2/8/19
2/15/19
- Downloaded/processing the databases in the study
- lots of midi data for piano parts and corresponding orchestral scores
- lots of variability in quality
- Working on implementing their alignment algorithm
2/22/19
- Looking at parallels between orchestration and language translations
- piano -> orchestra using RNN LSTM networks?
- training on each quanta of a midi file?
3/1/19
- Working on data tools
- Alignment algorithm creates folders of "aligned" midi files for the piano and orchestral scores
- writing tools to process these midi files into a format that can be read by ML model
- wrote code to store each song as a series of arrays
- song is divided into frames, each lasting one adjustable "quanta" of time (eg. sixteenth notes)
- frames have vales for each of the 122 midi notes.
3/15/19
- Trying to implement the RNN in pytorch
- Discovered more data issues
- code exists to process the midi into numpy arrays, but not the other way around
- wrote data tools and tested converting to and from midi
- Started working on the RNN
- Hard to find good examples/tutorials online
- Most examples are for language translation which isn't that similar to midi
4/12/19
- Trying to get the network training on multiple songs
- currently the network can only train on one of the 40 songs
- temporary solution—save model and load different data every time
- possible long term solution—use Keras generator not fit
4/19/19
- Working on training the model more/making it more accurate
- cuda not working - Docker?
- training on all 40 songs
- adjusting dropout layer
- results are either too dense or too sparse