jacopoteneggi/IBYDMT

Pre-computed CLIP embeddings

Embeddings are stored as HDF5 datasets with the following structure:

<DATASET_NAME>_<MODEL_NAME>_<OP>.hdf5

"""
    DATASET_NAME:   name of the dataset, e.g. "imagenette".
    MODEL_NAME:     name of the model, e.g. "open_clip:ViT-B-32".
    OP:             split of the dataset (either "train" or "val").
"""

dataset["embedding"]    contains the embeddings
dataset["label"]        contains the labels

To generate the dataset, run

$ python make_dataset.py -h
usage: make_dataset.py [-h] [--dataset DATASET [DATASET ...]] [--model MODEL [MODEL ...]]

options:
  --dataset DATASET [DATASET ...]   List of datasets to encode.
  --model MODEL [MODEL ...]         List of models to use.

Supported dataset names (see supported_datasets.txt):

• imagenette [dataset]

Supported model names (see supported_models.txt):

• open_clip:ViT-B-32 [model]
• open_clip:ViT-L-14 [model]
• clip:ViT-B/32 [model]
• clip:ViT-L/14 [model]

References

@misc{teneggi2024ibetdidmean,
  title={I Bet You Did Not Mean That: Testing Semantic Importance via Betting}, 
  author={Jacopo Teneggi and Jeremias Sulam},
  year={2024},
  eprint={2405.19146},
  archivePrefix={arXiv},
  primaryClass={stat.ML},
}