About ContaGan 2C loss

[WuDiDaBinGe]

this is 2c Loss, when i watch the code, i found the 2c loss code multiply a extra temperature.
denomerator = torch.cat([torch.unsqueeze(inst2proxy_positive, dim=1), instance_zone], dim=1).sum(dim=1)
criterion = -torch.log(temperature * (numerator / denomerator)).mean()

[mingukkang]

Hi,

The additional temperature (t) is multiplied for a balance between adversarial training and class conditioning.

However, as you can see in src/configs, all temperatures in ContraGAN.json are fixed at 1.0.

Sorry for the late reply.

Thank you.

[WuDiDaBinGe]

Thanks for you repling. But i still don't understand.
I don't know why use temperature to balance adversarial training and class conditioning.
gen_acml_loss += self.contrastive_lambda*self.contrastive_criterion(cls_embed_fake, cls_proxies_fake, fake_cls_mask, fake_labels, t, self.margin)
did't this value （contrastive_lambda i ） already balanced them?

[mingukkang]

Hello.

I noticed that multiplying both the temperature and self.contrastive_lambda is redundant.

I will revise the code in the main branch (actually it is already reflected in the renew_cfgs branch (worker.py line 293), which will officially be released at the end of this month).

Thank you.

[WuDiDaBinGe]

OK.Thanks a lot