scripts/process-dataset.py

import os
import random
import shutil
from pathlib import Path
import json

import numpy as np

SEED = 42
random.seed(SEED)
np.random.seed(SEED)


if __name__ == "__main__":
    root_indir = Path("./raw-datasets/DigitizePID_Dataset")
    imgs_indir = root_indir / "image_2"
    imgs_in = os.listdir(imgs_indir)

    root_outdir = Path("./processed-datasets/DigitizePID_Dataset")
    # labels_outdir = root_outdir / "labels"

    for split in ("train", "val"):
        (root_outdir / split).mkdir(parents=True, exist_ok=True)
        # (root_outdir / "labels" / split).mkdir(parents=True, exist_ok=True)

    imgs_in = os.listdir(imgs_indir)
    random.shuffle(imgs_in)

    n = len(imgs_in)
    train_end = int(0.8 * n)

    splits = (
        ("train", imgs_in[:train_end]),
        ("val", imgs_in[train_end:]),
    )

    for split, files in splits:
        metadata_lines = []
        for img_fname in files:
            idx = int(Path(img_fname).stem)
            shutil.copy(imgs_indir / img_fname, root_outdir / split / img_fname)

            symbols = np.load(
                root_indir / str(idx) / f"{idx}_symbols.npy", allow_pickle=True
            )

            metadata_lines.append({
                "file_name": img_fname,
                "symbols": {
                    "bbox": [[int(n) for n in symbol[1]] for symbol in symbols],
                    "labels": [int(symbol[2]) for symbol in symbols],
                },
            })

        with open(root_outdir / split / "metadata.jsonl", "w") as f:
            for line in metadata_lines:
                f.write(json.dumps(line) + "\n")