About temperature

[Muzhi Zhu]

Hi,
When I adjust the temperature, I observe a change in the relationship between the distributions. Specifically, when temperature=0.1, E(logitsA) is initially greater than E(logitsB), but when temperature=10, E(logitsA) becomes less than E(logitsB).

I use temp * torch.logsumexp(logits[:, :num_category] / temp, dim=1) to computer energy E()

logits = torch.randn(7, num_category)  
energy = temp * torch.logsumexp(logits[:, :num_category] / temp, dim=1)

Am I misunderstanding or implementing it incorrectly? I understand that temperature is only for smoothing and should not change the relative relationships, right?

[wetliu]

It is not just smoothing. The figure with temperature changes vs the FPR95 shows the relative different does change.