salu133445 / lakh-pianoroll-dataset

A collection of 174,154 multi-track piano-rolls

Source Code for Deriving Lakh Pianoroll Dataset (LPD)

The derived dataset using the default settings is available here.

1. Download Lakh MIDI Dataset (LMD) with the following script.

   ./scripts/download_lmd.sh

   (Or, download it manually here.)

2. Set the variables LMD_ROOT and LPD_ROOT in run.sh and variables in config.py to proper values.

3. Derive all subsets and versions of LPD, matched_ids.txt and cleansed_ids.txt with the following script.

   ./scripts/derive_lpd.sh

Derive the labels for the LPD

The derived labels can be found at data/labels.tar.gz.

1. Download the labels with the following script.

   ./scripts/download_labels.sh

2. Derive the labels with the following script.

   ./scripts/derive_labels.sh

Synthesize audio files for the LPD

1. Install GNU Parallel to run the synthesizer in parallel mode.

2. Synthesize audio files from multitrack pianorolls with the following script.

   ./scripts/batch_synthesize.sh ./data/lpd/lpd/lpd_cleansed/ \
     ./data/synthesized/lpd_cleansed 20

   (The above command will synthesize all the multitrack pianorolls in the LPD-cleansed subset with 20 parallel jobs.)