musdb18.py

import datasets
from pathlib import Path
import stempeg
import numpy as np


_DESCRIPTION = """\
    MUSDB18 music source separation dataset

    to open original stem file (mp4), which is done internally you need stempeg library.
    Outcome of mp4 file is a 3 dimensional np_array [n_stems, n_samples, sample_rate].
    
    firt dimension meanings: {
        0: mixture.
        1: drugs,
        2: bass,
        3: others,
        4:vocals,
    }

    Original dataset is not cutted in any parts, but here I cut each song in 10 seconds chunks with 1 sec overlap.
    """

_DESCRIPTION = "musdb dataset"


class Musdb18Dataset(datasets.GeneratorBasedBuilder):
    DEFAULT_WRITER_BATCH_SIZE = 300
    SAMPLING_RATE = 44100
    WINDOW_SIZE = SAMPLING_RATE * 10  # 10s windows
    INSTRUMENT_NAMES = ["mixture", "drums", "bass", "other", "vocals"]

    #! To configure different configurations (length of window is the only thing)
    # use datasets.BuilderConfig

    def _info(self):
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=datasets.Features(
                {
                    "name": datasets.Value("string"),
                    "n_window": datasets.Value("int16"),
                    **{
                        name: datasets.Audio(
                            sampling_rate=self.SAMPLING_RATE, mono=False
                        )
                        for name in self.INSTRUMENT_NAMES
                    },
                    "mean": datasets.Value("float"),
                    "std": datasets.Value("float"),
                }
            ),
        )

    def _split_generators(self, dl_manager):
        #! you must have your folder locally!
        archive_path = dl_manager.download_and_extract(
            "https://zenodo.org/record/1117372/files/musdb18.zip?download=1"
        )

        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={"audio_path": f"{archive_path}/train"},
            ),
            datasets.SplitGenerator(
                name=datasets.Split.TEST,
                gen_kwargs={"audio_path": f"{archive_path}/test"},
            ),
        ]

    def _generate_stem_dict(self, S, song_name, end):
        return {
            name: {
                "path": f"{song_name}/{name}",
                "array": S[i, end - self.WINDOW_SIZE : end, :],
                "sampling_rate": self.SAMPLING_RATE,
            }
            for i, name in enumerate(self.INSTRUMENT_NAMES)
        }

    def _generate_examples(self, audio_path):
        id_ = 0
        for stems_path in Path(audio_path).iterdir():
            song_name = stems_path.stem
            S, sr = stempeg.read_stems(
                str(stems_path), dtype=np.float32, multiprocess=False
            )

            mixture = S.sum(axis=0).T
            assert mixture.shape[0] == 2
            # from (n_instr, n_chann, n_samp) -> (n_chann, n_samp)
            mixture = mixture.mean(0)  # channel_wise mean -> (n_samples,)
            mean = mixture.mean().item()
            std = mixture.std().item()

            for idx, end in enumerate(
                range(self.WINDOW_SIZE, S.shape[1], self.WINDOW_SIZE)
            ):
                yield id_, {
                    "name": song_name,
                    "n_window": idx,
                    **self._generate_stem_dict(S, song_name, end),
                    "mean": mean,
                    "std": std,
                }

                id_ += 1

            # It's very rare for song to have exactly 3 minutes
            yield id_, {
                "name": song_name,
                "n_window": idx + 1,
                **self._generate_stem_dict(S, song_name, end=S.shape[1]),
                "mean": mean,
                "std": std,
            }

            id_ += 1