SigLIP image processing flow

[eternalding]

Greetings.

In SigLIP preprocess, the transform flow is defined in: link

For to_channel_dimension_format, the argument channel_dim and input_channel_dim are both set as self.data_format, so no matter it's set as channel_first or channel_last, it always returns the original image. Not sure if this is the correct behavior also.

Please correct me if I was wrong.
Thanks.

YC