Artificial Neural Networks made easy with the FANN library

/*
  Fast Artificial Neural Network Library (fann)
  Copyright (C) 2003 Steffen Nissen (lukesky@diku.dk)
  
  This library is free software; you can redistribute it and/or
  modify it under the terms of the GNU Lesser General Public
  License as published by the Free Software Foundation; either
  version 2.1 of the License, or (at your option) any later version.
  
  This library is distributed in the hope that it will be useful,
  but WITHOUT ANY WARRANTY; without even the implied warranty of
  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  Lesser General Public License for more details.
  
  You should have received a copy of the GNU Lesser General Public
  License along with this library; if not, write to the Free Software
  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
*/

#include <stdio.h>
#include <stdlib.h>
#include <stdarg.h>
#include <string.h>

#include "config.h"
#include "fann.h"
#include "fann_errno.h"

/*#define DEBUGTRAIN*/

#ifndef FIXEDFANN
/* INTERNAL FUNCTION
  Calculates the derived of a value, given an activation function
   and a steepness
*/
static fann_type fann_activation_derived(unsigned int activation_function,
	fann_type steepness, fann_type value)
{
	switch(activation_function){
		case FANN_LINEAR:
			return (fann_type)fann_linear_derive(steepness, value);
		case FANN_SIGMOID:
		case FANN_SIGMOID_STEPWISE:
			value = fann_clip(value, 0.01f, 0.99f);
			return (fann_type)fann_sigmoid_derive(steepness, value);
		case FANN_SIGMOID_SYMMETRIC:
		case FANN_SIGMOID_SYMMETRIC_STEPWISE:
			value = fann_clip(value, -0.98f, 0.98f);
			return (fann_type)fann_sigmoid_symmetric_derive(steepness, value);
		default:
			return 0;
	}
}

/* Trains the network with the backpropagation algorithm.
 */
FANN_EXTERNAL void FANN_API fann_train(struct fann *ann, fann_type *input, fann_type *desired_output)
{
	fann_run(ann, input);

	fann_compute_MSE(ann, desired_output);

	fann_backpropagate_MSE(ann);

	fann_update_weights(ann);
}
#endif

/* Tests the network.
 */
FANN_EXTERNAL fann_type * FANN_API fann_test(struct fann *ann, fann_type *input, fann_type *desired_output)
{
	fann_type neuron_value;
	fann_type *output_begin = fann_run(ann, input);
	fann_type *output_it;
	const fann_type *output_end = output_begin + ann->num_output;
	fann_type neuron_diff;

	/* calculate the error */
	for(output_it = output_begin;
		output_it != output_end; output_it++){
		neuron_value = *output_it;

		neuron_diff = (*desired_output - neuron_value);
		
		if(ann->activation_function_output == FANN_SIGMOID_SYMMETRIC ||
			ann->activation_function_output == FANN_SIGMOID_SYMMETRIC_STEPWISE){
			neuron_diff /= (fann_type)2;
		}
		
#ifdef FIXEDFANN
		ann->MSE_value += (neuron_diff/(float)ann->multiplier) * (neuron_diff/(float)ann->multiplier);
#else
		ann->MSE_value += (float)(neuron_diff * neuron_diff);
#endif
		
		desired_output++;
	}
	ann->num_MSE++;
	
	return output_begin;
}

/* get the mean square error.
   (obsolete will be removed at some point, use fann_get_MSE)
 */
FANN_EXTERNAL float FANN_API fann_get_error(struct fann *ann)
{
	return fann_get_MSE(ann);
}

/* get the mean square error.
 */
FANN_EXTERNAL float FANN_API fann_get_MSE(struct fann *ann)
{
	if(ann->num_MSE){
		return ann->MSE_value/(float)ann->num_MSE;
	}else{
		return 0;
	}
}

/* reset the mean square error.
   (obsolete will be removed at some point, use fann_reset_MSE)
 */
FANN_EXTERNAL void FANN_API fann_reset_error(struct fann *ann)
{
	fann_reset_MSE(ann);
}

/* reset the mean square error.
 */
FANN_EXTERNAL void FANN_API fann_reset_MSE(struct fann *ann)
{
	ann->num_MSE = 0;
	ann->MSE_value = 0;
}

#ifndef FIXEDFANN
/* INTERNAL FUNCTION
    compute the error at the network output
	(usually, after forward propagation of a certain input vector, fann_run)
	the error is a sum of squares for all the output units
	also increments a counter because MSE is an average of such errors

	After this train_errors in the output layer will be set to:
	neuron_value_derived * (desired_output - neuron_value)
 */
void fann_compute_MSE(struct fann *ann, fann_type *desired_output)
{
	fann_type neuron_value, neuron_diff, *error_it = 0, *error_begin = 0;
	struct fann_neuron *last_layer_begin = (ann->last_layer-1)->first_neuron;
	const struct fann_neuron *last_layer_end = last_layer_begin + ann->num_output;
	const struct fann_neuron *first_neuron = ann->first_layer->first_neuron;

	/* if no room allocated for the error variabels, allocate it now */
	if(ann->train_errors == NULL){
		ann->train_errors = (fann_type *)calloc(ann->total_neurons, sizeof(fann_type));
		if(ann->train_errors == NULL){
			fann_error((struct fann_error *)ann, FANN_E_CANT_ALLOCATE_MEM);
			return;
		}
	}
	
	/* clear the error variabels */
	memset(ann->train_errors, 0, (ann->total_neurons) * sizeof(fann_type));
	error_begin = ann->train_errors;
	
#ifdef DEBUGTRAIN
	printf("\ncalculate errors\n");
#endif
	/* calculate the error and place it in the output layer */
	error_it = error_begin + (last_layer_begin - first_neuron);

	for(; last_layer_begin != last_layer_end; last_layer_begin++){
		neuron_value = last_layer_begin->value;
		neuron_diff = *desired_output - neuron_value;

		if(ann->activation_function_output == FANN_SIGMOID_SYMMETRIC ||
			ann->activation_function_output == FANN_SIGMOID_SYMMETRIC_STEPWISE){
			neuron_diff /= 2.0;
		}
		
		ann->MSE_value += (float)(neuron_diff * neuron_diff);

		if(ann->train_error_function){ /* TODO make switch when more functions */
			if ( neuron_diff < -.9999999 )
				neuron_diff = -17.0;
			else if ( neuron_diff > .9999999 )
				neuron_diff = 17.0;
			else
				neuron_diff = (fann_type)log ( (1.0+neuron_diff) / (1.0-neuron_diff) );
		}
	
		*error_it = fann_activation_derived(ann->activation_function_output,
			ann->activation_steepness_output, neuron_value) * neuron_diff;

		
		desired_output++;
		error_it++;
	}
	ann->num_MSE++;
}

/* INTERNAL FUNCTION
   Propagate the error backwards from the output layer.

   After this the train_errors in the hidden layers will be:
   neuron_value_derived * sum(outgoing_weights * connected_neuron)
*/
void fann_backpropagate_MSE(struct fann *ann)
{
	fann_type neuron_value, tmp_error;
	unsigned int i;
	struct fann_layer *layer_it;
	struct fann_neuron *neuron_it, *last_neuron;
	
	fann_type *error_begin = ann->train_errors;
	fann_type *error_prev_layer;
	const fann_type activation_steepness_hidden = ann->activation_steepness_hidden;
	const struct fann_neuron *first_neuron = ann->first_layer->first_neuron;
	const struct fann_layer *second_layer = ann->first_layer + 1;
	struct fann_layer *last_layer = ann->last_layer;

	/* go through all the layers, from last to first.
	   And propagate the error backwards */
	for(layer_it = last_layer-1; layer_it > second_layer; --layer_it){
		last_neuron = layer_it->last_neuron;

		/* for each connection in this layer, propagate the error backwards*/
		if(ann->connection_rate >= 1 && !ann->shortcut_connections){
			/* optimization for fully connected networks */
			/* but not shortcut connected networks */
			error_prev_layer = error_begin + ((layer_it-1)->first_neuron - first_neuron);
			for(neuron_it = layer_it->first_neuron;
				neuron_it != last_neuron; neuron_it++){
				
				tmp_error = error_begin[neuron_it - first_neuron];
				for(i = neuron_it->num_connections ; i-- ; ){
					error_prev_layer[i] += tmp_error * neuron_it->weights[i];
				}
			}
		}else{
			for(neuron_it = layer_it->first_neuron;
				neuron_it != last_neuron; neuron_it++){
				
				tmp_error = error_begin[neuron_it - first_neuron];
				for(i = neuron_it->num_connections ; i-- ; ){
					error_begin[neuron_it->connected_neurons[i] - first_neuron] += tmp_error * neuron_it->weights[i];
				}
			}
		}

		/* then calculate the actual errors in the previous layer */
		error_prev_layer = error_begin + ((layer_it-1)->first_neuron - first_neuron);
		last_neuron = (layer_it-1)->last_neuron;
		
		switch(ann->activation_function_hidden){
			case FANN_LINEAR:
				for(neuron_it = (layer_it-1)->first_neuron;
					neuron_it != last_neuron; neuron_it++){
					neuron_value = neuron_it->value;
					*error_prev_layer *= (fann_type)fann_linear_derive(activation_steepness_hidden, neuron_value);
					error_prev_layer++;
				}
				break;
			case FANN_SIGMOID:
			case FANN_SIGMOID_STEPWISE:
				for(neuron_it = (layer_it-1)->first_neuron;
					neuron_it != last_neuron; neuron_it++){
					neuron_value = neuron_it->value;
					neuron_value = fann_clip(neuron_value, 0.01f, 0.99f);
					*error_prev_layer *= (fann_type)fann_sigmoid_derive(activation_steepness_hidden, neuron_value);
					error_prev_layer++;
				}
				break;
			case FANN_SIGMOID_SYMMETRIC:
			case FANN_SIGMOID_SYMMETRIC_STEPWISE:
				for(neuron_it = (layer_it-1)->first_neuron;
					neuron_it != last_neuron; neuron_it++){
					neuron_value = neuron_it->value;
					neuron_value = fann_clip(neuron_value, -0.98f, 0.98f);
					*error_prev_layer *= (fann_type)fann_sigmoid_symmetric_derive(activation_steepness_hidden, neuron_value);
					error_prev_layer++;
				}
				break;
			default:
				fann_error((struct fann_error *)ann, FANN_E_CANT_TRAIN_ACTIVATION);
				return;
		}
	}
}

/* INTERNAL FUNCTION
   Update weights for incremental training
*/
void fann_update_weights(struct fann *ann)
{
	struct fann_neuron *neuron_it, *last_neuron, *prev_neurons;
	fann_type tmp_error;
	struct fann_layer *layer_it;
	unsigned int i;
	
	/* store some variabels local for fast access */
	const float learning_rate = ann->learning_rate;
	const struct fann_neuron *first_neuron = ann->first_layer->first_neuron;
	struct fann_layer *first_layer = ann->first_layer;
	const struct fann_layer *last_layer = ann->last_layer;
	fann_type *error_begin = ann->train_errors;	

#ifdef DEBUGTRAIN
	printf("\nupdate weights\n");
#endif
	
	for(layer_it = (first_layer+1); layer_it != last_layer; layer_it++){
#ifdef DEBUGTRAIN
		printf("layer[%d]\n", layer_it - first_layer);
#endif
		last_neuron = layer_it->last_neuron;
		if(ann->connection_rate >= 1 && !ann->shortcut_connections){
			/* optimization for fully connected networks */
			/* but not shortcut connected networks */			
			prev_neurons = (layer_it-1)->first_neuron;
			for(neuron_it = layer_it->first_neuron;
				neuron_it != last_neuron; neuron_it++){
				tmp_error = error_begin[neuron_it - first_neuron] * learning_rate;
				for(i = neuron_it->num_connections ; i-- ; ){
					neuron_it->weights[i] += tmp_error * prev_neurons[i].value;
				}
			}
		}else{
			for(neuron_it = layer_it->first_neuron;
				neuron_it != last_neuron; neuron_it++){
				tmp_error = error_begin[neuron_it - first_neuron] * learning_rate;
				for(i = neuron_it->num_connections ; i-- ; ){
					neuron_it->weights[i] += tmp_error * neuron_it->connected_neurons[i]->value;
				}
			}
		}
	}
}

/* INTERNAL FUNCTION
   Update slopes for batch training
*/
void fann_update_slopes_batch(struct fann *ann)
{
	struct fann_neuron *neuron_it, *last_neuron, *prev_neurons;
	fann_type tmp_error, *weights_begin;
	struct fann_layer *layer_it;
	unsigned int i;
	
	/* store some variabels local for fast access */
	const struct fann_neuron *first_neuron = ann->first_layer->first_neuron;
	struct fann_layer *first_layer = ann->first_layer;
	const struct fann_layer *last_layer = ann->last_layer;
	fann_type *error_begin = ann->train_errors;
	fann_type *slope_begin, *neuron_slope;

	/* if no room allocated for the slope variabels, allocate it now */
	if(ann->train_slopes == NULL){
		ann->train_slopes = (fann_type *)calloc(ann->total_connections, sizeof(fann_type));
		if(ann->train_slopes == NULL){
			fann_error((struct fann_error *)ann, FANN_E_CANT_ALLOCATE_MEM);
			return;
		}
		memset(ann->train_slopes, 0, (ann->total_connections) * sizeof(fann_type));	
	}
	
	slope_begin = ann->train_slopes;
	weights_begin = fann_get_weights(ann);
	
#ifdef DEBUGTRAIN
	printf("\nupdate slopes\n");
#endif
	
	for(layer_it = (first_layer+1); layer_it != last_layer; layer_it++){
#ifdef DEBUGTRAIN
		printf("layer[%d]\n", layer_it - first_layer);
#endif
		last_neuron = layer_it->last_neuron;
		if(ann->connection_rate >= 1 && !ann->shortcut_connections){
			/* optimization for fully connected networks */
			/* but not shortcut connected networks */			
			prev_neurons = (layer_it-1)->first_neuron;
			for(neuron_it = layer_it->first_neuron;
				neuron_it != last_neuron; neuron_it++){
				tmp_error = error_begin[neuron_it - first_neuron];
				neuron_slope = slope_begin + (neuron_it->weights - weights_begin);
				for(i = neuron_it->num_connections ; i-- ; ){
					neuron_slope[i] += tmp_error * prev_neurons[i].value;
				}
			}
		}else{
			for(neuron_it = layer_it->first_neuron;
				neuron_it != last_neuron; neuron_it++){
				tmp_error = error_begin[neuron_it - first_neuron];
				neuron_slope = slope_begin + (neuron_it->weights - weights_begin);
				for(i = neuron_it->num_connections ; i-- ; ){
					neuron_slope[i] += tmp_error * neuron_it->connected_neurons[i]->value;
				}
			}
		}
	}
}

/* INTERNAL FUNCTION
   Clears arrays used for training before a new training session.
   Also creates the arrays that do not exist yet.
 */
void fann_clear_train_arrays(struct fann *ann)
{
	unsigned int i;
	
	/* if no room allocated for the slope variabels, allocate it now */
	if(ann->train_slopes == NULL){
		ann->train_slopes = (fann_type *)calloc(ann->total_connections, sizeof(fann_type));
		if(ann->train_slopes == NULL){
			fann_error((struct fann_error *)ann, FANN_E_CANT_ALLOCATE_MEM);
			return;
		}
	}
	memset(ann->train_slopes, 0, (ann->total_connections) * sizeof(fann_type));	
	/* if no room allocated for the variabels, allocate it now */
	if(ann->prev_steps == NULL){
		ann->prev_steps = (fann_type *)calloc(ann->total_connections, sizeof(fann_type));
		if(ann->prev_steps == NULL){
			fann_error((struct fann_error *)ann, FANN_E_CANT_ALLOCATE_MEM);
			return;
		}
	}
	memset(ann->prev_steps, 0, (ann->total_connections) * sizeof(fann_type));	
	
	/* if no room allocated for the variabels, allocate it now */
	if(ann->prev_train_slopes == NULL){
		ann->prev_train_slopes = (fann_type *)calloc(ann->total_connections, sizeof(fann_type));
		if(ann->prev_train_slopes == NULL){
			fann_error((struct fann_error *)ann, FANN_E_CANT_ALLOCATE_MEM);
			return;
		}
	}	

	if(ann->training_algorithm == FANN_TRAIN_RPROP){
		for(i = 0; i < ann->total_connections; i++){
			ann->prev_train_slopes[i] = (fann_type)0.0125;
		}
	} else {
		memset(ann->prev_train_slopes, 0, (ann->total_connections) * sizeof(fann_type));
	}
}

/* INTERNAL FUNCTION
   Update weights for batch training
 */
void fann_update_weights_batch(struct fann *ann, unsigned int num_data)
{
	fann_type *train_slopes = ann->train_slopes;
	fann_type *weights = fann_get_weights(ann);
	const float epsilon = ann->learning_rate/num_data;
	unsigned int i = ann->total_connections;
	while(i--){
		weights[i] += train_slopes[i] * epsilon;
		train_slopes[i] = 0.0;
	}
}

/* INTERNAL FUNCTION
   The quickprop training algorithm
 */
void fann_update_weights_quickprop(struct fann *ann, unsigned int num_data)
{
	fann_type *train_slopes = ann->train_slopes;
	fann_type *weights = fann_get_weights(ann);
	fann_type *prev_steps = ann->prev_steps;
	fann_type *prev_train_slopes = ann->prev_train_slopes;

	fann_type w, prev_step, slope, prev_slope, next_step;
	
	float epsilon = ann->learning_rate/num_data;
	float decay = ann->quickprop_decay; /*-0.0001;*/
	float mu = ann->quickprop_mu; /*1.75;*/
	float shrink_factor = (float)(mu / (1.0 + mu));

	unsigned int i = ann->total_connections;
	while(i--){
		w = weights[i];
		prev_step = prev_steps[i];
		slope = train_slopes[i] +  decay * w;
		prev_slope = prev_train_slopes[i];
		next_step = 0.0;
	
		/* The step must always be in direction opposite to the slope. */
	
		if(prev_step > 0.001) {
			/* If last step was positive...  */
			if(slope > 0.0) {
				/*  Add in linear term if current slope is still positive.*/
				next_step += epsilon * slope;
			}
		
			/*If current slope is close to or larger than prev slope...  */
			if(slope > (shrink_factor * prev_slope)) {
				next_step += mu * prev_step;      /* Take maximum size negative step. */
			} else {
				next_step += prev_step * slope / (prev_slope - slope); /* Else, use quadratic estimate. */
			}
		} else if(prev_step < -0.001){
			/* If last step was negative...  */  
			if(slope < 0.0){
				/*  Add in linear term if current slope is still negative.*/
				next_step += epsilon * slope;
			}
		
			/* If current slope is close to or more neg than prev slope... */
			if(slope < (shrink_factor * prev_slope)){
				next_step += mu * prev_step;      /* Take maximum size negative step. */
			} else {
				next_step += prev_step * slope / (prev_slope - slope); /* Else, use quadratic estimate. */
			}
		} else {
			/* Last step was zero, so use only linear term. */
			next_step += epsilon * slope;
		}


		/* update global data arrays */
		prev_steps[i] = next_step;
		weights[i] = w + next_step;
		prev_train_slopes[i] = slope;
		train_slopes[i] = 0.0;
	}
}

/* INTERNAL FUNCTION
   The iRprop- algorithm
*/
void fann_update_weights_irpropm(struct fann *ann, unsigned int num_data)
{
	fann_type *train_slopes = ann->train_slopes;
	fann_type *weights = fann_get_weights(ann);
	fann_type *prev_steps = ann->prev_steps;
	fann_type *prev_train_slopes = ann->prev_train_slopes;

	fann_type prev_step, slope, prev_slope, next_step, same_sign;

	/* These should be set from variables */
	float increase_factor = ann->rprop_increase_factor;/*1.2;*/
	float decrease_factor = ann->rprop_decrease_factor;/*0.5;*/
	float delta_min = ann->rprop_delta_min;/*0.0;*/
	float delta_max = ann->rprop_delta_max;/*50.0;*/

	unsigned int i = ann->total_connections;
	while(i--){	
		prev_step = fann_max(prev_steps[i], (fann_type)0.001); /* prev_step may not be zero because then the training will stop */
		slope = train_slopes[i];
		prev_slope = prev_train_slopes[i];
		next_step = 0.0;

		same_sign = prev_slope * slope;
	
		if(same_sign > 0.0) {
			next_step = fann_min(prev_step * increase_factor, delta_max);
		} else if(same_sign < 0.0) {
			next_step = fann_max(prev_step * decrease_factor, delta_min);
			slope = 0;
		}

		if(slope < 0){
			weights[i] -= next_step;
		}else{
			weights[i] += next_step;
		}

		/*if(i == 2){
			printf("weight=%f, slope=%f, next_step=%f, prev_step=%f\n", weights[i], slope, next_step, prev_step);
			}*/
	
		/* update global data arrays */
		prev_steps[i] = next_step;
		prev_train_slopes[i] = slope;
		train_slopes[i] = 0.0;
	}
}

#endif