Deep Symphony Overfitting (Case Study)

Posted on October 18, 2017

Repetition starts from 2:23

Repetition starts from 1:10

 

0035.mid is a proper example to demonstrate the overfitting problem. To exam whether the generated song is similar to some pieces in the dataset, I implemented a Longest Common Subsequence algorithm to check the maximum matching between songs.

The following result is the longest matching length between 0035.mid and other songs:

(340, './datasets/easymusicnotes/level9/cuando-los-recuerdos-te-hacen-llorar-victor-m-barba-piano-level-9.mid'),
(296, './datasets/easymusicnotes/level6/You_Decorated_My_Life_Kenny_Rodgers_Ballad_Piano_Level_6_Karaoke.mid'),
(295, './datasets/easymusicnotes/level13/concierto-triste-para-trompeta-II-victor-m-barba-instrumental-piano-level-13.mid'),
(269, './datasets/easymusicnotes/level13/concierto-triste-para-trompeta-I-victor-m-barba-instrumental-piano-level-13.mid'),
(255, './datasets/easymusicnotes/level11/no-hay-nada-mas-triste-victor-m-barba-piano-level-11.mid'),
(229, './datasets/easymusicnotes/level11/Como-yo-te-siento-victor-m-barba-piano-level-11.mid'),
(227, './datasets/easymusicnotes/level12/tonight-i-wanna-cry-keith-urban-country-piano-level-12.mid'),
(226, './datasets/easymusicnotes/level9/the-entertainer-scott-joplin-ragtime-piano-level-9.mid'),
(225, './datasets/easymusicnotes/level8/Almost_Lover_-_A_Fine_Frenzy_-_Piano_Level_8.mid'),
(213, './datasets/easymusicnotes/level13/piano-bebu-silvetti-instrumental-piano-level-13.mid'),
...

And I visualize the longest matching between cuando-los-recuerdos-te-hancen-llorar-victor-m-barba-piano-level-9.mid (first row) and 0035.mid (second row). The dark areas are the matched subsequences.

As shown in the figure, 0035.mid copies most of the pieces from an existing song (start from 1:45). There are a few modifications but they still sound the same.

 

 

None