scores_train/valid take different parameters than custom_scores

[DanChianucci]

custom_scores are defined as

function(target, prediction)

while scores_train and scores_valid are defined as

function(predict, target)

I'm not sure if this was intentional, but it makes it somewhat confusing

[dnouri]

I agree it's confusing and it's mostly an accident. custom_scores was first introduced and uses the scikit-learn metrics interface (an example here), while scores_train and scores_valid use Lasagne's objectives interface where the arguments are predictions, target, so the other way around.