README.md

Music Transformer Script: A ported script from Google Music Transformer notebook

This is a ported script from the original Google Music Transformer notebook. By porting from notebook to script, automating music generation creative process will be much easier. Note that this repo is only for music generation from pre-trained model only, not for training purpose.

Installation:

You need to install Magenta package (support only Python >= 3.5) with correct version:

pip install magenta==1.3.1

If you have any issues regarding installation, you can install via this method:

cd <path_to_this_repo>; pip install -r requirements.txt

Note: Some packages in here are redundant since this is my local environment.

You also need to install google cloud sdk to get Music Transformer pre-trained model on cloud bucket. To get Google Cloud SDK please follow this installation guide.

How to use

Download Music Transformer pre-trained model with Google Cloud SDK:

gsutil -q -m cp -r gs://magentadata/models/music_transformer/checkpoints/* <destination folder>

Unconditional Transformer:

You can generate music without any priming effect by simply type:

python unconditional_sample.py -model_path=path/to/model/checkpoints/unconditional_model_16.ckpt -output_dir=/tmp -decode_length=1024 -num_samples=1

or you can add primer by using primer_path parameter:

python unconditional_sample.py -model_path=path/to/model/checkpoints/unconditional_model_16.ckpt -output_dir=/tmp -decode_length=1024 -primer_path=path/to/primer_mid -num_samples=1

Conditional Transformer:

Generating music conditioned on midi file by typing:

python melody_sample.py -model_path=path/to/model/checkpoints/melody_conditioned_model_16.ckpt -output_dir=/tmp -decode_length=1024 -melody_path=path/to/melody_midi -num_samples=1

Music Generation Automation:

You can also create a whole new music melody by combining SmallMusicVAE to generate your own favorite melody and this Music Transformer to make your melody feel more natural and coherence. This repo also includes a bash script job to do just that:

sh music_generation.sh <path/to/midime_train.py> <path/to/midime_generate.py> <path/to/midime/tmp> <path/to/midime_training_tfrecord> <midime training step> <musicVAE model checkpoint> <number of melody samples to generate> <melody samples output path> <transformer uncoditioned checkpoint> <transformer conditioned checkpoint> <transformer output sample name> <transformer uncoditioned checkpoint> <transformer conditioned checkpoint>  

On how to create tfrecord for training SmallMusicVAE, you need to convert your melody midi file to tfrecord by following the instructions in Building your Dataset.