prepare_parquet.py

import pandas as pd
from pandas_image_methods import PILMethods

from pathlib import Path


SKILLS = {
    "single_skill/counting_only-paired-distance_and_counting": "Counting (w/ Distance)",
    "single_skill/counting_only-paired-position_and_counting": "Counting (w/ Position)",
    "single_skill/distance_only": "Distance",
    "single_skill/position_only": "Position",
    "single_skill/size_only": "Size",
    "combine_2_skill/distance_and_counting": "Distance + Counting",
    "combine_2_skill/distance_and_size": "Distance + Size",
    "combine_2_skill/position_and_counting": "Position + Counting",
    "reasoning/object_manipulation": "Object Manipulation",
    "reasoning/object_occlusion": "Object Occlusion",
}

pd.api.extensions.register_series_accessor("pil")(PILMethods)

for skill in SKILLS.keys():
    print(skill)
    fn_json = f"eval_datasets/coco_test2017_annotations/{skill}.json"
    df = pd.read_json(fn_json)
    df["image"] = df["metadata"].map(
        lambda dict: f"eval_datasets/coco_test2017/{dict['source_img_id']:0>12}.jpg"
    ).pil.open()

    fn_parquet = f"eval_datasets/coco_test2017_annotations_hf/{skill}.parquet"
    Path(fn_parquet).parent.mkdir(parents=True, exist_ok=True)
    df.to_parquet(fn_parquet)