wendellgul / agah Goto Github PK

Hello，thanks for the code.
I have a question about the learning rate of Adamax in the NUS-WIDE dataset. I am trying to use your code in the NUS-WIDE dataset, but I found that the result is so bad when using the same Adamax optimizer and learning rate as the MIRFLICKR-25K dataset.

Thanks for sharing your code!
I saw your paper "AGAH" using "CHN" as one of your compared methods. Could you tell me if you implement it by yourself? Or do you know where I can find the source code? Could you please share it to me?

could you please tell me how to download the dataset,thanks!

hi.
Have you considered modifying the feature extraction structure of images and text.
Do you think you can use VIT(Visioni transformer) to replace it.

Changing 64bit to 16bit can only reach about 74% map. What is the specific parameters of 16bit reaching 79% map?

Dear Author,
Could you please share the code of generating the dataset file (nus-wide-tc21-iall.mat, for example)? thanks a lot.

Excuse me, when I read your code, I found the loss for the discriminator looks like this:

D_txt_real = -D_txt_real.mean()
D_txt_fake = D_txt_fake.mean()
loss_D_txt = D_txt_real - D_txt_fake

Does this mean you train both real and fake items to the same direction? Or I get it wrong?