[FEATURE] About loss weights

[523997931]

for train data,
sometimes, I met the long tailed dataset, So I want to add the loss weight in loss function,
it can deal with some data imblance problem

[travishsu]

Hi Authors, I would also like to see this feature, is there any plan about this?

[rwightman]

There are countless possible customizations to loss, metrics, etc... the core train scripts are focused around reproducing imagenet training, which is quite balanced compared to many real world datasets. The scripts are very hackable, with minimal layers of abstraction, so it's easy to swap in different loss fn, add your own metrics, add args for weighting, etc for your given task