Ignore multiple labels in softmax loss

[klo...@vicomtech.org]

Hello,

Is it possible to ignore more than one label when computing the softmax loss using something like:

loss_param {
    normalize: true
    ignore_label: 1
    ignore_label: 2
  }

Thanks,

K

[Przemek D]

Currently the SoftmaxWithLoss layer can only store one index of an ignored class.

[Santiago Olivera]

How about now?

[Przemek D]

I'm sorry, to my knowledge there aren't plans for such a niche feature. Personally, I'd suggest writing a Python layer for this purpose - just run a simple loop over your outputs, and replace the values you want to ignore by some ignored value. If you use numpy (example here) it should be sufficiently fast too.