MMDS: A general-purpose multimodal dataset wrapper

This project is under construction, API may change from time to time.

Installation

Stable (not stable yet though)

pip install mmds

Latest

pip install mmds --pre

Example Usage

# example.py

import timeit
from mmds import MultimodalDataset, MultimodalSample
from mmds.modalities import RgbsModality, WavModality, MelModality, F0Modality
from mmds.utils.spectrogram import LogMelSpectrogram
from pathlib import Path
from multiprocessing import Manager


try:
    import youtube_dl
    import ffmpeg
    from torchvision import transforms
except:
    raise ImportError(
        "This demo requires youtube_dl, ffmpeg-python and torchvision, "
        "install them now: pip install youtube_dl ffmpeg-python torchvision"
    )


def download():
    Path("data").mkdir(exist_ok=True)

    ydl_opts = {
        "postprocessors": [
            {
                "key": "FFmpegExtractAudio",
                "preferredcodec": "mp3",
                "preferredquality": "192",
            }
        ],
        "postprocessor_args": ["-ar", "16000"],
        "outtmpl": "data/%(id)s.%(ext)s",
        "keepvideo": True,
    }
    with youtube_dl.YoutubeDL(ydl_opts) as ydl:
        ydl.download(["https://www.youtube.com/watch?v=BaW_jenozKc"])

    path = Path("data/BaW_jenozKc")

    if not path.exists():
        path.mkdir(exist_ok=True)

        (
            ffmpeg.input("data/BaW_jenozKc.mp4")
            .filter("fps", fps="25")
            .output("data/BaW_jenozKc/%06d.png", start_number=0)
            .overwrite_output()
            .run(quiet=True)
        )


class MyMultimodalSample(MultimodalSample):
    def generate_info(self):
        wav_modality = self.get_modality_by_name("wav")
        rgbs_modality = self.get_modality_by_name("rgbs")
        return dict(
            t0=0,
            t1=wav_modality.duration / 10,
            original_wav_seconds=wav_modality.duration,
            original_rgbs_seconds=rgbs_modality.duration,
        )


class MyMultimodalDataset(MultimodalDataset):
    Sample = MyMultimodalSample


def main():
    download()

    # optional multiprocessing cache manager
    manager = Manager()

    dataset = MyMultimodalDataset(
        ["BaW_jenozKc"],
        modality_factories=[
            RgbsModality.create_factory(
                name="rgbs",
                root="data",
                suffix="*.png",
                sample_rate=25,
                transform=transforms.Compose(
                    [
                        transforms.Resize((28, 28)),
                        transforms.ToTensor(),
                        transforms.Normalize(0.5, 1),
                    ],
                ),
                cache=manager.dict(),
            ),
            WavModality.create_factory(
                name="wav",
                root="data",
                suffix=".mp3",
                sample_rate=16_000,
                cache=manager.dict(),
            ),
            MelModality.create_factory(
                name="mel",
                root="data",
                suffix=".mel.npz",
                mel_fn=LogMelSpectrogram(sample_rate=16_000),
                base_modality_name="wav",
                cache=manager.dict(),
            ),
            F0Modality.create_factory(
                name="f0",
                root="data",
                suffix=".f0.npz",
                mel_fn=LogMelSpectrogram(sample_rate=16_000),
                base_modality_name="wav",
                cache=manager.dict(),
            ),
        ],
    )

    # first load
    print(timeit.timeit(lambda: dataset[0], number=1))

    # second load
    print(timeit.timeit(lambda: dataset[0], number=1))

    print(dataset[0])


if __name__ == "__main__":
    main()