SmallMusicVAE is simply a MidiMe model from Google open source Magenta but implemented in Python instead of JavaScript from original model. This model learns a smaller latent space from MusicVAE to encapsulate the general characteristic of a song. Thus, it can be trained faster and more energy efficient. For more detail behind the model, please take a look at this paper.

Additionally, this repo also includes LCMusicVAE which is a latent constraint MusicVAE model. Due to its training structure, each pre-trained model can be viewed as a module, making it easier to achieve latent space transfer between domains. This implementation is based from this paper but only on one domain music (For experimentation purpose).

Here is my medium post to briefly explain the architecture of both models as well as some melody samples generated from SmallMusicVAE model you can listen to.

Just like official Magenta repo, you can pip install magenta==1.3.1 package (tested only on Python == 3.7) or if you want to install with anaconda, just simply type:

curl https://raw.githubusercontent.com/tensorflow/magenta/master/magenta/tools/magenta-install.sh > /tmp/magenta-install.sh
bash /tmp/magenta-install.sh

Or you can simply create new python3.7 environment, change directory to this repo and type the following line:

pip install -r requirement.txt

Do note that it's my working environment so there might be some unused packages.

Get cat-mel_2bar_big pre-trained musicVAE model (download here)

Just like MusicVAE training pipeline, we also need to convert a collection of MIDI files into NoteSequences by following the instructions in Building your Dataset. Additionally, we also need to load encoder, decoder part and latent space from MusicVAE pre-trained model for training smaller latent space. Assuming your generated examples are data/notesequences.tfrecord, training ae-cat-mel_2bar_big will be something like this:

python midime_train.py \
--config=ae-cat-mel_2bar_big \
--run_dir=<model output directory> \
--mode=train \
--examples_path=<path to sample tfrecord> \
--pretrained_path=<path to cat-mel_2bar_big.ckpt> \
--num_steps=100

After training we can load the model for melody generation. The mechanism is just similar to MusicVAE but we also need to load MusicVAE model instead of only loading SmallMusicVAE model.

python midime_generate.py \
--vae_config=cat-mel_2bar_big \
--config=ae-cat-mel_2bar_big \
--checkpoint_file=<path to your train model ckpt file> \
--vae_checkpoint_file=<path to cat-mel_2bar_big.ckpt> \
--num_outputs=5 \
--output_dir=<generated melody output path>

Similar to SmallMusicVAE, we just need to change the config value to lc-cat-mel_2bar_big:

python midime_train.py \
--config=lc-cat-mel_2bar_big \
--run_dir=<model output directory> \
--mode=train \
--examples_path=<sample tfrecord file> \
--pretrained_path=<path to cat-mel_2bar_big.ckpt> \
--num_steps=100

Same thing applies here, we just need to change config parameter to lc-cat-mel_2bar_big:

python midime_generate.py \
--vae_config=cat-mel_2bar_big \
--config=lc-cat-mel_2bar_big \
--checkpoint_file=<path to your model ckpt file> \
--vae_checkpoint_file=<path to cat-mel_2bar_big.ckpt> \
--num_outputs=5 \
--output_dir=<generated melody output path>

For ease of testing, I have put some generated sample data in data folder. To test the training phase, you can try this command:

python midime_train.py \
--config=ae-cat-mel_2bar_big \
--run_dir=tmp/ \
--mode=train \
--examples_path=data/fur_elise.tfrecord \
--pretrained_path=<path to cat-mel_2bar_big.ckpt> \
--num_steps=200

Your model will be in tmp/train folder.

For generating melody, you can try:

python midime_generate.py \
--vae_config=cat-mel_2bar_big \
--config=ae-cat-mel_2bar_big \
--checkpoint_file=tmp/train/model.ckpt-200 \
--vae_checkpoint_file=<path to cat-mel_2bar_big.ckpt> \
--num_outputs=15 \
--output_dir=tmp/generated

Your generated melodies will be in tmp/generated folder.