adricl/midi_godzilla_piano_webdataset_1024

MIDI Godzilla Piano Webdataset split into 1024 chunks

Webdataset Midi of the Godzilla MIDI Dataset from Project Los Angeles

This dataset card aims to be a base template for new datasets. It has been generated using this raw template.

Dataset Details

Dataset Description

This dataset has been created by splitting the midi files into 1024 tokens. We then split the traning set into 70% traning 15% validation and 15% test. We augment the midi as per miditok.

        augment_dataset(
            subset_chunks_dir,
            pitch_offsets=[-12, 12],
            velocity_offsets=[-4, 4],
            duration_offsets=[-1, 1]
        )

Then output the files and save into Webdataset

            sample = {
                "__key__": str(midi_path.relative_to(root_dir_path)),
                "id": i,
                "midi_file": midi_data,
            }

The keyname is "mid.midi_file"

Dataset Sources [optional]

• —Repository: Godzilla MIDI Dataset from Project Los Angeles
• —Demo [optional]: Midi Jam Session

Uses

This is used for traning midi based transformers or anything that requires 1024 chunks of midi.

Dataset Structure

[More Information Needed]

Dataset Creation

Curation Rationale

<!-- Motivation for the creation of this dataset. -->

[More Information Needed]

Source Data

Godzilla MIDI Dataset