Multi-label loss function

For multi-label problem (not multi-class)

Should we use

  • Output layer :
    sigmoid for label 1
    sigmoid for label 2
- Then what should be the loss function?
       sum(logistic_regrssion_loss_function(I)) ?       

Also, what will be the TF name for this loss function?



Hi, @Lizhang_Qin!

You are right. You are trying to predict how much some output corresponds to each class. As you said, you are using a sigmoid function for the outputs (0 - 1 range), which means you will need binary crossentropy loss functions for them.

Great, thanks for confirming.