This is my first deep learning project in which I implemented a LSTM model on abc notation music data.
In this deep learning project, which is inspired by the Andrej Karpathy’s blog ‘The Unreasonable Effectiveness of Recurrent Neural Networks’, I implemented a LSTM model on abc notation music data.
This repository requires installation of Tensorflow, Numpy and Keras to run smoothly.
The input data is present in ‘./data’ named ‘input.txt’.
The accuracy and loss associated with every epoch is stored in ‘./logs’ in csv format named ‘training_log.csv’.
Model at every epoch is stored in ‘./model’ in h5 format.
A Jupyter Notebook is also presented in which I visualized the result generated by the model.
Use ‘sample.py’ to run the trained model with no number of epochs, length and seed as arguments to the program. The output generated by this script can be pasted here ‘https://abcjs.net/abcjs-editor.html‘ to play the generated music.
Use ‘train.py’ to train the model from scratch. Here the arguments are input, epochs and save_freq.
Use ‘model.py’ to change the parameters of the model.