Gold score

[hytseng0509]

Hi,

Thanks for sharing the source code!
According to the implementation here, it seems that you directly use the logits predicted from the discriminator as the weight for the re-weighting training. Does it mean that some weights might be negative (thus reversing the direction of the gradient computed from the re-weighted loss here)?

[sangwoomo]

Hi,

Your understanding is exactly right.
My intuition was to suppress the over-generated regions.
However, I did not investigated much variations of the loss; hence, one would find the better one :)

[hytseng0509]

Thanks for the reply!