Let's set up our system for the model:

1. In Piano, the lowest note is 24 and the highest is 102; hence, the range of notes is 78. Thus, the number of columns in the encoded matrix is 156 (that is, 78 for note-on and 78 for note-off):

lowest_note = 24L 
highest_note = 102L 
note_range = highest_note-lowest_note 

2. We will create notes for 15 number of steps at a time with 2,340 nodes in the input layer and 50 nodes in the hidden layer:

num_timesteps  = 15L 
num_input      = 2L*note_range*num_timesteps 
num_hidden       = 50L 
 

3. The learning rate (alpha) is 0.1:

alpha<-0.1