Loss_adv_p increasing, while loss_D decreasing. Is it correct? about advsemiseg HOT 6 CLOSED

loss_adv_p is essentially the inverse loss of loss_D. We expect them to achieve a good balance during the adversarial training. Please refer to equation (1) and (4) in our paper. We use lambda_adv to control the balance point between loss_adv_p and loss_D. Does that answer your question?

from advsemiseg.

Great. You have mentioned that your task tries to minimize loss_D and maximize loss_adv_p. However, when we maximize loss_adv_p, the loss_seg_total = loss_seg+ lambdax loss_adv_p in ep (2) tends to increase. But we want to minimize the eq (2). Does it make sense?

One more thing, your implementation use
You used

for param in model_D.parameters(): 
     param.requires_grad = False 
loss = loss_seg + args.lambda_adv_pred * loss_adv_pred 
loss.backward() 

Does it same as the below?

loss = loss_seg + args.lambda_adv_pred * loss_adv_pred 
loss_seg.backward() 

By delete the first two lines and change loss.backward() to loss_seg.backward()because loss_adv_pred does not do backward() so D not trained.

from advsemiseg.

from advsemiseg.

Thanks. I have updated something in the above question related to implementation. Could you also look at it? Thanks for your answer

from advsemiseg.

No. It's not equivalent. By doing that you are not performing the adversarial training since you are not minimizing loss_adv.

from advsemiseg.

@hfslyc : Thanks. I have one more question. In my case, the both loss_adv_p and loss_D are decreasing. What does happen in my case? Do we expect that loss_adv_p increases and loss_D decrease?

from advsemiseg.