Comparing Shallow Versus Deep Neural Network Architectures For Automatic Music Genre Classification

In this report we examine the paper by A. Schindler, T. Lidy and A. Rauber. Firstly detailing similar works and their relation to the paper, then re-implementing the architecture described in the paper in order to reproduce the results found. We then propose 2 further adaptations and show how these improve upon the results found.

Dataset

The dataset should be placed in the same directory and named 'music_genres_dataset.pkl'.

BlueCrystal Imports

Ensure numpy and tensorflow are on the correct versions.

module add libs/tensorflow/1.2
module load languages/anaconda2/5.0.1
pip install --user librosa

Batch File

We've provided a batch file to run with sbatch.

sbatch batch_classify.sh

Running

python classify_music.py [FLAG OPTIONS]

Flag	Options	Meaning
--decay	[const, exp]	Learning rate - constant or exponential decay.
--repr-func	[mel, cqt_note]	Spectrogram representation - mel or CQT.
--net-depth	[shallow, deep]	Network depth - shallow or deep.
--max-steps	[integer > 0]	Number of epochs to run for.
--learning-rate	[float > 0]	Learning rate (initial rate if exp. decay used)

Class Indices:

Name	Index
blues	0
classical	1
country	2
disco	3
hiphop	4
jazz	5
metal	6
pop	7
reggae	8
rock	9