Own BCELoss implementation gradients deviate slightly from pytorch version

¯\_(ツ)_/¯
Probably yes!