search

zhiCHEN96 / ConceptWhitening

MIT License
117 stars 28 forks source link

BinaryCrossEntropy instead of CrossEntropy for binary classification with ConceptWhitening #6

Closed irmavdbrandt closed 3 years ago

irmavdbrandt commented 3 years ago

Hi,

First of all, nice method and thank you for publishing the source code!

I noticed that you are performing the ISIC classification task, which is a binary one (malignant or not), with 2 output nodes and following cross-entropy loss instead of 1 output node and binary cross-entropy loss. Do you have a reason for choice or is it just convenience?

I am asking this since I would like to use the CW method in a binary classification problem with one output node.

Thank you!

Best, Irma

zhiCHEN96 commented 3 years ago

Hi Irma,

It's just convenience, using 1 output node and BCE loss should be fine.