PLAN.txt

Mini notes generator

1. First create MIDI loader that loads and transforms MIDI files into MIDI matrix
   - divide time into 100ms long periods and set value 0..1 if note X is played
     with volume 0..1 (if multiple instruments play same note add volume up to 1),
     if note is only partially played during the time interval mark it is p% of 
     its value
   - use note range (3 octaves) A-2 - A-6 [48 notes]
   - ignore drums for now (we generate only melodies) 
     // [later generate only drums then get them in sync]
   - ignore everything else (we are only interested in notes)

   DONE.

2. For testing purposes write code to play randomly generated or loaded
   MIDI note matrixes using piano keys and simple bass 
   [artificially test good boundary value when to play piano keys and 
    when to switch to bass or just use piano for all]
    	 
   DONE.

3. Export MIDI note matrixes into dinrhiw2 dataset format for training
   stacked BB-BRM (initially) BB-RBM 
   [discretize note values first with somekind of 30% cutoff level]
   	       
   - take 4+1 seconds long MIDI note matrix sequences and vectorize
     for dinrhiw2 dataset input format for RBM

   Use 100ms accuracy although 10ms would be much better (be enough). 
   100ms causes serious damage to melodies but keeps "NOTE picture" sizes "sane".

   DONE.
   
   - test GB-RBM with floating point note volume values
     [feed combined input+ouput into 1 layer GB-RBM: this will have 
      60*5*10 = 3000 input dimensions.

      - test running time and reconstruction error with different number
        of hidden units:

	- 10 hidden units (just test code runs ok)
	- 100, 1000, 1500, 3000 units

      

4. MIDI note synthesis

   - learn 2 stacked RBMs with and without latest (one second)
   - use feedforward neural networks to map feature vectors of 
     two stacked RBMs into another and synthesize extra one second forward
   - add new notes (500ms?) and then loop to synthesize more notes..