=Paper=
{{Paper
|id=Vol-1984/Mediaeval_2017_paper_49
|storemode=property
|title=DNN in the AcousticBrainz Genre Task 2017
|pdfUrl=https://ceur-ws.org/Vol-1984/Mediaeval_2017_paper_49.pdf
|volume=Vol-1984
|authors=Nicolas Dauban
|dblpUrl=https://dblp.org/rec/conf/mediaeval/Dauban17
}}
==DNN in the AcousticBrainz Genre Task 2017==
https://ceur-ws.org/Vol-1984/Mediaeval_2017_paper_49.pdf
DNN in the AcousticBrainz Genre Task 2017
Nicolas Dauban
IRIT, Université de Toulouse, CNRS, Toulouse, France
nicolas.dauban@irit.fr
ABSTRACT 3.2 Architecture of the network
This paper presents a method of genre classification using deep To establish the final topology of the neural network, different
neural networks for the AcousticBrainz genre classification task of architectures have been tried. The final architecture comprises 8
MediaEval 2017. full connected layers of 2000 neurons with ReLUs as activation
function (Figure 1). The input vector had size of all the data used
(for the 9 statistics on the 40 MEL bands, the input was a vector of
1 INTRODUCTION size 360), and an output with the total number of genre and sub-
The AcousticBrainz Genre Task 2017 is a music genre recognition genres (e.g. 315 for the Discogs subset) with a sigmoid activation
(MGR) task organised by MediaEval [1] where participants had to function. With the sigmoid output function, it was common that
make predictions on genre and subgenres, based on audio features the network did not predict any genre for some tracks, because not
extracted from Essentia [2]. The target labels are provided by 4 a single output neuron had a value superior to 0.5 for those tracks.
different sources: Discogs, Allmusic, Tagtraum and Lastfm. In order to have at least one genre prediction per track if no genre
was predicted, the genre with the maximum output value is chosen
as the predicted one. For the submission, the network was trained
2 RELATED WORK
with 40 epochs.
I participated in this challenge during my internship on content-
based music recommendation. During this internship, I worked
on music genre recognition using music features with deep neural
network [3]. This approach has been tested on the GTZAN [4]
corpus and then on the MagnaTagATune [5] corpus. The accuracy
results obtained were about 90% on the GTZAN corpus and about
80% with MagnaTagATune. As these results were satisfying, we
decided to take part of the challenge to try the neural network
approach.
3 APPROACH
As the time evolution of the music features was not supplied,
the convolutional approach - which needs correlations between
successive frames of the input image - has been quickly aborted. Figure 1: DNN architecture
Thus, we used a classic deep neural network instead of a convolu-
tional one. The neural network has been implemented using the
Theano framework “Lasagne” 1 . We tried to use different features 4 RESULTS AND ANALYSIS
and different architectures of neural networks. Only one submission was made for each dataset, for the task 1
only.
3.1 Features The Table 1 and Table 2 contain the results per tracks for all
During the development phase, the following features have been labels and for genre labels:
tried: Mel bands, spectral rolloff, zero crossing rate, spectral en-
Table 1: Results per track, for all labels
tropy, Harmonic Pitch Class Profile (HPCP), tempo, danceability,
key strength, dissonance, tuning diatonic strength. The choice of
those features has been made by relying on their definitions and by Precision Recall F-score
choosing those which seem to be the more relevant to characterize Allmusic 0.3751 0.1455 0.1813
the different music genres. In our preliminary experiments, Mel Discogs 0.3043 0.1243 0.1664
bands yielded the best results, thus, only Mel bands were used as LastFM 0.1845 0.0813 0.1073
input for our final submission. Tagtraum 0.2892 0.1214 0.1649
1. https://lasagne.readthedocs.io/en/latest/ The results for sub-genres and per label are not shown here be-
cause they were all between 0% and 10%. All the results are provided
on the AcousticBrainz Genre Task results page 2 .
Copyright held by the owner/author(s).
MediaEval’17, 13-15 September 2017, Dublin, Ireland
2. https://multimediaeval.github.io/2017-AcousticBrainz-Genre-Task/results/
MediaEval’17, 13-15 September 2017, Dublin, Ireland N. Dauban
Table 2: Results per track, for all labels RÉFÉRENCES
[1] Dmitry Bogdanov, Alastair Porter, Julián Urbano, and Hendrik Schreiber. The
Precision Recall F-score mediaeval 2017 acousticbrainz genre task: Content-based music genre recognition
from multiple sources. In Working Notes Proceedings of the MediaEval Workshop,
Allmusic 0.3962 0.3334 0.3513 Dublin, Ireland, September 13-15, 2017.
Discogs 0.3221 0.2674 0.2812 [2] Dmitry Bogdanov, Nicolas Wack, Emilia Gómez, Sankalp Gulati, Perfecto Herrera,
Oscar Mayor, Gerard Roma, Justin Salamon, José R Zapata, Xavier Serra, et al.
LastFM 0.1937 0.1764 0.1813 Essentia: An audio analysis library for music information retrieval. In ISMIR,
Tagtraum 0.3231 0.3055 0.311 pages 493–498, 2013.
[3] Christine Senac, Thomas Pellegrini, Florian Mouret, and Julien Pinquier. Music
feature maps with convolutional neural networks for music genre classification.
In Proceedings of ACM Workshop on Content Based Multimedia Indexing (CBMI).
During the development phase, the network obtained a 60% F- ACM, June 19-21, 2017.
Score on genre-only on Discogs. During this phase, the data has [4] George Tzanetakis and P Cook. Gtzan genre collection. Music Analysis, Retrieval
been split using the python script provided by organizers to ensure and Synthesis for Audio Signals, 2002.
[5] Edith Law, Kris West, Michael I Mandel, Mert Bay, and J Stephen Downie. Eva-
album filtering: 80% to train the network and 20% to test it. However, luation of algorithms using games: The case of music tagging. In ISMIR, pages
the best F-score obtained on genre labels of our final submission 387–392, 2009.
was 35% per track maybe due to a lack of generalization power of
our network.
5 DISCUSSION
We plan to explore our initial idea to use combinations of dif-
ferent acoustic feature types as we previously showed that these
were as successful as using Mel bands. Furthermore, some modifi-
cations on the network can be done: adding batch-normalization
after each dense layer, residual blocks, try some different initializa-
tion and decay strategies for the learning rate, and also lower the
threshold of the output sigmoid in order to give more predictions.