Closed napulen closed 2 years ago
aaaand that is exactly what it learned. Absolutely no chords written in my inference test :).
Dropout was removed. It created all sorts of issues in other experiments I ran. I still don't understand why it only affects the harmonicrhythm, but it is gone now.
Using dropout increases the accuracy across the SATB output tasks.
It does seem to get the
HarmonicRhythm2
stuck at around 94%.I made a quick check on a few BPS samples, and they have ~94% non-onset frames. So that's pretty terrible, basically, it could be learning a classifier that throws a
noOnset
on every frame.