This Repository contains the Semester project for the Neural Networks of Teun Boersma, Phillip Rode, Robert Hielkema and Alexia Zaharie. The sentence directory contains a basic RNN, running and training script for the prediction of a sentence. For the final project we have created an ESN to complete the Contrapunctus XIV Fugue by Bach. This repository contains the code used. On this branch under ESN_Bach, all the code implementing the model (music_model.py) and analysing the data can be found. Should you wish to run any script, I enjoy using command line arguments a lot. The code operates using names, as an example the main file. Here running 'python3 main.py grid name' will run a grid search as specified in the code and save it under name. The same goes for the key words train and run replacing grid, which - perhaps self-explanatorily - train a model and run an existing model.
Any model trained using main.py train ... will automatically be saved in the models directory. This can then be run using main.py run ..., upon doing so some cool numbers are printed and the sound files are saved in the wav output directory.
Should you for whatever reason wish to run a grid search this can be done using the main file, the output will be saved in the gridsearch_raw directory. Each folder contains the results of that model, if the model can predict duration well enough to not risk crashing the pc trying to create a sound file, the outputs and weights of the model can also be found in this folder. To get an analysable dataset the data collector should be run using the name of the grid search conducted.
If you are interested, on the other branch of this repo, our original code can be found before switching to ESNs. Additionally an elementary sentence completion model can be found here. Its an RNN which can quickly be trained to copy any 'normal' sentence of around 50 words.
Also one last note, everything in this repo is based around python 3.10.12