brangerbriz / midi-glove

Create MIDI note vector embeddings using GloVe (Global Vectors for Word Representation)

MIDI GloVe

Create GloVe embeddings from monophonic MIDI tracks. You can make your own embeddings or use our pre-trained embeddings created from 263,438 monophonic MIDI tracks from the Lakh MIDI Dataset. You can download the files we used to create our pre-trained embeddings here (139MB).

Pre-trained Embeddings

Pre-trained embeddings are included for all monophonic tracks (including monophonic tracks pulled from polyphonic MIDI files) from the Lakh MIDI Dataset (LMD). We've included both csv and binary representations of these embedding in the following dimensions:

• 10-D (bin | csv)
• 25-D (bin | csv)
• 50-D (bin | csv)
• 100-D (bin | csv)
• 200-D (bin | csv)
• 300-D (bin | csv)

You can find our pre-trained embeddings in data/pre-trained-embeddings. Below you will find the training loss for our pre-trained embeddings of each dimension.

Embeddings were created from a 159MB file that contained all note sequences from the monophonic instrument tracks in LMD seperated by an additional START_TRACK token. Note values were expressed as general MIDI note values 0-127 and no attempt to represent timing/rythm information was made. See notebooks/process_midi_for_glove.ipynb to see how MIDI data was prepaired before being processed by GloVe.

Creating Your Own Embeddings

Setup

# clone this repo
git clone https://github.com/brangerbriz/midi-glove.git
cd midi-glove

# initialize and clone the GloVe repo
git submodule init
git submodule update

# build GloVe
cd GloVe
make

# return to midi-glove project root
cd ../
pip install -r requirements.txt

Pre-process MIDI

Place a folder of MIDI files you would like to use create the embeddings inside of data/. Next start an Jupyter notebook server.

jupyter notebook

Open notebooks/process_midi_for_glove.ipynb and change the value of the midi_dir variable to point to the folder containing your MIDI files. Run the notebook. Processing can take a while (263,438 took ~36 min on my 4GHz CPU). When processing completes it will output a file to data/notes.txt. This will be used as the input to GloVe.

Alternatively you can also copy the contents of the Jupyter notebook to a python script and run it without Jupyter if you prefer.

Create Embeddings

To create the GloVe embeddings, run:

./create_embeddings.sh

This script will save the embeddings to data/embeddings by default, using these settings:

WORD_FILE=data/notes.txt
OUT_DIR=data/embeddings
MIN_COUNT=5
WINDOW_SIZE=15
DIMENSIONS=( 10 25 50 100 200 300 )
ITERATIONS=200

Edit create_embeddings.sh to change the default settings.

Plotting Loss

To plot the training loss of the GloVe algorithm, save the output of create_embeddings.sh.

./create_embeddings.sh &> data/embeddings/train.out

Use notebooks/plot_glove_training_loss.ipynb to load, parse, and graph the loss from this saved file.

Citations

Lakh MIDI Dataset

Colin Raffel. "Learning-Based Methods for Comparing Sequences, with Applications to Audio-to-MIDI Alignment and Matching". PhD Thesis, 2016.

GloVe: Global Vectors for Word Representation

Jeffrey Pennington, Richard Socher, and Christopher D. Manning. 2014. GloVe: Global Vectors for Word Representation.