to know how to train with only one GPU, can you help me? thanks a lot.

wo0423 avatar Oct 17 '22 02:10 wo0423

Hi, we highly recommend using 4 or 8 GPUs to train the model, since the performance is very sensitive to the batch size.

To train with one GPU, you could set a smaller batch size and linearly reduce the learning rate.

charlesCXK avatar Oct 18 '22 12:10 charlesCXK

got it, due to limited conditions, I only have one GPU at my disposal, very appreciate for your patient reply!

wo0423 avatar Oct 18 '22 13:10 wo0423

Hello,i’m sorry to bother you again,there is a problem that has puzzled me for a long time. I really can't figure out where to correct it.I trained with one gpu. I made the following modifications to the code(train.py): image and image and image and image The problem is that my training results are poor.I don't know what the problem is.

wo0423 avatar Nov 02 '22 14:11 wo0423

do you solve this problem?i want to know how to deal with that too.

XieZhiyao21 avatar Apr 01 '24 07:04 XieZhiyao21