NeuralNetwork.cpp

/**
 * DESCRIPTION OF THE FILE
 *
 * @author Michal Kravčenko
 * @date 13.6.18 - 
 */

#include <iostream>
#include <boost/random/mersenne_twister.hpp>
#include <boost/random/uniform_real_distribution.hpp>
#include "NeuralNetwork.h"
#include "../NetConnection/ConnectionWeightIdentity.h"

NeuralNetwork::NeuralNetwork() {
    this->neurons = new std::vector<Neuron*>(0);
    this->connection_weights = new std::vector<double>(0);

    this->connection_weights->reserve(0);

    this->delete_weights = true;
}

NeuralNetwork* NeuralNetwork::get_subnet(std::vector<size_t> &input_neuron_indices, std::vector<size_t> &output_neuron_indices){
    NeuralNetwork *output_net = nullptr;

    Neuron * active_neuron, * target_neuron;

    size_t n = this->neurons->size();
    bool *part_of_subnetwork = new bool[n];
    std::fill(part_of_subnetwork, part_of_subnetwork + n, false);

    bool *is_reachable_from_source = new bool[n];
    bool *is_reachable_from_destination = new bool[n];
    std::fill(is_reachable_from_source, is_reachable_from_source + n, false);
    std::fill(is_reachable_from_destination, is_reachable_from_destination + n, false);

    bool *visited_neurons = new bool[n];
    std::fill(visited_neurons, visited_neurons + n, false);

    size_t active_set_size[2];
    active_set_size[0] = 0;
    active_set_size[1] = 0;
    size_t * active_neuron_set = new size_t[2 * n];
    size_t idx1 = 0, idx2 = 1;

    /* MAPPING BETWEEN NEURONS AND THEIR INDICES */
    size_t idx = 0, idx_target;
    for(Neuron *v: *this->neurons){
        v->set_idx( idx );
        idx++;
    }

    /* INITIAL STATE FOR THE FORWARD PASS */
    for(size_t i: input_neuron_indices ){

        if( i < 0 || i >= n){
            //invalid index
            continue;
        }
        active_neuron_set[idx1 * n + active_set_size[idx1]] = i;
        active_set_size[idx1]++;

        visited_neurons[i] = true;
    }

    /* FORWARD PASS */
    while(active_set_size[idx1] > 0){

        //we iterate through the active neurons and propagate the signal
        for(int i = 0; i < active_set_size[idx1]; ++i){
            idx = active_neuron_set[i];

            is_reachable_from_source[ idx ] = true;

            active_neuron = this->neurons->at( idx );

            for(Connection* connection: *(active_neuron->get_connections_out())){

                target_neuron = connection->get_neuron_out( );
                idx_target = target_neuron->get_idx( );

                if( visited_neurons[idx_target] ){
                    //this neuron was already analyzed
                    continue;
                }

                visited_neurons[idx_target] = true;
                active_neuron_set[active_set_size[idx2] + n * idx2] = idx_target;
                active_set_size[idx2]++;
            }
        }
        idx1 = idx2;
        idx2 = (idx1 + 1) % 2;
        active_set_size[idx2] = 0;
    }


    /* INITIAL STATE FOR THE FORWARD PASS */
    active_set_size[0] = active_set_size[1] = 0;
    std::fill(visited_neurons, visited_neurons + n, false);

    for(size_t i: output_neuron_indices ){

        if( i < 0 || i >= n){
            //invalid index
            continue;
        }
        active_neuron_set[idx1 * n + active_set_size[idx1]] = i;
        active_set_size[idx1]++;

        visited_neurons[i] = true;
    }

    /* BACKWARD PASS */
    size_t n_new_neurons = 0;
    while(active_set_size[idx1] > 0){

        //we iterate through the active neurons and propagate the signal
        for(int i = 0; i < active_set_size[idx1]; ++i){
            idx = active_neuron_set[i];

            is_reachable_from_destination[ idx ] = true;

            active_neuron = this->neurons->at( idx );

            if(is_reachable_from_source[ idx ]){
                n_new_neurons++;
            }

            for(Connection* connection: *(active_neuron->get_connections_in())){

                target_neuron = connection->get_neuron_in( );
                idx_target = target_neuron->get_idx( );

                if( visited_neurons[idx_target] ){
                    //this neuron was already analyzed
                    continue;
                }

                visited_neurons[idx_target] = true;
                active_neuron_set[active_set_size[idx2] + n * idx2] = idx_target;
                active_set_size[idx2]++;
            }
        }
        idx1 = idx2;
        idx2 = (idx1 + 1) % 2;
        active_set_size[idx2] = 0;
    }

    /* FOR CONSISTENCY REASONS */
    for(size_t in: input_neuron_indices){
        if( !is_reachable_from_destination[in] ){
            n_new_neurons++;
        }
        is_reachable_from_destination[in] = true;
    }
    /* FOR CONSISTENCY REASONS */
    for(size_t in: output_neuron_indices){
        if( !is_reachable_from_source[in] ){
            n_new_neurons++;
        }
        is_reachable_from_source[in] = true;
    }

    /* WE FORM THE SET OF NEURONS IN THE OUTPUT NETWORK  */
    if(n_new_neurons > 0){
//        printf("Number of new neurons: %d\n", n_new_neurons);
        output_net = new NeuralNetwork();
        output_net->set_weight_array( this->connection_weights );

        std::vector<size_t > local_inputs(0), local_outputs(0);
        local_inputs.reserve(input_neuron_indices.size());
        local_outputs.reserve(output_neuron_indices.size());

        std::vector<Neuron*> local_n_arr(0);
        local_n_arr.reserve( n_new_neurons );

        std::vector<Neuron*> local_local_n_arr(0);
        local_local_n_arr.reserve( n_new_neurons );

        int * neuron_local_mapping = new int[ n ];
        std::fill(neuron_local_mapping, neuron_local_mapping + n, -1);
        idx = 0;
        for(size_t i = 0; i < n; ++i){
            if(is_reachable_from_source[i] && is_reachable_from_destination[i]){
                neuron_local_mapping[i] = (int)idx;
                idx++;

                Neuron *new_neuron = this->neurons->at(i)->get_copy( );

                output_net->add_neuron( new_neuron );
                local_local_n_arr.push_back( new_neuron );
                local_n_arr.push_back( this->neurons->at(i) );
            }
        }
        for(size_t in: input_neuron_indices){
            local_inputs.push_back(neuron_local_mapping[in]);
        }
        for(size_t in: output_neuron_indices){
            local_outputs.push_back(neuron_local_mapping[in]);
        }

//        printf("%d\n", local_n_arr.size());
//        printf("inputs: %d, outputs: %d\n", local_inputs.size(), local_outputs.size());
        int local_idx_1, local_idx_2;
        for(Neuron* source_neuron: local_n_arr){
            //we also add the relevant edges
            local_idx_1 = neuron_local_mapping[source_neuron->get_idx()];

            for(Connection* connection: *(source_neuron->get_connections_out( ))){
                target_neuron = connection->get_neuron_out();

                local_idx_2 = neuron_local_mapping[target_neuron->get_idx()];
                if(local_idx_2 >= 0){
                    //this edge is part of the subnetwork
                    Connection* new_connection = connection->get_copy( local_local_n_arr[local_idx_1], local_local_n_arr[local_idx_2] );

                    local_local_n_arr[local_idx_1]->add_connection_out(new_connection);
                    local_local_n_arr[local_idx_2]->add_connection_in(new_connection);

//                    printf("adding a connection between neurons %d, %d\n", local_idx_1, local_idx_2);
                }

            }

        }
        output_net->specify_input_neurons(local_inputs);
        output_net->specify_output_neurons(local_outputs);


        delete [] neuron_local_mapping;
    }

    delete [] is_reachable_from_source;
    delete [] is_reachable_from_destination;
    delete [] part_of_subnetwork;
    delete [] visited_neurons;
    delete [] active_neuron_set;


    return output_net;
}


NeuralNetwork::~NeuralNetwork() {
    if(this->neurons){
        delete this->neurons;
        this->neurons = nullptr;
    }
    if(this->output_neurons){
        delete this->output_neurons;
        this->output_neurons = nullptr;
    }
    if(this->input_neurons){
        delete this->input_neurons;
        this->input_neurons = nullptr;
    }
    if(this->active_eval_set){
        delete this->active_eval_set;
        this->active_eval_set = nullptr;
    }

    if(this->connection_weights && this->delete_weights){
        delete this->connection_weights;
        this->connection_weights = nullptr;
    }
}

int NeuralNetwork::add_neuron(Neuron *n) {
    this->neurons->push_back(n);
    this->in_out_determined = false;
    n->set_idx(this->neurons->size() - 1);

    return n->get_idx();
}

size_t NeuralNetwork::add_connection_simple(int n1_idx, int n2_idx) {
    return add_connection_simple(n1_idx, n2_idx, -1);
}

size_t NeuralNetwork::add_connection_simple(int n1_idx, int n2_idx, size_t weight_idx) {
    return add_connection_simple(n1_idx, n2_idx, weight_idx, 1);
}

size_t NeuralNetwork::add_connection_simple(int n1_idx, int n2_idx, size_t weight_idx, double weight_value) {
    // TODO generate weight_value automatically from normal distribution

    if(weight_idx < 0 || weight_idx >= this->connection_weights->size()){
        //this weight is a new one, we add it to the system of weights
        this->connection_weights->push_back(weight_value);
        weight_idx = (int)this->connection_weights->size() - 1;
        this->n_weights++;
    }
    Neuron *neuron_out = this->neurons->at(n1_idx);
    Neuron *neuron_in = this->neurons->at(n2_idx);

    ConnectionWeightIdentity *con_weight_u1u2 = new ConnectionWeightIdentity(this->connection_weights);
    con_weight_u1u2->SetParamIndex(weight_idx, 0);

    Connection *u1u2 = new Connection(neuron_out, neuron_in, con_weight_u1u2);

    neuron_out->add_connection_out(u1u2);
    neuron_in->add_connection_in(u1u2);

    return weight_idx;
}

void NeuralNetwork::add_connection_shared_power1(int n1_idx, int n2_idx, size_t weight_idx) {
    std::function<double(double *, size_t*, size_t)> weight_function = [](double * weight_array, size_t * index_array, size_t n_params){
        return weight_array[index_array[0]];
    };

    size_t* weight_indices = new size_t[1];
    double* weight_values = new double[1];

    weight_indices[0] = weight_idx;
    this->add_connection_general( n1_idx, n2_idx, &weight_function, weight_indices, weight_values, 1);

    delete [] weight_indices;
    delete [] weight_values;
}

void NeuralNetwork::add_connection_shared_power2(int n1_idx, int n2_idx, size_t weight_idx) {
    std::function<double(double *, size_t*, size_t)> weight_function = [](double * weight_array, size_t * index_array, size_t n_params){
        return weight_array[index_array[0]] * weight_array[index_array[0]];
    };

    size_t* weight_indices = new size_t[1];
    double* weight_values = new double[1];

    weight_indices[0] = weight_idx;
    this->add_connection_general( n1_idx, n2_idx, &weight_function, weight_indices, weight_values, 1);

    delete [] weight_indices;
    delete [] weight_values;
}

void NeuralNetwork::add_connection_general(int n1_idx, int n2_idx, std::function<double(double *, size_t*, size_t)> *f,
                                           size_t* weight_indices, double* weight_values, size_t n_weights) {

    ConnectionWeight *con_weight_u1u2 = new ConnectionWeight(n_weights, this->connection_weights);
    //we analyze weights
    size_t weight_idx = 0;
    double weight_value = 0.0;
    for(size_t wi = 0; wi < n_weights; ++wi){
        weight_idx = weight_indices[wi];
        weight_value = weight_values[wi];

        if(weight_idx < 0 || weight_idx >= this->connection_weights->size()){
            //this weight is a new one, we add it to the system of weights
            this->connection_weights->push_back(weight_value);
            weight_indices[wi] = (int)this->connection_weights->size() - 1;
        }

        con_weight_u1u2->SetParamIndex(weight_indices[wi], wi);
    }

    Neuron *neuron_out = this->neurons->at(n1_idx);
    Neuron *neuron_in = this->neurons->at(n2_idx);
    Connection *u1u2 = new Connection(neuron_out, neuron_in, con_weight_u1u2);

    neuron_out->add_connection_out(u1u2);
    neuron_in->add_connection_in(u1u2);

}

void NeuralNetwork::determine_inputs_outputs() {
    if(this->output_neurons){
        delete this->output_neurons;
    }

    if(this->input_neurons){
        delete this->input_neurons;
    }

    this->output_neurons = new std::vector<Neuron*>(0);
    this->input_neurons = new std::vector<Neuron*>(0);

    if(this->active_eval_set == nullptr){
        this->active_eval_set = new std::vector<Neuron*>(this->neurons->size() * 2);
    }
    else{
        this->active_eval_set->resize(this->neurons->size() * 2);
    }

    for(Neuron* neuron: *this->neurons){
        if(neuron->get_connections_out()->empty()){
            //this neuron has no outgoing connections, it is the output neuron
            this->output_neurons->push_back(neuron);
        }
        else if(neuron->get_connections_in()->empty()){
            //this neuron has no incoming connections, it is the input neuron
            this->input_neurons->push_back(neuron);
        }
    }

    this->n_inputs = this->input_neurons->size();
    this->n_outputs = this->output_neurons->size();

    this->in_out_determined = true;
}

void NeuralNetwork::set_weight_array(std::vector<double> *weight_ptr) {
    if( this->connection_weights && this->delete_weights ){
        delete this->connection_weights;
    }
    this->connection_weights = weight_ptr;
    this->delete_weights = false;

    this->n_weights = this->connection_weights->size();
}

void NeuralNetwork::eval_single(std::vector<double> &input, std::vector<double> &output, double * custom_weights) {
    if(!this->in_out_determined && this->n_inputs * this->n_outputs <= 0){
//        this->determine_inputs_outputs();
        std::cerr << "Input and output neurons have not been specified\n" << std::endl;
        exit(-1);
    }


    if(this->n_inputs != input.size()){
        std::cerr << "Error, input size != Network input size\n" << std::endl;
        exit(-1);
    }

    if(this->n_outputs != output.size()){
        std::cerr << "Error, output size != Network output size\n" << std::endl;
        exit(-1);
    }

//    std::vector<double> *weight_ptr = this->connection_weights;
//    std::vector<double> *custom_weight_ptr = nullptr;
//
//
    if(custom_weights != nullptr){
//        custom_weight_ptr = new std::vector<double>(custom_weights, custom_weights + this->n_weights);
//        this->connection_weights = custom_weight_ptr;
        this->connection_weights->assign(custom_weights, custom_weights + this->n_weights);
    }



    std::fill(output.begin(), output.end(), 0.0);

    //reset of the potentials
    for(Neuron* neuron: *this->neurons){
        neuron->set_potential(0.0);
        neuron->set_saturation_in(0);
        neuron->set_saturation_out(0);
    }

    if(this->active_eval_set == nullptr){
        this->active_eval_set = new std::vector<Neuron*>(this->neurons->size() * 2);
    }
    else{
        this->active_eval_set->resize(this->neurons->size() * 2);
    }

    int active_set_size[2];
    active_set_size[0] = 0;
    active_set_size[1] = 0;

    int idx1 = 0, idx2 = 1;

    active_set_size[0] = this->n_inputs;
    int i = 0;
    auto n = this->neurons->size();

    for(Neuron* neuron: *this->input_neurons){

        this->active_eval_set->at(i) = neuron;

        neuron->set_potential(input[i]);

#ifdef VERBOSE_NN_EVAL
        printf("INPUT NEURON %d, POTENTIAL: %f\n", neuron, input[i]);
#endif
        ++i;
    }
    Neuron* active_neuron;
    Neuron* target_neuron;

    while(active_set_size[idx1] > 0){

        //we iterate through the active neurons and propagate the signal
        for(i = 0; i < active_set_size[idx1]; ++i){
            active_neuron = this->active_eval_set->at(i + n * idx1);
            active_neuron->activate();//computes the activation function on its potential
#ifdef VERBOSE_NN_EVAL
            printf(" active neuron %d, potential: %f, state: %f\n", active_neuron, active_neuron->get_potential(), active_neuron->get_state());
#endif
            for(Connection* connection: *(active_neuron->get_connections_out())){
                connection->pass_signal();

                target_neuron = connection->get_neuron_out();
                target_neuron->adjust_saturation_in(1);

                if(target_neuron->is_saturated_in()){
                    this->active_eval_set->at(active_set_size[idx2] + n * idx2) = target_neuron;
                    active_set_size[idx2]++;
                }
            }
        }
#ifdef VERBOSE_NN_EVAL
        printf("-----------\n");
#endif

        idx1 = idx2;
        idx2 = (idx1 + 1) % 2;

        active_set_size[idx2] = 0;
    }

    i = 0;
    for(Neuron* neuron: *this->output_neurons){

        output[i] = neuron->get_state();

#ifdef VERBOSE_NN_EVAL
        printf("OUTPUT NEURON %d, VALUE: %f\n", neuron, output[i]);
#endif

        ++i;
    }

//    this->connection_weights = weight_ptr;
}

void NeuralNetwork::copy_weights(double *weights) {
    for(unsigned int i = 0; i < this->connection_weights->size(); ++i){
        (*this->connection_weights)[i] = weights[i];
    }
}

std::vector<double>* NeuralNetwork::get_weights_ptr(){
    return this->connection_weights;
}

void NeuralNetwork::randomize_weights() {

    if( this->n_weights <= 0 ){

        return;
    }

    boost::random::mt19937 gen;

    // Init weight guess ("optimal" for logistic activation functions)
    double r = 4 * sqrt(6./(this->n_weights));

    boost::random::uniform_real_distribution<> dist(-r, r);

    for(size_t i = 0; i < this->n_weights; i++) {
        this->connection_weights->at(i) = dist(gen);
    }
}

size_t NeuralNetwork::get_n_inputs() {
    return this->n_inputs;
}

size_t  NeuralNetwork::get_n_outputs() {
    return this->n_outputs;
}

size_t NeuralNetwork::get_n_weights() {
    if(!this->n_weights) {
        this->n_weights = this->connection_weights->size();
    }
    return this->n_weights;
}

void NeuralNetwork::specify_input_neurons(std::vector<size_t> &input_neurons_indices) {
    if( !this->input_neurons ){
        this->input_neurons = new std::vector<Neuron*>();
    }
    this->input_neurons->reserve( input_neurons_indices.size() );

    for( size_t e: input_neurons_indices ){
        if( e < 0 || e >= this->neurons->size() ){
            continue;
        }
        this->input_neurons->push_back( this->neurons->at(e) );
    }
    this->n_inputs = this->input_neurons->size();
}

void NeuralNetwork::specify_output_neurons(std::vector<size_t> &output_neurons_indices) {
    if( !this->output_neurons ){
        this->output_neurons = new std::vector<Neuron*>();
    }
    this->output_neurons->reserve( output_neurons_indices.size() );

    for( size_t e: output_neurons_indices ){
        if( e < 0 || e >= this->neurons->size() ){
            continue;
        }
        this->output_neurons->push_back( this->neurons->at(e) );
    }
    this->n_outputs = this->output_neurons->size();
}

void NeuralNetwork::print_weights() {
    printf("Connection weights: ");
    if(this->connection_weights){
        for( size_t i = 0; i < this->connection_weights->size() - 1; ++i){
            printf("%f, ", this->connection_weights->at(i));
        }
        printf("%f", this->connection_weights->at(this->connection_weights->size() - 1));
    }

    printf("\n");
}

void NeuralNetwork::print_stats(){
    printf("Number of neurons: %d, number of weights: %d\n", this->neurons->size(), this->n_weights);
}