DCGAN/SNGAN architecture with discriminator conditioning

[zhendong-wang]

DCGAN/SNGAN architecture with discriminator conditioning

[kongj5]

Sorry, I don't know your email address, mine is [email protected]

[Zhendong-Wang]

We merge the DCGAN/SNGAN into the stylegan2 framework for training. We found t-embedding doesn't influence much about the performance of DCGAN/SNGAN on CIFAR-10. If you want to add the t_embedding, in your case, following the way we did in stylegan2, the t-embedding is concatenated with the image feature after conv layers as follows,

out = self.model(img)
out = torch.cat((out.view(out.shape[0], -1), t_embedding), dim=1)
validity = self.adv_layer(out)

[ariel415el]

Oh I see.
Is self.adv_layer a binary linear classifier?

[Zhendong-Wang]

Yes, it is a MLP model but not a binary linear classifier. In our case, the discriminator output the logits while not the probability.

gen_logits = self.run_D(gen_img, gen_c, sync=False)
loss_Gmain = torch.nn.functional.softplus(-gen_logits)  # -log(sigmoid(gen_logits))

If you are interested in playing around on DCGAN, you can send me an email. I will share with you the code for that. The code is not cleaned and mighe be a little messy, lol.

[ariel415el]

Thanks, I've sent you a an email to your @columbia.edu address

[Zhendong-Wang]

Sent.

[kongj5]

can you share me a copy of the diffusion training code for DCGAN,

[Zhendong-Wang]

Sure!

[Sarah-2021-scu]

Hey @Zhendong-Wang, could you please share the code of DCGAN, my email is [email protected].
Thanks in advance