Discount parameter in the loss layer

[Corey Nolet]

I have a custom loss layer that discounts the loss for a specific class (because of class imbalance) and decreases the discount over the entire training period (until, eventually, there is no discount at all).

How would I decrement this value from epoch to epoch? I'm specifying my loss function like this:

def fcrn_loss(y_true, y_pred):
  ....

Thanks!

[Klemen Grm]

You could do it with a global keras variable that you change at the end of every training epoch, eg. with a custom callback.

[Corey Nolet]

I tried your solution with a Theano shared variable that I update in a custom callback and it looks like it's working. Thanks!