Currently I 've implemented the Rnn-rbm network described in the paper Modeling Temporal Dependencies in High-Dimensional Sequences:Application to Polyphonic Music Generation and Transcription. A very simple implementation can be found on deeplearning.net
I am trying to make it work better using blocks and make it more modular
- Pretraining of Rbm
- Pretraining of Rnn
- Variable layer number Rbm
- Modular Rnn block with any number/kind of rnns. This would solve it
Unimportant stuff
- Get file of midis and train on them
- Somehow sample mp3 to create dataset