train.cpp

几种神经网络的源程序

/**
  Train a MLP by a supplied GREN error network
  $Id: train.cpp,v 1.2 2004/06/16 12:32:21 opx Exp $
  @author OP
*/
#include <ctime> 
#include "annie/args.h"
#include "annie/random.h"
#include "annie/sys.h"
#include "annie/listeners.h"
#include "annie/MultiLayerNetwork.h"
#include <iomanip>
using namespace annie;
using namespace std;

static NumberParameter npars[]	=	{
                                     { "hidden1", "count of hidden neurons (0 = no first hidden layer) in the trained MLP", 0. },
                                     { "hidden2", "count of hidden neurons in second hidden layer (0=no layer) in the trained MLP", 0. },
                                     DEFAULT_MLP_NPARS
                                     { "verbose", "be a bit verbose", 0. },
                                     { "filter", "filter frequent changes", 0. },
                                     { "verboseX", "examples verbose", 0. },
                                     { NULL }
                                 };

enum { GREN, INPUTS, TRAINED };
static StringParameter spars[]	=	{
                                     { "gren", "file to load the trained GREN from", "../data/fromPolar.gren.net" },
                                     { "inputs", "file of training inputs", "../data/fromPolar.inputs.ts" },
                                     { "trained", "file to save the trained net to", "../data/fromPolar.trained.net" },
                                     { NULL }
                                 };
class TrainByGREN	{
public:
	enum { TRAINED_OFFSET = 5 };
	/// @param ts contains also the output pairs, but these are only used to compute real error
	TrainByGREN(MultiLayerNetwork &trainer, TrainingSet &ts, uint hidden1, uint hidden2) : net(ts.getInputSize(), TRAINED_OFFSET), gren(trainer), _ts(ts)	{
		ASSERT(!hidden2 || hidden1);
		if(hidden1)	net.addLayer(hidden1);
		if(hidden2)	net.addLayer(hidden2);
		uint outs = gren.getInputCount() - _ts.getInputSize();
		ASSERT(outs);
		if(_ts.getOutputSize())	ASSERT(outs == _ts.getOutputSize());
		net.addLayer(outs);	//out

		//full connection
		net.connectLayer(0);
		if(hidden1)	net.connectLayer(1);
		if(hidden2)	net.connectLayer(2);
		//cout << "initial net: " << net.verbose();
	}

	void train()	{
		net.getErrorGREN(_ts);
		net.trainGREN(gren, _ts, defaultControl["epochs"] , defaultControl["learningRate"] , defaultControl["momentum"] );
		defaultControl["real epoch error"] = -1;
		defaultControl["normalized real epoch error"] = -1;
		cout << " resulting error: \n";
		net.getErrorGREN(_ts);
	}

	void save(string &file)	{
		net.save(file);
	}

	const MultiLayerNetwork &getNet() const {
		return net;
	};
private:
	MultiLayerNetwork net, &gren;
	TrainingSet &_ts;
};

//// init/run stuff
int main(int argc, char *argv[])	{
	try {
		sysInit();
		parseArgs(argc, argv, npars, spars);
		//srand((unsigned)time(NULL));
		srand(1);	//const
		//srand(2);	//arith
		defaultControl["momentum"] = 0;
		SimpleVisualiser sv;
		RigidVisualiser rv;
		ValueUpdateListener *visualiser;
		if(defaultControl["filter"] ) visualiser = &sv;
		else visualiser = &rv;
		defaultControl.addListener(visualiser);
		cerr.setf(ios::fixed, ios::floatfield);
		cerr << std::setprecision(6);
		cout << "preparing" << endl;
		TrainingSet inputs(spars[INPUTS].value);
		cout << "loaded inputs: " << (string) inputs << endl;
		MultiLayerNetwork gren(spars[GREN].value);
		cout << "loaded GREN: " << (string) gren << endl;
		cout << gren.getLayer(1)[0] << endl;
		TrainByGREN tr(gren, inputs, defaultControl["hidden1"] , defaultControl["hidden2"] );
		cout << "MLP to train:: " << (string) tr.getNet() << endl;
		cout << "Training inputs: " << (string) inputs << endl;
		cout << "training " << endl;
		tr.train();
		cout << "saving to " << spars[TRAINED].value << endl;
		tr.save(spars[TRAINED].value);
		defaultControl.triggerAll();
	}	catch (Exception &e)	{
		cout << ">>> heh?\n" << e.what() << endl;
		exit(-1);
	}	catch (...)	{
		cout << ">>> heh? unknown exception\n";
		exit(-1);
	}

	return 0;
}