This repository contains 3 proposed Convolutional Neural Networks (CNNs) that aims to classify classical music composers based on data from International Music Score Library Project (IMSLP). The first CNN accepts spectrogram of the audio clip as inputs. The second CNN extracts various features from the audio clip using Python's library Librosa and pass these numpy arrays in as inputs. The last CNN is an ensemble as it performs classification with both spectrogram and features.
contains the initialization and training loop for all the proposed CNNs
should contain the training, validation and testing set. However, as these files are too large to upload, the folder only contains a csv file with all the classical pieces used for our training and model evaluation
contains code for downloading data, converting them into arrays, train-test split and so on.
contains 2 other models (Residual Network and Autoencoder) that could be implemented to further improve the classification accuracies.
As the dataset is too large to be put on github, please utilize the data_preprocessing folder to obtain the dataset. You could attain audio clips with various lengths through manipulating sample_length and sample_ratio in data_convert.py. You could also change the train-test split ratio in data_divide.py through train and test.