Custom weighted NLL_loss

1

Hello,
I want to develop my own NLL_loss which will penalize certain errors by their severity (I want some errors to be punished more than others).
To this end I need to alter the source code of pytorch’s NLL_loss, but it is written in C and I can’t seem to find it.
I found an implementation of CrossEntropyLoss here:
https://gist.github.com/mjdietzx/50d3c26f1fd543f1808ffffacc987cbf
and when I test it against the CrossEntropyLoss in Pytorch it gives the same output, although when I use weighted samples the output is different, thus my starting point for altering the code is bad.

My main points are:

  1. Is there any out of the box functionality in Pytorch to penalize certain mistakes over others?
  2. If not, I can implement this myself, but I don’t understand how NLL_loss in Pytorch deals with sample weights, which prevents me from achieving my first goal of creating the exact same loss function with my own tweakable implementation.

Thanks all.

2

This is the source code (and 2D version).

The weights in NLLLoss are used to compute the weighted average (one weight per target class). I think the docs explain it clearly.