wjun0830 / localizable-rotation Goto Github PK

Hello,
I want to ask about the implementation of the PGD attack. From the original implementation of PGD attack that you mention in the paper, calculating the perturbed image uses the gradient of the loss from classifier and auxilliary rotation. Is the Lorot loss value used in PGD or does it only use the loss from the classifier? Then, what is the ratio of the Lorot loss used in PGD?

Thank you

You've perfectly integrated self-supervision into supervised learning through multi-task learning, in a similar way to SLA! But I was not clear about your experimental setup when I reproduced Table 1 (CIFAR, Classification), so if you would like to provide a version of the implementation or some details, it would be appreciated!

Hello, thank you very much for your great work!
I have a doubtful question that I would like to ask you. In the ood_lorot-E file, you use idx to control which part of the image is rotated (a total of four parts), and use idx2 to control the angle of rotation, right? After doing so it will be divided into 16 categories. But I think when the value of idx2 is 0, no matter what idx is, the picture will not be rotated, right? In other words, there are 4 situations in which rotation does not occur. So how does the model distinguish between these 4 situations? I hope you can answer my questions despite your busy schedule. Thank you very much!