Patch-level labels with SAM

[soniajoseph]

• Given patchxdimension activations and segmentation map, return more granular labels
• High-level design decisions need to be made about whether to compute our own segmentation maps, use pre-computed datasets, and returning hierarchical labels (e.g. eye -> face -> person).

[soniajoseph]

@themachinefan

[soniajoseph]

Done here: https://huggingface.co/datasets/Prisma-Multimodal/segmented-imagenet1k-subset