For a custom Keras loss function, I need to create a float tensor from a bool tensor. Unfortunately, K.cast() is not differentiable and therefore can't be used. Is there an alternative way to do this that is differentiable?

less_than_tau = y_pred < tau
less_than_tau = K.cast(less_than_tau, 'float32')

CodePudding user response：

Dr. Snoopy is right.

The way you solve for this in deep learning is "soft" functions, such as softmax instead of max.

In your case, if you want to minimize y-pred relative y-tau, you'd do something like

switch = sigmoid(y_pred - y_tau)
loss = switch * true_case   (1. - switch) * false_case