This project aims to predict hotttnesss of a song. It is a decimal score, between zero and one, based on The Echo Nest data from 2010. It is limited by the subset provided by millionsongdataset. By simplicity, it was only considered the songs attributes of all the information. The major groups are metadata, analysis, and musicbrainz
Setup your project folders and files
chmod +x install.sh
./install.shDataset of 10k songs, random by MSD. It has three groups: metadata, analysis, musicbrainz
Metadata group with 20 attributes
dtype([
('analyzer_version', 'S32'), #0
('artist_7digitalid', '<i4'), #1
('artist_familiarity', '<f8'), #2 - Yes
('artist_hotttnesss', '<f8'), #3 - Yes
('artist_id', 'S32'), #4
('artist_latitude', '<f8'), #5
('artist_location', 'S1024'), #6
('artist_longitude', '<f8'), #7
('artist_mbid', 'S40'), #8
('artist_name', 'S1024'), #9
('artist_playmeid', '<i4'), #10
('genre', 'S1024'), #11 - Maybe
('idx_artist_terms', '<i4'), #12
('idx_similar_artists', '<i4'), #13
('release', 'S1024'), #14
('release_7digitalid', '<i4'), #15
('song_hotttnesss', '<f8'), #16 - Target - Can be a NaN | Filter
('song_id', 'S32'), #17
('title', 'S1024'), #18
('track_7digitalid', '<i4') #19
])Analysis group with 31 attributes
dtype([
('analysis_sample_rate', '<i4'), #0
('audio_md5', 'S32'), #1
('danceability', '<f8'), #2 - X Empty [0, 1]
('duration', '<f8'), #3 - Yes
('end_of_fade_in', '<f8'), #4 - Yes
('energy', '<f8'), #5 - X Empty [0, 1]
('idx_bars_confidence', '<i4'), #6 - X Empty
('idx_bars_start', '<i4'), #7 - X Empty
('idx_beats_confidence', '<i4'), #8 - X Empty
('idx_beats_start', '<i4'), #9 - X Empty
('idx_sections_confidence', '<i4'), #10 - X Empty
('idx_sections_start', '<i4'), #11 - X Empty
('idx_segments_confidence', '<i4'), #12 - X Empty
('idx_segments_loudness_max', '<i4'), #13 - X Empty
('idx_segments_loudness_max_time', '<i4'), #14 - X Empty
('idx_segments_loudness_start', '<i4'), #15 - X Empty
('idx_segments_pitches', '<i4'), #16 - X Empty
('idx_segments_start', '<i4'), #17 - X Empty
('idx_segments_timbre', '<i4'), #18 - X Empty
('idx_tatums_confidence', '<i4'), #19 - X Empty
('idx_tatums_start', '<i4'), #20 - X Empty
('key', '<i4'), #21
('key_confidence', '<f8'), #22
('loudness', '<f8'), #23
('mode', '<i4'), #24
('mode_confidence', '<f8'), #25
('start_of_fade_out', '<f8'), #26
('tempo', '<f8'), #27
('time_signature', '<i4'), #28
('time_signature_confidence', '<f8'), #29
('track_id', 'S32') #30
])dtype([
('idx_artist_mbtags', '<i4'), #0
('year', '<i4') #1
])Some values are undefined, before the training part, there was a validation process to verify non-nullable or missing values. The 10k records downloaded from MSD, are not provided with danceability and energy. In consequence, the model will be missing important attributes.
The model implemented is linear regression. A numeric score should be the result of the model, therefore it is a regression problem. It is limited to attributes that could not be related to the real calc of the measure.
It will help us to measure the behavior of the model, in other words, what is the loss. As in an article mentioned in the references, the loss(not the same as cost) function describes the form of the error, and the descent will be the downside process. For gradient descent, the loss function should be derivable.
The chosen cost function to measure our model is RMSE. It is because a linear relation of units will facilitate the analysis, outliers could affect it. Important this function is only for analysis purposes and not for improving the model.
The cost function to train the model
The purpose is minimize the error between the prediction, so the model needs to move down towards the function. For that, it can be derivated, this value means how much our descent should move, it will be also affected by the learning rate.
It will be removed to the previous param value
Using 14 attributes(including bias), with a learning rate of 10, it diverges. Parameters oscillate between positive and negative values, but they can not converge
We can get better results with a lower learning rate of 0.01
With 5000 epochs, looking for less than 0.01 with a learning rate of 0.1; Train size 2950, Test size 1264
# Final Parameters
Bias = 4.5689969077910070e-01
audio_md5 = 5.4573519380770839e-03
year = 3.4212693288237568e-02
loudness = 1.3203266354525294e-02
mode = -4.3669295455111487e-03
tempo = 2.6598180921980005e-03
time_signature = 4.1232899012358566e-04
artist_familiarity = 4.6258212788267319e-02
artist_hotttnesss = 3.4633849986531004e-02After applying the model with output parameters
Samples from 562 to 572
hyp 0.26586104921065007 y 0.401203522922599
hyp 0.4137398358601097 y 0.5684640847172688
hyp 0.2416552517044926 y 0.6336199597135092
hyp 0.3552855364230894 y 0.435398565057618
hyp 0.52549711888017 y 0.43113513759439415
hyp 0.6582414367557253 y 0.5033091661911661
hyp 0.4176416475968032 y 0.4514920652295682
hyp 0.541551997892655 y 0.39091498285641035
hyp 0.46384907791765656 y 0.404715590122965
hyp 0.5616924111789523 y 0.5144792508848043
The previous results demonstrate that the variance is significant, but it still keeps in the expected range
The previous model was tested with multiple epoch values, but the result with higher values does not have a relevant improvement. A fixed learning rate gives a problem if it is bigger than 2. To improve it, a dynamic learning rate can be implemented.
- https://github.com/mdeff/fma
- http://millionsongdataset.com/musixmatch/
- http://millionsongdataset.com/pages/field-list/
- https://medium.com/atchai/in-search-of-the-perfect-music-dataset-ed7e111d3b7e
- https://heartbeat.fritz.ai/5-regression-loss-functions-all-machine-learners-should-know-4fb140e9d4b0
- https://towardsdatascience.com/difference-between-batch-gradient-descent-and-stochastic-gradient-descent-1187f1291aa1
- https://towardsdatascience.com/batch-mini-batch-stochastic-gradient-descent-7a62ecba642a
- https://www.dataquest.io/blog/understanding-regression-error-metrics/
Thierry Bertin-Mahieux, Daniel P.W. Ellis, Brian Whitman, and Paul Lamere. The Million Song Dataset. In Proceedings of the 12th International Society for Music Information Retrieval Conference (ISMIR 2011), 2011.
#/mnt/d/tmp_ml/msd/MillionSongSubset/data
find . -type f -name \*.h5 -exec cp \{\} /mnt/d/tmp_ml/msd/MillionSongSubset/tracks/ \;
# find . -type f -name \*.txt -exec cp \{\} /home/jupyter/xseed-test/data/domains \;
# ls | xargs -I{ cat { > /home/jupyter/xseed-test/data/domains.txt- Look for music datasets
- Understand datasets
- Define the scope for regression algorithm
- Implement model
- Write documentation