Support PIL resizse with PIL interpolation to train transformer.
Almost all vision transformer models need using PIL.Image.BICUBIC
as interpolation in resize.