BCEWithLogitsLosscombines aSigmoidlayer and theBCELossin one single class that makes it more numerically stable. This makes it suitable for a set of predictions (real numbers) from 0 to 1, which represents the probabilities of a class.- you should use this over BCELoss because pytorch’s implementation makes gradients stable: