nickewing/netz

Clojure Neural Network Library

Netz - Clojure Neural Network Library

Description

Netz is a Clojure implementation of a multilayer perceptron (MLP), a type of feedforward artificial neural network. Netz provides functions for training and running MLPs. Training is accomplished via gradient descent batch Rprop or standard backpropagation.

Netz implements Rprop as described by Riedmiller in Rprop - Description and Implementation Details. A description of Netz's standard backpropagation algorithm can be found in docs/backpropagation.pdf.

Netz uses Incanter for matrix operations.

WARNING: Netz has had very little real world testing and training convergence is still slower than more sophisticated implementations. For production ready implementations, try:

• FANN (C, Java, Ruby, Python and many more...)
• Neuroph (Java)
• Encog (Java and CLR)

Usage

(ns your-namespace
  (:require [netz.core :as netz]))

(def examples [[[0 0] [1]]
               [[0 1] [0]]
               [[1 0] [0]]
               [[1 1] [1]]])

(def network (netz/train examples {:hidden-neurons [2]}))

(netz/run network [0 0]) ; => [0.9176]
(netz/run network [0 1]) ; => [0.0549]
(netz/run network [1 0]) ; => [0.0728]
(netz/run network [1 1]) ; => [0.9307]

Options

• :hidden-neurons - A vector containing the number of neurons in each hidden layer. Set to [2 2] for two hidden layers with two neurons each, or [] for no hidden layers. Setting this option is recommended. Default: One hidden layer with the same number of hidden neurons as inputs.

• :learning-algorithm - The algorithm to use while training. Choose either :rprop for the Rprop algorithm or :bprop for standard back propagation. Default: :rprop.

• :bprop

  • :learning-rate - The learning rate used while training with the standard backpropagation algorithm. Default: 0.25.
  • :learning-momentum - The learning momentum used while training with the standard backpropagation algorithm. Default: 0.0.

• :rprop

  • :init-update - Initial update value used for Rprop. Default: 0.1.
  • :update-min - Minimum update value used for Rprop. Default: 1e-6.
  • :update-max - Maximum update value used for Rprop. Default: 50.0.
  • :increase-factor - Increase factor for Rprop. Default: 1.2.
  • :decrease-factor - Decrease factor for Rprop. Default: 0.5.

• :regularization-constant - The regularization constant (lambda) used to penalize large weights. Default: 0.0.

• :callback - A callback function. If provided, Netz will call this function after every epoch of training. Returning false or nil from this callback will cause training to stop. See netz.core/report-callback for an example. Default: netz.core/report-callback.

• :callback-resolution - An integer specifying how often the callback function is invoked. Default: 100.

• :max-epochs - An integer specifying the maximum number of training epochs. Default: 20,000.

• :desired-error - A float specifying the desired training set mean squared error (MSE) used while training. Training will stop once the MSE drops below the desired error. Default: 0.005

• :calc-batch-error-in-parallel - Calculate example batch errors in parallel for each epoch. Default: true.

• :weight-initialization

  • :method - The weight initialization method. Choose either :random for randomly initialized weights or :nguyen-widrow to use the Nguyen-Widrow initialization method. Default: :nguyen-widrow.
  • :range - Randomly initialized weights will be between [-:weight-initialization-range .. :weight-initialization-range]. Default: 0.5.

License and Copyright

Netz is distributed under the MIT License. See LICENSE.

Copyright © 2012 Nick Ewing

netz's Issues

Incanter is slow...

Hi I played with matrix libraries for sometimes and Incanter is the slowest.
I obtained better result wrapping directly Jblas with an increment of performance about 20~50 times.