batch proportion instead of batch size

Open topepo opened this issue 5 years ago • 1 comments

This would be much easier to tune

Dec 04 '20 01:12 topepo

I think we usually tune the batch_size to be small as possible while making good estimates of the gradient and that shouldn't depend on the dataset size.

Maybe we could accept both integers and fractions and just interpret (0-1] as fractions and >1 as batch_size?

Dec 04 '20 13:12 dfalbel