Adaptive with the normalization of [-1, 1]

[cxmscb]

Hi, I found the normalization used for images is [0, 1]. If the normalization of images is [-1, 1], how could I revise the attack codes?

[fra31]

Hi,

I think the easiest solution (assuming PyTorch models) is to use images in [0, 1] and create a wrapper for the models like

class NewModel():
    def __init__(self, model):
        self.model = model

    def __call__(self, x):
        z = x * 2. - 1. # z in [-1, 1] if x in [0, 1]
        return self.model(z)

which takes input in [0, 1], but rescale it to [-1, 1] before inference. Please refer also to #13 for a similar case.

For TensorFlow models, I guess one could do something similar in the model definition, i.e. adding z = x * 2. - 1. on the input to rescale it to [-1, 1].

Let me know if this helps!

[cxmscb]

Thanks!