gt-ripl / geometric-sensitivity-decomposition Goto Github PK

Dear,

I'd like to know why the a and b parameters need to go through the sigmoid and softplus functions.
Because when I use this for training, the loss doesn't decrease and it doesn't help calibration.
If sigmoid and softplus are not used, it can improve calibration, and the result is close to temperature scaling in in-distribution data. (I think it is correct that the result in in-distribution data is close to temperature scaling)

Thanks,
Yung-Chen