Synopsis

This repo is a fork of: https://github.com/deepsound-project/genre-recognition

Two models are provided for training and validation:

• The first model is a modification on the LSTM model developed by Piotr Kozakowski, Jakub Królak, Łukasz Margas and Bartosz Michalak. The model relies on the weighted average of the probability distribution of genres at each time step.

• The second model was implemented by Nikhil George Titus based on http://benanne.github.io/2014/08/05/spotify-cnns.html This model achieved a validation accuracy of 83% on the current data split of the GTZAN data set.

Usage

In a fresh virtualenv type:

pip install -r requirements.txt

You can train your own model by modifying and running train_model.py. If you wish to train a model by yourself, download the GTZAN dataset (or provide analogous) to the data/ directory, extract it, run create_data_pickle.py to preprocess the data and then run train_model.py to train the model:

cd data
wget http://opihi.cs.uvic.ca/sound/genres.tar.gz
tar zxvf genres.tar.gz
cd ..
python create_data_pickle.py
 python train_model.py

By default the model 2 is chosen. The model can be changed using an optional model_choice parameter. eg:

python train_model.py -c 1

Acknowledgments

• This repo is a fork of: https://github.com/deepsound-project/genre-recognition. Many thanks to Piotr Kozakowski & Bartosz Michalak.
• Models were trained and tested as part of the neural networks project of COMPSCI 682 at UMASS Amherst. https://compsci682.github.io/. Please read the project report here: REPORT
• The second model was trained based on http://benanne.github.io/2014/08/05/spotify-cnns.html

Contact info

• Feel free to contact me at nikhiltitusk@gmail.com if you need any further info.