rezer0dai/zer0nets

experimenting with neural network (FFNN) architecture

• inspired by courses of AndrewNg

  • vectorization of batches
  • universal layered approach

• and udacity template

  • class encapsulation approach

• architecture is encapsulated more over to FeatureSpaces

  • each space has own weights & activation fuction
  • define derivative, back-prop and forward-pass on this level

• NeuralNetwork class is responsible only for puting it all together

  • initializing spaces
  • interaction with correct spaces during forward and backward prop

Example of usage

class NeuralNetwork(zer0nn.NeuralNetwork):
    def __init__(self, input_nodes, hidden_nodes, output_nodes, learning_rate):
        super(NeuralNetwork, self).__init__(
                [(None, input_nodes), ("sig", hidden_nodes), ("lin", output_nodes)],
                "mse",
                learning_rate,
                .995)