Question about data normalization

[MoeinSorkhei]

Hi, in your implementation data is always normalised with ImageNet stats: https://github.com/thuml/LogME/blob/main/ranking.py#L75

Since the target dataset might be in a different domain than source (e.g aerial images), the stats of target might not necessarily be close to ImageNet stats, and normalizing with ImageNet stats might lead the input not to be exactly in [0-1] that pretrained models were trained on.

With this in mind, do you recommend normalizing the data with ImageNet stats or target dataset stats?

Thanks in advance.