Regarding the preprocessing steps for BiT and ViT models inside timm

[sayakpaul]

As far as I know the BiT and ViT (prefixed with jx) models were converted from the official checkpoints released by Google. So, I wanted to know do BiT and ViT follow the same preprocessing steps as the official implementations do during inference? Generally, we use the respective mean and std to carry out normalization but that might not be the case here.

For example, BiT just scales the pixel values to [0, 1] range (source). ViT scales to [-1, 1] (source).

A clarification would be helpful.

[rwightman]

@sayakpaul the code here is correct wrt to preprocessing, I matched what the official code impl did. Any discrepencies with the paper is a separate issue. Both are the same [-1, 1]

[sayakpaul]

I see. Shouldn't the BiT ones be [0, 1]?