focal-loss-keras | Binary and Categorical Focal loss implementation | Machine Learning library

I'm as puzzled as you are as to why they multiply the loss by a constant. The code you supplied has alpha as a parameter with the default value set to 0.25, but perhaps when you call the function you're supposed to instead supply a tensor (with the same shape as y_pred/y_true) with the weights to this parameter? That's the only explanation I can think of.

However, I do think you might be able to simply omit those weights in the code altogether, and instead supply your weights to tf.fit()'s class_weight parameter, and then that'll do the weighting for you. Could you keep me updated on whether that works?