OK COMPOSER: SYMBOLIC MUSIC GENERATION WITH RECURRENT NEURAL NETWORKS

Author: Cristian Mitroi

June, 2019

Code instructions

Requires Python 3.6

Modules:

keras_self_attention
keras
music21
tensorflow
sklearn
keras_tqdm
pypianoroll
tqdm
librosa
pretty_midi
matplotlib
numpy

Pretrained models

Can be found in here.

Follow the instruction in sampling to generate samples, based on your own input.

PREPROCESSING A DATASET

NOTE: There is a small subset of the folk dataset available here, in the folkdataset_small folder.

MAKING YOUR OWN

You need to install the Magenta fork from this github: https://github.com/cristianmtr/magenta

Place it a folder and edit preprocessing\procedure.py (and other files where the magenta folder is added to the path):

from

sys.path.append("D:\\data\\magenta-1.0.2\\magenta\\music")

to

sys.path.append("YOUR PATH")

Download the dataset and place the MIDI files in a folder called 1orig in an parent empty folder.

Create a script (your dataset)_main.py in the folder preprocessing.

Put the code below in that script

import subprocess
from glob import glob
import os
import filter_four_four
import encode
import transpose

import procedure

if __name__ == "__main__":
    max_sequences = 100000
    nr_bars = 4
    bar_len = 16  # based on steps_per_quarter=4 in midi_file_to_melody in D:\data\magenta-1.0.2\magenta\music\melodies_lib.py
    max_seq_len = nr_bars * bar_len

    # top dir
    top_dir = "(PATH TO TOP DIR)"
    procedure.main(top_dir, max_sequences, nr_bars, bar_len, max_seq_len)

Run the script

It will generate two datasets, one in pianoroll and one in magenta melody format.

Edit the config.py file to point to these.

TRAINING A MODEL:

NOTE: A pre-trained simple model is included in model_folksmall_pianoroll_1lstm10_noattention.zip

python main.py --new --layers (nr of layers) --cells (nr of cells per layer) --dataset (which dataset)

add --bi if you want to use bidirectional wrapper for each layer add --att if you want to add an attention layer on top

list of datasets can be found in config.py

model will be saved in a folder named model_(dataset)_(encoding)_(bidir.)(nr layers)lstm(nr cells)_(attention) (e.g. model_folk100k_melody_1lstm10_noattention or model_folk100k_melody_bi1lstm10_attention)

CONTINUE TRAINING OF A MODEL:

python main.py --id (id of model)

SAMPLING

python sampling.py (MODEL FOLDER/ID) (PATH TO MIDI FILE) --nr=(NR OF SAMPLES) --temp=(TEMP)

e.g.

python sampling.py model_folk100k_melody_1lstm10_noattention d:\data\thesis_model2\lamb.mid --nr=20 --temp=1.0

Prompt will indicate where files are stored

EVALUATION

NOTE: You need at least 100 samples in the 7_comparison file and 100 generated samples.

You need a separate python 2.7 environment to run the mgeval comparison tool.

Modules required:

seaborn
matplotlib
pypianoroll
numpy

cd mgeval2

Comparison with training set

python compare.py (PATH TO DATASET)\7_comparison training_set (PATH TO SAMPLES DIR) model (NAME OF COMPARISON)

e.g.

python compare.py ..\folkdataset_small\7_comparison training_set ..\model_folksmall_pianoroll_1lstm10_noattention\samples\sessiontune102_transposed\ model debug_comparison

PLAGIARISM CHECKER

You need to install Java

open cmd prompt in plagiarism_checker

java -jar melodyshape-1.4.jar -q (PATH TO MODEL SAMPLES) -c (PATH TO MIDI DATASET TOP DIR)\4_mono -a 2015-shapeh -k 2 -t 4 | paste -s -d '\n' - > match.tsv

This will output scores in the match.tsv file

To generate a Latex table with stats:

python scripts\extract_plagiarism.py plagiarism_checker\match.tsv model