In going through the 02 jupyter notebook, I find that when ‘item_tfms=Resize(128)’ , the code errors on my linux server with 1080IT. It however works when I change the parameter to ‘item_tfms=Resize(64)’. This is probably related to amount of memory with gpu.
My question is as follows: "Do the item_tfms setting have to be powers of 2? Can I set it at 119 or 101 and how does this affect the training of the model in the code in 02 jupyter notebook? In other words, do the resnet32 pretrained models in fastai need to work with item_tfms=Resize(128) for optimal performance?