Closed napulen closed 2 years ago
I'm sure someone will say "what's the deal with just using the data/tools provided in the paper", I need to make my case by showing examples like this. This kind of thorough revision is what makes my algorithm work better than other models. More than the architecture/etc.
During the time I was correcting the corpus manually, I recorded a few videos showcasing scores that break one or more assumptions made by encoders, and that's why I need to inspect/curate them by hand or develop automated tools. I need to find those videos, but in the meantime, here is an example from DCML's Mozart Piano Sonatas:
K333-3 - measure 200
What's the length of that measure? How does the musicxml encode that? How does
music21
read that? How does theCSV
annotation files provided by DCML encode that? Usually instances like these are what breaks and misaligns everything. Then the performance of the network goes down. That's why I curate the data by hand.