midi_dataset.py

import os
import pandas as pd
from datasets import DatasetBuilder, GeneratorBasedBuilder, Split, Value, Features, ClassLabel

class MidiDataset(GeneratorBasedBuilder):
    """Hugging Face Dataset for MIDI files and their labels"""

    def _info(self):
        return DatasetBuilder.info(
            features=Features({
                "file_path": Value("string"),  # MIDI file path
                "label": ClassLabel(names=["1", "2", "3", "4"]),  # 4Q labels as class labels
                "annotator": Value("string")  # Annotator information
            })
        )

    def _split_generators(self, dl_manager):
        """Split the dataset. Assumes all files are local."""
        # Set paths to midis/ and label.csv
        midi_path = "midis"
        label_path = "label.csv"

        return [
            Split(
                name=Split.TRAIN,
                gen_kwargs={
                    "midi_dir": midi_path,
                    "label_file": label_path
                }
            )
        ]

    def _generate_examples(self, midi_dir, label_file):
        """Yield examples from MIDI files and the label.csv"""
        # Read the label.csv into a pandas DataFrame
        df = pd.read_csv(label_file)
        
        for index, row in df.iterrows():
            midi_file = os.path.join(midi_dir, f"{row['ID']}.mid")
            if os.path.exists(midi_file):
                yield index, {
                    "file_path": midi_file,
                    "label": str(row["4Q"]),  # Convert label to string for ClassLabel compatibility
                    "annotator": row["annotator"]
                }

# Usage Example:
# You can now load the dataset using this script as follows:
# from datasets import load_dataset
# dataset = load_dataset("path/to/your/script", split="train")