AutoRBM automatically creates a RBM object that follows the given graph description. It dynamically creates object functions, functions names, functions docstrings and functions signatures according to the graph description. This results in a fully operational RBM architecture with functions inputs & functions names named following the layers names given in the graph description.
AutoRBM is able to :
- build Factored RBM (gated factored RBM only, no gated RBM)
- build Deep RBM (multi-hidden layer)
- deal with gaussian and binary layers (Relu layers are coming)
- automatically detect conditional layers
- automatically creates parameters
- build probabilities and samples functions wrt to needed inputs (first order and second order inputs if there are message-passing/linear layer)
- perform a network stabilisation using Gibbs sampling (for deep propagation through the network and for co-dependant layers)
- CD-k gradients estimation
This project use JAX as backend but it can easily be adapted to other ML libraries or Numpy
I added a notebook test_autoRBM you can play with to check how it works.
Feel free to take this code to play with and thanks for any help you could give me :)