Why not use RandomSampler instead of SequentialSampler for validation set?

[ucalyptus2]

Hi @endernewton, @KaimingHe great paper.

[netw0rkf10w]

@forkbabu Why do you need the validation to be random?

[ucalyptus2]

@netw0rkf10w because AUC score is 0 if the minibatch has only one type of labels. if labels are 0 and 1 and my minibatch has all 1s, the AUC score is 0.

[netw0rkf10w]

@forkbabu Why do you need to compute AUC for each minibatch?

[ucalyptus2]

@netw0rkf10w just need the running mean of auc from each minibatch as my final auc after every validation epoch

[netw0rkf10w]

I guess you should keep counting (using accumulating sums) the numbers of true positives and false positives, etc. and only compute AUC at the end.

[ucalyptus2]

@netw0rkf10w ok thanks