Each subdirectory contains the configuration file for a single model. v01_drums_vel
is the main proposed model,
the rest are from the ablation study (section VI-C in the paper). Generally, we find that v01_drums
(which does not model
velocity) performs considerably better on real-world MIDI files than the main proposed model.
v01_drums_vel
('dr. + vel.') supports both drums and velocityv01_drums
('drums') does not support velocityv01_vel
('velocity') does not support drumsv01
('none') supports neither of the abovev01_drums_vel_perf
('dr. + vel. + 𝛥') is likev01_drums_vel
, but uses the 𝛥-encoding
To use the pre-trained model checkpoints, extract them inside this directory (the checkpoint files for each model should end up next to the corresponding model.yaml
file).