MIDIFoundationModel/data_representation

Dataset Download

Our model supports four different datasets:

• Symbolic Orchestral Database (SOD): Link
• Lakh MIDI Dataset (Clean version): Link
• Pop1k7: Link
• Pop909: Link

Download Instructions

You can download the datasets via the command line:

# SOD
wget https://qsdfo.github.io/LOP/database/SOD.zip

# LakhClean
wget http://hog.ee.columbia.edu/craffel/lmd/clean_midi.tar.gz

For Pop1k7, the official repository link is currently unavailable. However, you can download it from this Google Drive link: Download Pop1k7

For Pop909, the dataset is uploaded in the official Github repository: Repository link

Using Your Own Dataset

If you plan to use your own dataset, you can modify the dataset class in the data_utils.py script under symbolic_encoding folder inside the nested_music_transformer folder. Alternatively, for a simpler approach, rename your dataset to match one of the following options:

• SOD: Use this for score-based MIDI datasets that require finer-grained quantization (supports up to 16th note triplet level quantization; 24 samples per quarter note).
• LakhClean: Suitable for score-based MIDI datasets requiring coarse-grained quantization (supports up to 16th note level quantization; 4 samples per quarter note).
• Pop1k7, Pop909: Ideal for expressive-based MIDI datasets requiring coarse-grained quantization (supports up to 16th note level quantization; 4 samples per quarter note).

Data Representation

This document outlines our standard data processing pipeline. By following the instructions and running the corresponding Python scripts, you can generate a data representation suited to your specific needs.

We focus on symbolic music and limit the use of musical features to a select few. Each feature set size corresponds to specific musical attributes. Through various experiments, we decided to use 7 features for the Pop1k7 and Pop909 datasets, which consist of pop piano music requiring velocity for expression, and 5 features for the Symbolic Orchestral Database (SOD), Lakh MIDI, and SymphonyMIDI datasets.

• 4 features: ["type", "beat", "pitch", "duration"]
• 5 features: ["type", "beat", "instrument", "pitch", "duration"]
• 7 features: ["type", "beat", "chord", "tempo", "pitch", "duration", "velocity"]
• 8 features: ["type", "beat", "chord", "tempo", "instrument", "pitch", "duration", "velocity"]

Parse Argument

• -d, --dataset: This required argument specifies the dataset to be used. It takes one of the following values: "BachChorale", "Pop1k7", "Pop909", "SOD", "LakhClean", or "SymphonyMIDI".

• -e, --encoding: This required argument specifies the encoding scheme to use. It accepts one of the following: "remi", "cp", "nb", or "remi_pos".

• -f, --num_features: This required argument specifies the number of features. It can take one of the following values: 4, 5, 7, or 8.

• -i, --in_dir: This optional argument specifies the input data directory. It defaults to ../dataset/represented_data/corpus/ if not provided.

• -o, --out_dir: This optional argument specifies the output data directory. It defaults to ../dataset/represented_data/events/.

• --debug: This flag enables debug mode when included. No additional value is needed.

1. MIDI to Corpus

In this step, we convert MIDI files into a set of events containing various musical information. The MIDI files should be aligned with the beat and contain accurate time signature information. Place the MIDI files in <nmt/dataset/MIDI_dataset> and refer to the example files provided. Navigate to the <nmt/data_representation> folder and run the script. The converted data will be stored in <nmt/dataset/represented_data/corpus>.

• Example usage: python3 step1_midi2corpus.py --dataset SOD --num_features 5

2. Corpus to Event

We provide three types of representations: REMI, Compound Word (CP), and Note-based Encoding (NB). The converted data will be stored in <nmt/dataset/represented_data/events>.

• Example usage: python3 step2_corpus2event.py --dataset SOD --num_features 5 --encoding nb

3. Creating Vocabulary

This script creates a vocabulary in the <nmt/vocab> folder. The vocabulary includes event-to-index pair information.

• Example usage: python3 step3_creating_vocab.py --dataset SOD --num_features 5 --encoding nb

4. Event to Index

In this step, we convert events into indices for efficient model training. The converted data will be stored in <nmt/dataset/represented_data/tuneidx>.

• Example usage: python3 step4_event2tuneidx.py --dataset SOD --num_features 5 --encoding nb