Standalone single-image pipeline. It reads only rows whose image_type is
dermoscopic; clinical images and clinical metadata are never loaded.
--data-dir contains the metadata and ground-truth CSV files. Images may be
under that directory or its parent; otherwise pass --input-dir explicitly.
Run from this directory. The split manifest is created by the first run and
reused verbatim by the second… See the full description on the dataset page:
https://huggingface.co/datasets/duyle2408/dermoscopic_backbone_code.