Effect of randomize_temperature on gFID of RAR

[RohollahHS]

Hi,

Did you conduct any ablation studies on the effect of randomize_temperature on gFID? For instance, if you use a very low randomize_temperature (equivalent to argmax), does the gFID improve or worsen?

[cornettoyu]

Hi,

Thanks for your interests in our work. We have not tried that, but generally speaking, using "argmax" or very low temp should leads to a much worse gFID, since the diversity should be much worse.

[RohollahHS]

Thanks for your reply. I have developed a model whose gFID improves when I use argmax (temperature=0) but worsens when I use a high temperature. For example, I trained my model and the original MaskGIT (from your repo) on 100,000 images of imagenet (with each class containing 100 images). With temperature=0, my model achieves a gFID of about 10, whereas MaskGIT, with temperature=0, gets a gFID of about 50. However, when I use a higher temperature (like the one in the MaskGIT config file), the gFID of MaskGIT improves significantly to about 10, while the gFID of my model worsens to around 15.

I wanted to hear your thoughts on this. Do you think my model is generally better than MaskGIT (since it achieves much better results when temperature=0), or worse (since its results worsen when using a higher temperature, like 1)?

Thanks.