Chord2Melody - Automatic Music Generation AI

日本語

demonstration1 | demonstration2

samples

paper

What is Chord2Melody?

It is an AI that composes music, with MIDI output.

It is based on GPT-2. You can generate music of arbitrary length, and you can also specify the chord progression to generate music.

Or they can compose a continuation of a music they've been working on.

The output music can be used as free content without any copyright or usage restrictions.

Pretrained Models

There are two models that have been trained: the "base_5tr" with 5 output tracks and the "base_17tr" with 17 tracks.

Model Name	output tracks	total number of parameters
base_5tr (backup url)	Drums, Piano, Guitar, Bass, Strings	86167296
base_17tr (backup url)	Drums, Piano, Chromatic Percussion, Organ, Guitar, Bass, Strings, Ensemble, Brass, Reed, Pipe, Synth Lead, Synth Pad, Synth Effects, Ethnic, Percussive, Sound Effects	86941440

Usage

First, clone chord2melody from GitHub.

$ git clone https://github.com/tanreinama/chord2melody
$ cd chord2melody

Then, download and extract the pretrained model from the link above.

$ wget https://www.nama.ne.jp/models/chord2melody-base_5tr.tar.bz2
$ tar xvfj chord2melody-base_5tr.tar.bz2

Launch chord2melody.py with specifying the model , a MIDI file is created.

$ python3 chord2melody.py --model base_5tr

There is no limit to the length of music that can be output. The total number of bars of music to be generated is specified with the "--num_bars" option.

$ python3 chord2melody.py --num_bars 48

Chord to Melody

To specify the chord progression, use the "--chord" option, and use the "--chordbeat" option to specify how many chords to put in a measure.

$ python3 chord2melody.py --chord "C|C|C|C|Dm|Dm|Dm|Dm|G7|G7|G7|G7|Am|Am|Am|Am" --chordbeat 4

Chord" option, you can specify from Available Chord or "auto" connected with "|".

###Compose a continuation of a music

The program "melody2melody.py" will automatically compose a continuation of a music you have been working on. Use the "--input" option in "melody2melody.py" to specify the MIDI file you want to create a continuation of.

$ python3 melody2melody.py --input halfway.mid

Specifies the fluctuation of the melody

By specifying "--top_p", you can specify the fluctuation of the song.

$ python3 chord2melody.py --top_k 25 --top_p 0

Put one or two numbers in "--top_p". The first number is used when there's a chord progression and the second (if specified) when the chord progression is "auto".

Learning Methods

The data for training is from Lakh Pianoroll Dataset. To train, download lpd_5_full.tar.gz or [lpd-17-full.tar.gz](https://drive. google.com/uc?export=download&id=1bJAC2SKhdKbKvpLL1V1l66cCgWX8eQEm) and extract it.

Next, go to the train directory and run "encode.py" to create a training data file.

The "--da" option allows you to specify data augmentation by modulation. The randomly modulated data is used to increase the training data.

$ cd train
$ python3 encode.py --dataset lpd_5 --output lpd_5_dataset

Run "train.py" with specify the type of dataset (lpd_5/lpd_7) in "--dataset" option and the encoded training data file in "--input" option.

$ python3 train.py --dataset lpd_5 --input lpd_5_dataset

Fine Tuning

To fine-tune your own data, you must first edit the data to 5 or 17 tracks of MIDI data.

Then save the data in pypianoroll format, with the tracks in the same order as the original model.

Then you can create a training data file in encoder.py and fine tune it by specifying the original trained model in the "--restore_from" field.

$ python3 train.py --dataset lpd_5 --input lpd_5_dataset --restore_from ../base_5tr