Class weights for weighted cross entropy loss

[Aakash Kumar Nain]

Hi,

I am a little confused about how the `class weights` thing is implemented in the source code. When we pass a dictionary of class weights, as per the source code, first the dictionary is passed to the `_standardize_user_data()` function is called which returns a list of class weights. After this, I wasn't able to get how the mask from this list is generated

and how it is applied to the final cross_entropy function. Can someone please elaborate on this?

On the keras github repo, I found this: https://github.com/keras-team/keras/issues/2115#issuecomment-204060456

 but this seems to penalize the loss in a different way. Can someone please confirm and if possible, provide a little code for the same? 

Regards,

Aakash Nain