Description

This dataset split contains 20 track_ids of each artist for 17,591 artists. We filtered out 20 songs for each artist which are randomly selected including various albums to prevent recording environment effects.

• This data consists of 7Digital_id/artist_index/song_index.

Usage

• In python 3.6.1,

  import pickle
  import numpy as np
  
  def load_label(num_artist, train_size=15, valid_size=18, test_size=20):
  
      file_list = pickle.load(open('msd-artist-list.pkl','rb'))
      file_list = file_list[:num_artist*20]  # Cut as many artists as you want to use.
      np.random.shuffle(file_list)
  
  ... 
  

• train_size=15, valid_size=18, test_size=20 means that we used 1-15 songs for training, 16-18 songs for validation and 19-20 songs for testing of each artist.

This dataset split was used in following works:

• Representation Learning of Music Using Artist Labels [pdf, code]
  Jiyoung Park, Jongpil Lee, Jangyeon Park, Jung-Woo Ha and Juhan Nam
  Proceedings of the 19th International Society for Music Information Retrieval Conference (ISMIR), 2018

• A Hybrid of Deep Audio Feature and i-vector for Artist Recognition [pdf]
  Jiyoung Park, Donghyun Kim, Jongpil Lee, Sangeun Kum and Juhan Nam
  Joint Workshop on Machine Learning for Music, the 34th International Conference on Machine Learning (ICML), 2018

• Representation Learning Using Artist Labels for Audio Classification Tasks [pdf, code]
  Jiyoung Park, Jongpil Lee, Jangyeon Park, Jung-Woo Ha and Juhan Nam
  Music Information Retrieval Evaluation eXchange (MIREX) in the 18th International Society for Musical Information Retrieval Conference (ISMIR), 2017
  (1st place in the Music Mood Classification task across all algorithms submitted so far)