diff --git a/ChangeLog b/ChangeLog index 29ca1c0..c227769 100644 --- a/ChangeLog +++ b/ChangeLog @@ -1,21 +1,47 @@ -------------------------------------------------------------------------------- -0.2.2: -Added initXML(), XMLFromSet() and closeXML() methods to auto-generate training -XML. +--- 0.3 release --- + +2009-08-09 BlackLight + + * Makefile: Totally changed + * neural++.hpp: Changed header name, added BETA0 macro + * synapsis.cpp: Added momentum() method to compute the inertial momentum + of a synapsis + * everything: Data type changed from float to double for everything + +--- Release 0.2.2 --- + +2008-11-04 BlackLight + + * all: Added initXML(), XMLFromSet() and closeXML() methods to auto- + generate training XML. train() method can now get XML both from a file and from a simple string. -------------------------------------------------------------------------------- -0.2.1: -Added `train()` method to NeuralNet class, that allows you to train a neural -network using an XML containing pre-saved training sets (input values, -expected output). See examples/adder.xml for an example of such XML. -------------------------------------------------------------------------------- -0.2: -Added `save()` method to NeuralNet class, that allows you to save a trained -neural network to a binary file, to load anytime you need using -NeuralNet(const char *fname) constructor. -------------------------------------------------------------------------------- -0.01b: -First release ^^ -------------------------------------------------------------------------------- + +--- Release 0.2.1 --- + +2008-10-22 BlackLight + + * all: Added `train()` method to NeuralNet class, that allows you to + train a neural network using an XML containing pre-saved training + sets (input values, expected output). See examples/adder.xml for an + example of such XML. + +--- Release 0.2 --- + +2008-10-12 BlackLight + + * all: Added `save()` method to NeuralNet class, that allows you to + save a trained neural network to a binary file, to load anytime you + need using NeuralNet(const char *fname) constructor. + +--- Release 0.01b --- + +2008-04-03 BlackLight + + * all: First beta release ^^ + +Copyright 2008, 2009, BlackLight +Copying and distribution of this file, with or without modification, +are permitted provided the copyright notice and this notice are +preserved. diff --git a/VERSION b/VERSION index ee1372d..be58634 100644 --- a/VERSION +++ b/VERSION @@ -1 +1 @@ -0.2.2 +0.3 diff --git a/examples/Makefile b/examples/Makefile index 27cf000..b995a37 100644 --- a/examples/Makefile +++ b/examples/Makefile @@ -7,3 +7,4 @@ clean: rm learnAdd rm doAdd rm adderFromScratch + rm adder.net diff --git a/include/neural++.hpp b/include/neural++.hpp index ad4f2c3..217d17c 100644 --- a/include/neural++.hpp +++ b/include/neural++.hpp @@ -31,6 +31,9 @@ using namespace std; namespace neuralpp { //! Default rand value: |sin(rand)|, always >= 0 and <= 1 #define RAND ( (double) abs( sinf((double) rand()) ) ) + + //! Initial value for the inertial momentum of the synapses + #define BETA0 0.7 class Synapsis; class Neuron; @@ -221,42 +224,43 @@ namespace neuralpp { */ class Synapsis { double delta; + double prev_delta; double weight; Neuron *in; Neuron *out; - NeuralNet *net; double (*actv_f)(double); double (*deriv)(double); public: - Synapsis(Neuron* i, Neuron* o, NeuralNet* n, double w, double d) { - in=i; out=o; - weight=w; delta=d; - net=n; - } + /** + * @brief Constructor + * @param i Input neuron + * @param o Output neuron + * @param w Weight for the synapsis + * @param d Delta for the synapsis + */ + Synapsis(Neuron* i, Neuron* o, double w, double d); /** * @brief Constructor * @param i Input neuron * @param o Output neuron - * @param n Reference to the neural network * @param a Activation function * @param d Derivate for activation function */ - Synapsis (Neuron* i, Neuron* o, NeuralNet* n, double(*a)(double), double(*d)(double)); + Synapsis (Neuron* i, Neuron* o, double(*a)(double), double(*d)(double)); /** * @brief Constructor * @param i Input neuron * @param o Output neuron - * @param n Reference to the neural network * @param w Weight for the synapsis (default: random) * @param a Activation function * @param d Derivate for activation function */ - Synapsis (Neuron* i, Neuron* o, NeuralNet* n, + Synapsis (Neuron* i, Neuron* o, double w, double(*a)(double), double(*d)(double)); /** @@ -289,6 +293,24 @@ namespace neuralpp { * @return Delta of the synapsis */ double getDelta(); + + /** + * @brief Get the delta of the synapsis at the previous iteration + * @return The previous delta + */ + double getPrevDelta(); + + /** + * @brief Get the inertial momentum of a synapsis. This value is inversely proportional + * to the number of steps done in the learning phase (quite high at the beginning, decreasing + * to zero towards the end of the learning algorithm), and is needed to avoid strong + * oscillations in output values at the beginning, caused by the random values assigned to + * the synaptical weights + * @param N The number of iterations the network must have to adjust the output values + * @param x The number of iterations already taken + * @return The inertial momentum of the synapsis + */ + double momentum (int N, int x); }; /** @@ -386,7 +408,6 @@ namespace neuralpp { */ class Layer { vector< Neuron > elements; - NeuralNet *net; void (*update_weights)(); double (*actv_f)(double); @@ -396,17 +417,16 @@ namespace neuralpp { /** * @brief Constructor * @param sz Size of the layer - * @param n Reference to the neural network * @param a Activation function * @param d Its derivate */ - Layer (size_t sz, NeuralNet* n, double (*a)(double), double(*d)(double)); + Layer (size_t sz, double (*a)(double), double(*d)(double)); /** * @brief Alternative constructor. It directly gets a vector of neurons to build * the layer */ - Layer (vector< Neuron >&, NeuralNet* net, double(*a)(double), double(*d)(double)); + Layer (vector< Neuron >&, double(*a)(double), double(*d)(double)); /** * @brief Redefinition for operator []. It gets the neuron at i diff --git a/src/layer.cpp b/src/layer.cpp index 635bcf1..bce339a 100644 --- a/src/layer.cpp +++ b/src/layer.cpp @@ -21,7 +21,7 @@ using namespace neuralpp; * @param a Activation function * @param d Its derivate */ -Layer::Layer (size_t sz, NeuralNet* n, double(*a)(double), double(*d)(double)) { +Layer::Layer (size_t sz, double(*a)(double), double(*d)(double)) { for (size_t i=0; i &el, NeuralNet* n, double (*a)(double), double(*d)(double)) { +Layer::Layer (vector< Neuron > &el, double (*a)(double), double(*d)(double)) { elements=el; actv_f=a; deriv=d; - net=n; } /** @@ -65,7 +63,7 @@ void Layer::link (Layer& l) { for (size_t j=0; jpush_out(s); n2->push_in(s); diff --git a/src/neuralnet.cpp b/src/neuralnet.cpp index 18afe89..ac76130 100644 --- a/src/neuralnet.cpp +++ b/src/neuralnet.cpp @@ -43,9 +43,9 @@ NeuralNet::NeuralNet (size_t in_size, size_t hidden_size, size_t out_size, doubl actv_f=__actv; deriv=__deriv; - input = new Layer(in_size, this, __actv, __deriv); - hidden = new Layer(hidden_size, this, __actv, __deriv); - output = new Layer(out_size, this, __actv, __deriv); + input = new Layer(in_size, __actv, __deriv); + hidden = new Layer(hidden_size, __actv, __deriv); + output = new Layer(out_size, __actv, __deriv); link(); } @@ -70,9 +70,9 @@ NeuralNet::NeuralNet (size_t in_size, size_t hidden_size, size_t out_size, actv_f=a; deriv=d; - input = new Layer(in_size, this, a, d); - hidden = new Layer(hidden_size, this, a, d); - output = new Layer(out_size, this, a, d); + input = new Layer(in_size, a, d); + hidden = new Layer(hidden_size, a, d); + output = new Layer(out_size, a, d); link(); } @@ -153,7 +153,13 @@ void NeuralNet::updateWeights() { for (size_t j=0; jnIn(); j++) { Synapsis *s = &(n->synIn(j)); - out_delta = s->getIn()->getActv() * error(ex) * (-l_rate); + + if (ref_epochs - epochs > 0) + out_delta = s->getIn()->getActv() * error(ex) * (-l_rate) + + s->momentum(ref_epochs, ref_epochs-epochs) * s->getPrevDelta(); + else + out_delta = s->getIn()->getActv() * error(ex) * (-l_rate); + s->setDelta(out_delta); } } @@ -164,7 +170,12 @@ void NeuralNet::updateWeights() { for (size_t j=0; jnIn(); j++) { Synapsis *s = &(n->synIn(j)); - s->setDelta((-l_rate) * d * s->getIn()->getActv()); + + if (ref_epochs - epochs > 0) + s->setDelta((-l_rate) * d * s->getIn()->getActv() + + s->momentum(ref_epochs, ref_epochs-epochs) * s->getPrevDelta()); + else + s->setDelta((-l_rate) * d * s->getIn()->getActv()); } } } diff --git a/src/synapsis.cpp b/src/synapsis.cpp index fa4ee8e..0bef72d 100644 --- a/src/synapsis.cpp +++ b/src/synapsis.cpp @@ -15,67 +15,55 @@ #include "neural++.hpp" using namespace neuralpp; -Synapsis::Synapsis (Neuron* i, Neuron* o, NeuralNet* n, double(*a)(double), double(*d)(double)) { +Synapsis::Synapsis(Neuron* i, Neuron* o, double w, double d) { + in=i; out=o; + weight=w; + delta=d; prev_delta=0; +} + +Synapsis::Synapsis (Neuron* i, Neuron* o, double(*a)(double), double(*d)(double)) { srand((unsigned) time(NULL)); delta=0; + prev_delta=0; weight=RAND; in=i; out=o; actv_f=a; deriv=d; - net=n; } -/** - * @brief Constructor - * @param i Input neuron - * @param o Output neuron - * @param w Weight for the synapsis (default: random) - * @param a Activation function - * @param d Derivate for activation function - */ -Synapsis::Synapsis (Neuron* i, Neuron* o, NeuralNet* n, +Synapsis::Synapsis (Neuron* i, Neuron* o, double w, double(*a)(double), double(*d)(double)) { delta=0; + prev_delta=0; weight=w; in=i; out=o; actv_f=a; deriv=d; - net=n; } -/** - * @return Reference to input neuron of the synapsis - */ Neuron* Synapsis::getIn() { return in; } -/** - * @return Reference to output neuron of the synapsis - */ Neuron* Synapsis::getOut() { return out; } -/** - * @return Weight of the synapsis - */ double Synapsis::getWeight() { return weight; } -/** - * @return Delta of the synapsis - */ double Synapsis::getDelta() { return delta; } -/** - * @brief It sets the weight of the synapsis - */ +double Synapsis::getPrevDelta() { return prev_delta; } + void Synapsis::setWeight(double w) { weight=w; } -/** - * @brief It sets the delta (how much to change the weight after an update) - * of the synapsis - */ -void Synapsis::setDelta(double d) { delta=d; } +void Synapsis::setDelta(double d) { + prev_delta=delta; + delta=d; +} + +double Synapsis::momentum(int N, int x) { + return (BETA0*N)/(20*x + N); +}