NearOCR

Class TrainingManager

java.lang.Object

NearOCR.TrainingManager

public abstract class TrainingManager
extends java.lang.Object

Handles all aspects relating to training techniques for both the encapsulated Neural Net and the symbol table.

Version:

0.1

Author:

Stephen Wattam

Constructor Summary

Constructors

Constructor and Description
TrainingManager()

Method Summary

Methods

Modifier and Type	Method and Description
static void	adjustSymbolWeights(NeuralNet[] nets, Symbol symbol, ProgressListener listener, ReceptorManager receptorManager) Adjusts all symbol weights in a single symbol.
static double	runSingleTrainingCycle(NeuralNet net, Symbol target, int iterations, ProgressListener listener, int netOffset, double entropy, ReceptorManager receptorManager) Runs a single training cycle for a single net and a single symbol.
static double	trainAllSymbols(NeuralNet net, Symbol[] symbols, int iterationsPerSymbol, int iterations, ProgressListener listener, double minMSE, int netOffset, double entropy, ReceptorManager receptorManager) Trains the net in a cyclic fashion with all symbols given.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Constructor Detail

TrainingManager

public TrainingManager()

Method Detail

adjustSymbolWeights

public static void adjustSymbolWeights(NeuralNet[] nets,
                       Symbol symbol,
                       ProgressListener listener,
                       ReceptorManager receptorManager)

Adjusts all symbol weights in a single symbol. Used to adapt a symbol table to a net stack.

Parameters:

nets - The whole net stack to run through, ordered.

symbol - The symbol whose output node weights are to be edited

listener - A ProgressListener which can watch the progress of the algorithm

trainAllSymbols

public static double trainAllSymbols(NeuralNet net,
                     Symbol[] symbols,
                     int iterationsPerSymbol,
                     int iterations,
                     ProgressListener listener,
                     double minMSE,
                     int netOffset,
                     double entropy,
                     ReceptorManager receptorManager)

Trains the net in a cyclic fashion with all symbols given. This takes only one net, not the entire stack, yet trains it with all chosen symbols.

Parameters:

net - The neural net to train

symbols - An unordered array of symbols

iterationsPerSymbol - How many times to train the net with each symbol. Best results are low, around 1 to 10.

iterations - The number of times to train the whole symbol set, best results are large, depending on training rate and how complex the symbol set is

listener - A ProgressListener that can watch the operation

minMSE - The mean standard error value at which to abort training. Set to 0 in order to guarantee completion of a full cycle.

netOffset - The offset in the output stack of neurons that point to the current net being run, for example, if net 0 on the stack had 23 outputs and we were training net 1 then this would be 23 - pointing to the first element which is owned by the current net. Getting this wrong can totally destroy the training process or shift symbols around.

Returns:

The average mean standard error for the training

runSingleTrainingCycle

public static double runSingleTrainingCycle(NeuralNet net,
                            Symbol target,
                            int iterations,
                            ProgressListener listener,
                            int netOffset,
                            double entropy,
                            ReceptorManager receptorManager)

Runs a single training cycle for a single net and a single symbol.

Parameters:

net - The neural net to train

target - The symbol to train with

iterations - The number of times to train

listener - A ProgressListener that can watch the operation

netOffset - The offset in the output stack of neurons that point to the current net being run, for example, if net 0 on the stack had 23 outputs and we were training net 1 then this would be 23 - pointing to the first element which is owned by the current net. Getting this wrong can totally destroy the training process or shift symbols around.

Returns:

The average mean standard error for the training