Problem with Geman-McClure loss implementation

[JokerYan]

I have been testing the Geman-McClure loss implemented in the repo, more specifically the model_utils.general_loss_with_squared_residual() with alpha = -2.

However, the output, as shown below, is very different from my expectation. (alpha = -2, scale = 0.1) Shouldn't it be like the quadratic shape when |x| < scale

[keunhong]

Did you square the residual before passing it to the function?

[JokerYan]

No, I didn't. I think that is the problem. Thank you very much!