I have a multi class classification problem so I am going with Cross entropy I have 12 label: initially categorical labels , but I perform an integer encoding, I was wondering if should I use one hot encoding instead and how that would fit or not my case ? thanks
Hi Oussama!
I can’t think of a good use case in pytorch where one-hot encoding
makes sense. (Of course, if you are using some third-party code
that expects one-hot encoding, then you’ll need it.)
In the case of pytorch’s CrossEntropyLoss
, unless you are using
probabilistic values for target
(the labels), integer (categorical)
class labels are expected.
So, no, don’t use one-hot encoding. Stick with your “integer encoding.”
That’s what CrossEntropyLoss
expects, and it’s a perfectly fine approach.
Best.
K. Frank
1 Like