ekalinicheva/plot_vegetation_coverage

Possible inversion of operations in calculation of admissibility loss

Closed this issue · 2 comments

CharlesGaydon commented

plot_vegetation_coverage/model/loss_functions.py

Line 63 in f703325

return ((pred_adm - gt_adm[:, -1]).pow(2) + 0.0001).pow(0.5).mean()

This loss functions looks a lot like a RMSE, except that the root-squaring is performed after averaging instead of before. In this case, pow(2) and pow(0.5) would cancel each other. Is it possible that this operation should look like this instead :

((pred_adm - gt_adm[:, -1]).pow(2) + 0.0001).mean().pow(0.5)

?
This is probably not a big deal since admissibility loss is usually not considered at the moment, but I felt better to share my observation !

ekalinicheva commented

Everything is fine with this formula! It's an absolute loss error and as we add epsilon=0.0001 pow(2) and pow(0.5) do not cancel each other.

CharlesGaydon commented

All right, thanks for claryfing 👍
Closing the issue.