search

imatge-upc / salgan

SalGAN: Visual Saliency Prediction with Generative Adversarial Networks
https://imatge-upc.github.io/salgan
MIT License
368 stars 106 forks source link

Two problems about the discriminator network #19

Closed suntao498253097 closed 6 years ago

suntao498253097 commented 7 years ago

Dear friend, I have two problems about the discriminator network : Firstly, I found that the drop of discriminator network's loss was very slow when I was training salgan. Secondly, I found the model size of discriminator network I trained was different from yours.(5.5MB vs.3.4MB) why? Did you have done some change to your discriminator? Can you give me some suggestion? I am looking forward to your reply.Thanks. Sincerely yours

junting commented 7 years ago

Hi @suntao498253097 ,

The drop of discriminator loss should be very slow, otherwise it can be very unstable. But you can always change the hyperparamers to accelerate the training. I have just make a minor change in the code now. Could you please try if you can use the pretrained model? And also I suggest you to load both model and check the number of elements in it, to make sure they belong to the same model architecture.

Best,