Build music classification Neural Network models using spectrograms as input data with Python 3.6 + Tensorflow + Keras
In order to get the dataset used and some utilery, you'll have to clone the fma github repo inside this project's directory. Download the dataset on the data section of the repo and place it inside the fma directory.
Create a fresh virtual environment, conda preferably, for Python 3.6 and run from terminal pip install -r requirements.txt.
The model implemented is based on Sander Dielemann's architecture and is implemented on models.Dielemann.py.
To train it, from your terminal, at the root level of this repo, run
python -m training.Dielmann_windows hop_length_param number_of_windows_param model_name_param number_of_epochs_param.
For example, python -m training.Dielemann_windows 355 10 "hop_length_355.h5" 5 corresponds to build spectrograms with a hop length of 355, building 10 windows, model saved will be 'hop_length_355.h5' and will be trained for 5 epochs.
For a comprehensive evaluation of the model, use the Jupyter Notebook Evaluation and confusion matrix.ipynb to see the performance of the model on each subset of the dataset.
For feature visualization of the model, use the Jupyter Notebook Feature visualization.ipynb