Try different architectures.
Try with smaller - image size and batch size.
I prefer AWS because I am used to it, but also on AWS I have cuda runtime error - specially with resnext50 or resnet101 - but I dont mind since resnet50 gives good enough results.
@alessa thank you will, will keep your advice. I did contact paperspace and one of their recommendations was to switch to P5000. I might switch over to this, if I keep getting out of memory errors.
Is there any comparison on cost? I know after last month trying to train a model ofr the kaggle furniture competition, i got a bit of trouble at home in regards to my paperspace bill :-/