Clarification on loss function names

[bilal2vec]

Hi,

I'm assuming cross_entropy_logits() is meant to be used on the raw logits of the model and sigmoid_cross_entropy_logits() is the version that's meant to be used on logits that have been normalized to [0, 1] with a sigmoid or softmax layer?

Or is that sigmoid_cross_entropy_logits() is meant to be used for binary classification on the raw logits of the model and cross_entropy_logits() is meant for multi-class classification?

[david-berthelot]

Your second interpretation is the right one: sigmoid for binary, cross for multiclass.