FMA Website | Dataset | References

This repository contains PyTorch implementation of a CNN model for music genre classification of the FMA dataset. Specifically the large subset of data was used and only tracks having the top genre label were considered.

The audio_clips is a PyTorch Dataset class that reads one track at a time. It takes the Mel spectrogram of the raw waveform and converts it to a 128 by 128 image. This image is then the input to the CNN model. Since decoding the mp3 files is very slow, it dumps the generated spectrogram image into a pickle file for later use. The mp3 decoding is only done once, on subsequent fetches if the pickle file is availale, the image is directly read from there.

Helpful to get familiar with the dataset by going through some of the jupyter notebooks here

Also look at usage.py for reading a track and generating a mel-spectrogram

The data module contains fma_data.py which contains the class definition of audio_clips The models module contains a sample CNN model which operates on 128 x 128 spectrogram images main.py is the script to train the model test_mymodel.py is to test the trained model on the test set

Install PyTorch according to the directions at the PyTorch Website for your operating system and CUDA setup.

In addition following libraries are needed -

• Librosa

1. Defferrard, Michäel, Kirell Benzi, Pierre Vandergheynst, and Xavier Bresson (2017). FMA: ADataset for Music Analysis. In:18th International Society for Music Information RetrievalConference (ISMIR). arXiv:1612.01840.
2. Defferrard, Michäel, Sharada P. Mohanty, Sean F. Carroll, and Marcel Salath́e (2018). Learning to Recognize Musical Genre from Audio. Challenge Overview. In:The 2018 Web ConferenceCompanion. ACM Press.isbn: 9781450356404.doi:10.1145/3184558.3192310. arXiv:1803.05337.
3. Dieleman, Sander and Benjamin Schrauwen (2014). “End-to-end learning for music audio”. In:2014 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP),pp. 6964–6968.doi:10.1109/ICASSP.2014.6854950.