This project involves the implementation of a music generating model trained by a Long Short-Term Memory artificial neural network. The final model generates music which displays characteristics common in that of the folk music genre:
- monophonic, a simple melody without any accompanying harmony or chords
- of the AABB form (binary form) which represents the repetition of a melody A and a melody B
- ends on the tonic, if the key of a song is in G, the song should end with a G note
Music in general is the result of a sequence of musical notes played and organised within a length of time. In this way, music is structured and sequential, much like natural language, where musical notes are words and measures are sentences. The final generator is a result of data preparation, data processing, and neural network training, all implemented in Python.
This project requires that PyTorch be installed
The dataset that was used for this project was the Nottingham Database which contains 1200 monophonic British and American folk tunes all written in abc notation. nottingham_data_prep.py prepares the data by removing all comments, chords (written between quotation marks), and meta-data excluding the meter M, the key K, and the tune itslef. An example of a prepared tune is as follows
M:4/4
K:G
f|: || |:g3/2a/2 gf|gd2e|dc BA|BG2A|BG2A/2B/2|\
cA2B/2c/2|Bd cB| [1Ad ef:|[2A2 A2||
BG2A/2B/2|cA dB|ec2d/2e/2|fzde/2f/2|
g3/2a/2 ge|df gB|ce d/2e/2d/2c/2|BG G2:|
Here is what the abc notated text looks like on a traditional music staff
All the tunes are saved in a single file: nottingham_parsed.txt in the datasets/nottingham_database/ directory.
The LSTM is trained in the music_train.py file. data_proc.py takes care of all token encoding for the neural network.
All hyperparameters are configured in the model_configuration.py file. The hyperparameters that are set were determined by means of a grid search and the best results were chosen by comparing learning curves. The tuned and tested hyperparameters are as follows:
- number of epochs = 100
- number of layers = 3
- number of hidden units = 512
- learning rate = 0.001
- train-validation cut = 0.14
- dropout rate = 0.4
The batch and sequence sizes were chosen by calculating an appropriate value based on the total length of the nottingham dataset.
model_configuration.py can be modified to train a model of your own.
To train a model execute the following:
python3 music_train.py [model_name.pth]
If no model file path is provided the model will be saved in the working directory as model.pth allong with a log of the the same name and a .png image of the training and validation learning curves.
Music generation is all implemented in music_gen.py . To generate a folk tune, run:
python3 music_gen.py -m <model_path> [-f file_path] [-p prompt]
model_path is the path to the trained model which must be provided in order to generate music. This means that the provided model ./models/musicmodel.pth must be used or a model must be trained.
file_path is the the path including the name of the file where the tune data will be saved in abc notation.
prompt is what will be fed to the model in order for it to predict the tokens that would most likely follow. By default, it is a predefined end of sequence (eos) token: </s>. However the prompt can be any of the following:
- The eos token
</s> - A meter, ex:
M:4/4 - A meter and a key, ex:
'M:3/4\nK:Dm'(note the single quotes to acknowledge the newline)
To generate a tune of meter 3/4 and in the D minor key:
python3 music_gen.py -m ./models/musicmodel.pth -f ./mytune.abc -p 'M:3/4\nK:Dm'
Here is the output, written in abc notation:
M:3/4
K:Dm
G/2G/2|BB A2|A/2B/2A/2G/2 FG|A3/2A/2 EG|E3A|G3/2A/2 Bc|
d3/2c/2 Bd|ed ef|g2 f3/2e/2|dc Bd|fe ef|\
g3/2a/2 ga/2a/2|
af ag|f3 f3/2a/2|af ga|a2 d2|e3/2g/2 f/2e/2d/2e/2|f2 d:|
Here is the generated tune written on a staff:
- EasyABC - Open source software for creating, editing, and visualising ABC data
- Direct MIDI to MP3 Converter - Used in this project to convert to and from MIDI and WAV audio formats