question for the formular in _diag_fisher() function

[HYJrrr]

Hello, I have read your code and it is really helpful for me to understand the logic of EWC. I still have a doubt for the input of nll_loss in _diag_fisher function. Should the label be the ground truth, not the output of the model? Thanks for your code again and looking forward to your answer!

[moskomule]

Hi, indeed, I don't remember why I did so because it was two years ago I implemented this.

[Albert0147]

Hello, I have read your code and it is really helpful for me to understand the logic of EWC. I still have a doubt for the input of nll_loss in _diag_fisher function. Should the label be the ground truth, not the output of the model? Thanks for your code again and looking forward to your answer!

Hi. I think both way can get the fisher information matrix, while it is called empirical Fisher if you use the provided GT label for computing the matrix.