Hi, there
I am trying to train RankHardLoss on coco dataset.
First, I train a lstmla with random sequence generated by 'get_rand_seq()'.
Then I lock the net and use 'RankHardLoss' provided to train on coco.
Following vse++'s 2-step training strategy, I failed to get better recall then hard negative triplet.
I also tried to unlock the lstmla after prior 2-step training is done, but only got slightly improvement, still less then triplet.
I guess the key point is how to use UpdatingWrapper, but updating every epoch only makes things worse.
Would you please elaborate on the details of training strategy on coco?
Hi, there
I am trying to train RankHardLoss on coco dataset.
First, I train a lstmla with random sequence generated by 'get_rand_seq()'.
Then I lock the net and use 'RankHardLoss' provided to train on coco.
Following vse++'s 2-step training strategy, I failed to get better recall then hard negative triplet.
I also tried to unlock the lstmla after prior 2-step training is done, but only got slightly improvement, still less then triplet.
I guess the key point is how to use UpdatingWrapper, but updating every epoch only makes things worse.
Would you please elaborate on the details of training strategy on coco?