Open SinDongHwan opened 5 years ago
Hi, SinDongHwan, thanks for following my work. To train the generator with the discrminator, you should train with both seq2seq loss and policy gradient (RL) loss. If you only train with RL loss, the generator will easily fall into a terrible local minimum so that you can see the generator only outputs duplicate frame IDs.
Thank you for kind answer.^^ I'm interested in the Video Summarization using Deep Learning.
Could I some questions? Is different your other statement and the following statement?
In a "model_vsum-ptr-gan.ipynb", only the generator trains with seq2seq loss first. And then, the generator with the discriminator trains using policy gradient(RL) loss.
the attached above picture is result in a "model_vsum-ptr_gan.ipynb" final generator model with discriminator.
But when I set is_tr=True, is_pg=True, the results is not good...Is there anything wrong with this?
@dashuzhilin i've gotten same result. i don't know... so i couldn't solve... i'm waiting writer...
i've run code on jupyter-notebook step by step, and didn't edit codes.
when only generator model train, result is good.
however, when generator and discriminator model train, very bad.TT
After Generator&Discriminator train, Generator model evaluates.
How train model, i can get good result.
Thank you^^