Custom Criterion

[nina-wiegers]

Hey everyone,

I want to change the criterion for every experience. I used the code from #604 (comment) and changed it a bit. I need different weights for each experience, because I have got a dataset with unbalanced data. In the below, i posted my code. My question is, how can I validate, that this code works. When I print out strategy.criterion in the train loop, I cannot verify, that these loss functions are in use. Has anybody got any idea or had this problem before?

loss_fun_0 = CrossEntropyLoss(weight=exp_0_weights)
loss_fun_1 = CrossEntropyLoss(weight=all_weights)

class ClassWeights(SupervisedPlugin):

  def __init__(self, loss_fun_0, loss_fun_1):
      super().__init__()
      self.loss_fun_0 = loss_fun_0
      self.loss_fun_1 = loss_fun_1

  def criterion(self):
      if self.experience == 0:
          return self.loss_fun_0(self.mb_pred, self.mb_y)
      else:
          return self.loss_fun_1(self.mb_pred, self.mb_y)

[AntonioCarta]

You can use a debugger and put a breakpoint in your function. Alternatively, you can put a print inside your function to check if it is called.