Mirex 2019 Audio Classification (Train/Test) Tasks

Tasks

Task	No of samples	No of Classes
Audio Classical Composer Identification	2772	11
Audio US Pop Music Genre Classification	7000	10
Audio Latin Music Genre Classification	3227	10
Audio Music Mood Classification	600	5
Audio K-POP Mood Classification	1438	5
Audio K-POP Genre Classification	1894	7

Requirements

• Python >= 3.6
• Python packages:
  • librosa, numpy, pandas, joblib, tqdm, sklearn, albumentations, runstats
  • PyTorch >= 1.1

Setting up environment

Use the provided packaged archive (created using conda-pack):

• mkdir -p mirex
• tar -xzf mirex.tar.gz -C mirex
• source mirex/bin/activate
• conda-unpack

Or, create a new conda environment using the provided environment.yml file: conda create -f environment.yml

Running commands

generating sample data

python generate_sample_data.py -d data/ -i data/sample.wav

Feature extraction

python extract_features.py -s /home/scratch -i data/features_extraction.txt -n 4

Training

python train.py -s /home/scratch -i data/train.txt -n 4 -t kpop_mood

Classifying

python classify.py -s /home/scratch -i data/test.txt -o test_preds.txt -n 4 -t kpop_mood

Time taken

Features extraction

• 4 threads ~ 5 min
• ~ 1.5 GB memory for extracted features # maychange with parameters