How is the model trained when labeled target domain data is available? about aspect_adversarial HOT 4 CLOSED

@QueenT Thanks for the question! In the current implementation I'm not using the labeled target domain data. This is because I didn't do that semi-supervised experiment before. Perhaps I will add that later.

Yes, the unlabeled data is larger than the labeled data. Say there are 2000 unlabeled docs and 1000 labeled docs. The schedule is simply alternating between the labeled and the unlabeled set, and moving the batch pointer of each set (see s_train_ptr and t_ul_batch_ptr). So when the training procedure passes through the unlabeled data once, it will pass through the labeled data twice. Hope it's clear. Thanks!

from aspect_adversarial.

@yuanzh Hi, thanks for your detailed reply. Assuming there are 20,000 unlabeled docs and 1000 labeled docs, will the domain classifier and thus corresponding document encoder part be over-trained? If so, to avoid this situation, is it necessary to drop some unlabeled docs used in the training procedure? Thank you.

from aspect_adversarial.

@QueenT Yes, over-fitting is a possible issue. One option is to directly limit the number of iterations over the labeled data, such as 20 iterations, and this means the unlabeled data are used only once. Another option is to use a different batch size, like 10 for labeled and 200 for unlabeled.

from aspect_adversarial.

@yuanzh Thank you! Now it's much more clear to me.

from aspect_adversarial.