This repository contains the source code to reproduce the results of the ISMIR'19 paper Learning to Generate Music with Sentiment. This paper presented a generative LSTM that can be controlled to generate symbolic music with a given sentiment (positive or negative). The LSTM is controlled by optimizing with a Genetic Algorithm the values of specific neurons that are responsible for the sentiment signal. Such neurons are found plugging a Logistic Regression to the LSTM and training the Logistic Regression to classify sentiment of symbolic music encoded with the mLSTM hidden states.
This project depends on tensorflow (2.0), numpy and music21, so you need to install them first:
$ pip3 install tensorflow tensorflow-gpu numpy music21A new dataset called VGMIDI was created for this paper. It contains 95 labelled pieces according to sentiment as well as 728 non-labelled pieces. All pieces are piano arrangements of video game soundtracks in MIDI format. The VGMIDI dataset has its own GitHub repo that is constantly updated. To reproduce the results of the ISMIR'19 paper, use the VGMIDI 0.1 that can be downloaded as follows:
wget https://github.com/lucasnfe/vgmidi/archive/0.1.zipTo make sur the paths will match in the rest of this tutorial. Rename the dataset directory from vgmidi-0.1 to vgmidi and make sure the vgmidi directory is inside the music-sentneuron directory.
To process the MIDI data with neural networks, one needs to encode the MIDI data as a sequence of vectors. To augment encode the In this work, we use a one-hot encoding strategy:
python3.7 midi_encoder.py --path vgmidi/unlabelled/train --transp 10 --strech 10
python3.7 midi_encoder.py --path vgmidi/unlabelled/test --transp 10 --strech 10Note that this will create a .txt file for each .mid file in the train and test directories. Each line in these files represents an encoded augmented version of the original mid piece. For example, in the train directory there should be a Dragon_Warrior_Battle_Theme.txt file that starts with:
t_80 v_68 d_16th_0 n_38 n_41 w_1 v_76 n_41 n_44 n_47 w_1 ...Make sure these txt files exist and are not empty before procedding.
This paper has two main results: (1) sentiment analysis of symbolic music and (2) generation of symbolic music with sentiment.
To train a sentiment classifier of symbolic music:
python3.7 train_generative.py --train ../vgmidi/unlabelled/train/ --test ../vgmidi/unlabelled/test/ --embed 256 --units 512 --layers 4 --batch 64 --epochs 15 --lrate 0.00001 --seqlen 256 --drop 0.05This script saves a checkpoint of the trained model after every epoch in the "trained/" folder. Note that the model/training parameters are slightly different than the ones presented in the paper. These parameters should achieve similar results and are faster to train. Feel free to adjust the pameters to the original ones (--embed 64, --units 4096 --batch 32 --epochs 4 --lrate 0.000001).
To sample from this trained generative model (without sentiment control):
python3.7 midi_generator.py --model trained --ch2ix trained/char2idx.json --embed 256 --units 512 --layers 4The following script will encode the labelled pieces with the final cell states of the generative LSTM and train the logistic regression model:
python3.7 train_classifier.py --model trained --ch2ix trained/char2idx.json --embed 256 --units 512 --layers 4 --train ../vgmidi/labelled/vgmidi_sent_train.csv --test ../vgmidi/labelled/vgmidi_sent_test.csv --cellix 4After running this script, a binary file named "classifier_ckpt.p" containing the trained classifier logistic regression is saved in the "trained/" folder.
To control the sentiment of the trained generative LSTM, we evolve the values of the neurons activated during training the logistic regression classifier.
evolve_generative.py --ch2ix trained/char2idx.json --embed 256 --units 512 --layers 4 --genmodel trained/ --clsmodel trained/classifier_ckpt.p --cellix 4 --elitism 0.1 --epochs 10 --sent 1After running this script, a json file named "neurons_positive.json" containing the neuron values that control the generative model to be positive is saved in the "trained/" folder.
python3.7 midi_generator.py --model trained --ch2ix trained/char2idx.json --embed 256 --units 512 --layers 4 --cellixAfter running this script, a json file named "neurons_negative.json" containing the neuron values that control the generative model to be netagive is saved in the "trained/" folder.
python3 midi_generator.py --model trained/ --ch2ix trained/char2idx.json --embed 256 --units 512 --layers 4 --seqlen 512 --override trained/neurons_positive.json --cellix 4python3 midi_generator.py --model trained/ --ch2ix trained/char2idx.json --embed 256 --units 512 --layers 4 --seqlen 512 --override trained/neurons_negative.json --cellix 4If you use this method in your research, please cite:
@article{ferreira_ismir_2019,
title={Learning to Generate Music with Sentiment},
author={Ferreira, Lucas N. and Whitehead, Jim},
booktitle = {Proceedings of the Conference of the International Society for Music Information Retrieval},
series = {ISMIR'19},
year={2019},
}