convert_dataset.py

from pathlib import Path

import pandas as pd
from datasets import Dataset, DatasetDict, Image


DATA_ROOT = Path("FBHM")


def create_split(split_name):
    csv_path = DATA_ROOT / f"{split_name}.csv"

    print(f"Loading: {csv_path}")

    df = pd.read_csv(csv_path)

    # Current value:
    # F1/memes/0151.jpg
    #
    # Convert it to:
    # /full/local/path/.../FBHM/F1/memes/0151.jpg
    df["img"] = df["img"].apply(
        lambda x: str((DATA_ROOT / x).resolve())
    )

    # Check whether images actually exist
    missing = [
        p for p in df["img"]
        if not Path(p).exists()
    ]

    if missing:
        print(f"Missing images in {split_name}: {len(missing)}")
        print(missing[:10])
        raise FileNotFoundError("Some images could not be found.")

    dataset = Dataset.from_pandas(
        df,
        preserve_index=False
    )

    # VERY IMPORTANT
    dataset = dataset.cast_column(
        "img",
        Image()
    )

    return dataset


train_dataset = create_split("train")
test_dataset = create_split("test")


dataset = DatasetDict({
    "train": train_dataset,
    "test": test_dataset,
})


print(dataset)

print("\nFeatures:")
print(dataset["train"].features)

print("\nTesting first image:")
print(dataset["train"][0]["img"])


dataset.push_to_hub(
    "nrizwan/FBHM",
    max_shard_size="300MB"
)