Module contrib.classification.transforms

Contains the transforms used in every spot:

  • Training
  • Making figures
  • Web app

For both figures and the webapp, the transform is:

  1. Resize the image so that the shortest size is 512 pixels.
  2. Take the middle 448x448 as a crop

This gives us object-centric images that are 448x448 (fixed pixel size is important for the web app) and not distorted.

Functions

def for_figures()
def for_training(vit_ckpt: str)
def for_webapp(img: PIL.Image.Image) ‑> PIL.Image.Image