Simple comparison with the three tokenization method.
REMI&MIDILike: not too much difference, at the end runs to 85-90%. CPword: it is combined, and under the same configuration it doesn't work as nice as the the other two. But it should require larger architecture.
(They all seems to overfit at some point. and the current model is not large at all. )
Section 1
val_acc
val_acc
Select runs that logged val_acc to visualize data in this line chart.
val_loss
val_loss
Select runs that logged val_loss to visualize data in this line chart.
val_fscore
val_fscore
Select runs that logged val_fscore to visualize data in this line chart.