Why (input * 2 - 1) in all experiments?

[futakw]

pre-training/robustness/adversarial/train_imagenet.py

Line 127 in 10cd02a

logits = net(adv_bx * 2 - 1)

Hi, I assume this is the input normalization of mean=0.5 and std=0.5.

• Is this a common way for transfer learning?
• If you have tried, what happens if there's no normalization?

[hendrycks]

It's convenient for mapping back and forth between [-1,1], and [0,1].

…

On Mon, Apr 10, 2023 at 2:17 AM futawaseda ***@***.***> wrote: https://github.com/hendrycks/pre-training/blob/10cd02a3c21698dfc790a50fe442fddf417ad17b/robustness/adversarial/train_imagenet.py#L127 Hi, I assume this is the input normalization of mean=0.5 and std=0.5. - Is this a common way for transfer learning? - If you have tried, what happens if there's no normalization? — Reply to this email directly, view it on GitHub <#7>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACZBITU7EQDNSNIUHCZT7EDXAPF3LANCNFSM6AAAAAAWYYZODI> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[futakw]

Thank you for your answer.
Usually, images are mean/std normalized ((x - mean)/std), instead of ((x - 0.5)/0.5) which is yours.
But I assumed that, since mean and std are different for pre-train and target datasets, you simply used 0.5 for mean and std.