network.cpp

也是遗传算法的源代码

/***************************************************************************
                          network.cpp  -  description
                             -------------------
    copyright            : (C) 2001, 2002 by Matt Grover
    email                : [e-mail]
 ***************************************************************************/

/***************************************************************************
 *                                                                         *
 *   This program is free software; you can redistribute it and/or modify  *
 *   it under the terms of the GNU General Public License as published by  *
 *   the Free Software Foundation; either version 2 of the License, or     *
 *   (at your option) any later version.                                   *
 *                                                                         *
 ***************************************************************************/

// Uncomment for debugging messages
//#define DEBUG_NETWORK

using namespace std;
#include "types.h"
#include <math.h>
#include <stdio.h>
#include <iostream>
#include <string>
#include <iterator>
#include <algorithm>
#include <vector>
#include <list>
#if GCC_VERSION >= 30000
    #include <ext/hash_map>
#else
    #include <hash_map>
#endif
#include "neuron.h"
#include "basicneuron.h"
#include "layer.h"
#include "network.h"
#include "simplespikeinput.h"
#include "functionlookup.h"
#include "utilities.h"

#ifdef DEBUG_NETWORK
#include <sys/time.h>
#endif

class SimpleSpikeInput;

Network::Network()
{
    // initialize the lookup tables
    SetDefaults();
    pspLRes = 100;  // unit = microseconds => 0.1 ms
    pspLSize = 1000;

    functionRef = new FunctionLookup();
    spikeInput = new SimpleSpikeInput(this);    // default input object
    netSize = 0;
    eventRequestCount = 0;
    isLayered = false;
    runCount++;
    trainingMode = true;
    maxSpikeDelay = 0;
    maxSpikeDelaySyn = 0;
    currSpikeDelayOffset = 0;
    maxOffset = 0;
    spikeDelaysOn = false;
    spikeBatchCount = 0;
}

void Network::SetDefaults()
{
    nextNeuronId = 1;
    nextLayerId = 1;
    streamingInput = false;
}

Network::~Network()
{
    // We don't know if the keys are contiguous
    // in net, so use an iterator to delete
    // Neurons instead of subscripting
    hash_map<AmIdInt, Neuron*>::iterator netItr;
    netItr = net.begin();
    while (netItr != net.end()) {
        delete netItr->second;
        netItr->second = 0;
        netItr++;
    }
    net.clear();
    runCount--;
    delete functionRef;
    delete spikeInput;

    for(layer_iterator layer = layers.begin(); layer != layers.end(); layer++){
        delete layer->second;
    }
}

// Initialize statics
AmTimeInt Network::simTime = 0;
unsigned int Network::runCount = 0;
AmTimeInt Network::simStepSize = 100;  // Default must match Neuron::simStepSize default!

void Network::ResetSimTime()
{
    // FIXME: Some checks need to be developed to make
    // sure that it is safe to reset simTime right now.
    // For example, simTime should not be reset if
    // any Networks are in the run loop.  This should
    // maybe be done as part of a more general reset routine
    // that clears out old data, etc.  A flag would have to be
    // set to indicate to other instances that they need to reset, also.
    simTime = 0;
}

void Network::SetTimeStepSize(AmTimeInt stepSize)
{
    // FIXME: Need a static flag to keep this from running after
    // neurons have been added to a net.
    simStepSize = stepSize;
    Neuron::simStepSize = stepSize;
}

Neuron* Network::AddNeuron(LayerType lType, AmIdInt nId)
{
    if (nId >= nextNeuronId) {
        nextNeuronId = nId + 1;
    }
    //cout << "Adding neuron " << nId << endl;
    BasicNeuron* nrn = new BasicNeuron(nId);
    net[nId] = nrn;

    nrn->SetTableDimensions(pspLSize, pspLRes);
    nrn->SetLookupTables(functionRef);

    nrn->SetLayerType(lType);
    nrn->SetParentNet(this);

    nrn->TrainingOn(trainingMode);
    return nrn;
}

Neuron* Network::AddNeuron(LayerType lType, Neuron* nrn)
{
    AmIdInt nId = 0;

    nId = nrn->GetID();
    if (nId >= nextNeuronId) {
        nextNeuronId = nId + 1;
    }

    //cout << "Adding neuron " << nId << endl;
    net[nId] = nrn;
    nrn->SetTableDimensions(pspLSize, pspLRes);
    nrn->SetLookupTables(functionRef);

    nrn->SetLayerType(lType);
    nrn->SetParentNet(this);
    nrn->TrainingOn(trainingMode);
    return nrn;
}

bool Network::ConnectNeurons(Neuron* preSynapticNeuron,
                                Neuron* postSynapticNeuron,
                                float weight,
                                AmTimeInt delay=0)
{
// FIXME: Sign enforcement is not working. There is confusion
//  as to when a neuron should be designated as inhibitory.
//  This code should maybe run in the neuron instead, and
//  a static function can set the enforceSign flag.
// NOTE: EnforceSign() has been made a static member of Neuron,
//  but this section of code has still not been tested. Remove
//  the fixme once this is confirmed to work correctly.
    if ( Neuron::EnforceSign() ) {
        // Make sure weight has correct sign
        if ( preSynapticNeuron->Inhibitory() ) {
            if (weight > 0.0) {
                cerr << "Inhibitory neuron " << preSynapticNeuron->GetID() <<
                    " has a positive weight!\n";
                return false;
            }
        }
        else if (weight < 0.0) {
            cerr << "Excitatory neuron " << preSynapticNeuron->GetID() <<
                " has a negative weight!\n";
            return false;
        }
    }

    try {
        Synapse* syn = new Synapse(postSynapticNeuron, weight, delay);
        preSynapticNeuron->AddSynapse(syn);

        if (delay > maxSpikeDelay) {
            maxSpikeDelay = delay;
            maxSpikeDelaySyn = syn;
        }
    }
    catch(string& e) {
        cerr << e << endl;
        return false;
    }
    catch(...) {
        cerr << "An error occured while connecting the neurons.\n";
        return false;
    }

    return true;
}

bool Network::ConnectNeurons(AmIdInt preSynapticNeuron,
                                AmIdInt postSynapticNeuron,
                                float weight,
                                AmTimeInt delay=0)
{
    return ConnectNeurons(net[preSynapticNeuron], net[postSynapticNeuron], weight, delay);
}

void Network::AddLayer(Layer* newLayer)
{
    unsigned int lId = newLayer->LayerId();

    if (!lId) {
        lId = nextLayerId;
        newLayer->SetLayerId(lId);
    }

    if (lId >= nextLayerId) {
        nextLayerId = lId + 1;
    }
    // we don't want duplicate Layer IDs
    for(layer_iterator layer = layers.begin(); layer != layers.end(); layer++){
        Layer *l = layer->second;
        if(l->LayerId() == lId) throw string("Layer ID: " + Utilities::itostr(lId) + " already in use");
    }

    layers[newLayer->LayerId()] = newLayer;
    newLayer->SetLayerParent(this);

    isLayered = true;
}

void Network::ScheduleNEvent(NEvent eventType,
                            AmTimeInt eventTime,
                            Neuron* reqNrn)
{
    SpikeRequest newSpike;
    if (eventType > RMSPIKE) {
        // initialize a new event request
        newSpike.requestTime = simTime;
        newSpike.requestor = reqNrn;
        newSpike.spikeTime = eventTime;
        newSpike.requestOrder = eventRequestCount++;

        // insert into the queue
        if (eventType == SPIKE || eventType == RESPIKE) {
            eventQ.push(newSpike);
        }
        else if (eventType == INPUTSPIKE) {
            inputQ.push(newSpike);
        }
    }
}

void Network::ScheduleNEvent(NEvent eventType,
                            AmTimeInt eventTime,
                            AmIdInt reqNrnId)
{
    Neuron* nrn = net.find(reqNrnId)->second;

    if (nrn) {
        ScheduleNEvent(eventType, eventTime, nrn);
    }
    else {
        string errMsg = "Neuron ID could not be found.";
        throw errMsg;
    }
}

void Network::Run(AmTimeInt maxRunTime)
{
    // This is the main loop
    // If streaming input is being used, keep on going until simTime >= maxRunTime.
    // Otherwise, run until the event queue is empty or
    // until simTime >= maxRunTime -- whichever happens first.
    bool stopRun = false;
    const unsigned int stopTime = maxRunTime + simTime;
    AmTimeInt nextInputTime = 0;

  	#ifdef DEBUG_NETWORK
  	timeval time1;
  	timeval time2;
    int totalTime;
    #endif

    if (simTime == 0)
        simTime = simStepSize;

    // Initialize the delayed spike queue if it has not been done.
    if (!delayedSpikeQ.size()) {
        InitializeDelayedSpikeQ();
    }

    if (eventQ.empty() && inputQ.empty()) {
        if(!streamingInput) {
            stopRun = true;
        }
        else {
            stopRun = false;
        }
    }
    else {
        if (!inputQ.empty()) {
            nextInputTime = inputQ.top().spikeTime;
        }
        else {
            nextInputTime = 0;
        }
        stopRun = false;
    }
    while (!stopRun) {
        while (!eventQ.empty()) {

            const SpikeRequest& topSpike = eventQ.top();
            if (topSpike.spikeTime != simTime)
                break;

            #ifdef DEBUG_NETWORK
            cout << "\nSending spike from Neuron: " << topSpike.requestor->GetID() << endl;
            gettimeofday(&time1, NULL);
            #endif

            topSpike.requestor->SendSpike(simTime);
            eventQ.pop();

            #ifdef DEBUG_NETWORK
            gettimeofday(&time2, NULL);
            totalTime = time2.tv_usec - time1.tv_usec;
            cout << "\nTotal time for sending spike: " << totalTime << "us" << endl;
            #endif
  	
        }
        if (streamingInput) {
            spikeInput->ReadInputBuffer();
            if (!inputQ.empty()) {
                nextInputTime = inputQ.top().spikeTime;
            }
            else {
                nextInputTime = 0;
            }
        }

        while (nextInputTime <= simTime) {
            if (nextInputTime == 0)
                break;

            inputQ.top().requestor->SendSpike(simTime);
            inputQ.pop();
            if (!inputQ.empty()) {
                nextInputTime = inputQ.top().spikeTime;
            }
            else {
                nextInputTime = 0;
            }
        }

        // send the delayed spikes
        // this must be called after Neuron::SendSpike()
        // has been called for the last time during this
        // time step
        if (spikeDelaysOn) {
            SendDelayedSpikes();
        }

        // increment simTime
        if (runCount > 1) {
            // A call to IncrementSimTime() is made if more than one
            // instance of Network is running in a process.  This is done
            // to keep all of the Networks synchronized and to handle any
            // threading issues that may arise due to simTime being static.
            IncrementSimTime();
        }
        else {
            simTime += simStepSize;
        }

        if (simTime >= stopTime) {
            stopRun = true;
        }
        else if (!streamingInput) {
            if (eventQ.empty() && inputQ.empty() && !spikeBatchCount) {
                stopRun = true;
            }
        }
    }
}

void Network::SetSpikeInput(SpikeInput* sIn)
{
    delete spikeInput;
    spikeInput = sIn;
}

void Network::IncrementSimTime()
{
    simTime += simStepSize;
}

void Network::SetTrainingMode(bool tMode)
{
    if (tMode == trainingMode) {
        return;
    }
    trainingMode = tMode;

    hash_map<AmIdInt, Neuron*>::iterator itr;
    itr = net.begin();

    while (itr != net.end()) {
        itr->second->TrainingOn(tMode);
        itr++;
    }
}

void Network::ScheduleSpikeDelay(vector<Synapse*>& axon)
{
    unsigned int maxOffset = delayedSpikeQ.size() - 1;
    for (vector<Synapse*>::iterator it=axon.begin(); it!=axon.end(); ++it) {
        AmTimeInt offset = (*it)->GetOffset();
        offset += currSpikeDelayOffset;

        // If the offset goes past the end of the queue, then
        // start back at the beginning
        if (offset > maxOffset) {
            offset -= (maxOffset + 1);
        }

        delayedSpikeQ[offset].push_back((*it));
    }
    spikeBatchCount += axon.size();
}


void Network::SendDelayedSpikes()
{
    vector<Synapse*>& spikeBatch = delayedSpikeQ[currSpikeDelayOffset];
    if (!spikeBatch.size()) {
        IncrementDelayOffset();
        return;
    }

    // sort the elements in the current offset batch
    // according to neuronId and get a pointer to
    // the first element
    vector<Synapse*>::iterator beginItr = spikeBatch.begin();
    vector<Synapse*>::iterator endItr = spikeBatch.end();
    static CompareSynapse comp;
    sort(beginItr, endItr, comp);
    //Synapse* firstSyn = spikeBatch[0];
    SynapseItr firstSyn = spikeBatch.begin();
    unsigned int numSyn = 0;

    // parse the spikeBatch vector and send one group of spikes
    // to each neuron in the vector.
    for (unsigned int i=0; i<spikeBatch.size(); ++i) {
        if (spikeBatch[i]->GetPostNeuron() != (*firstSyn)->GetPostNeuron()) {
            (*firstSyn)->GetPostNeuron()->InputSpike(firstSyn, simTime, numSyn);
            //firstSyn = spikeBatch[i];
            firstSyn += numSyn;
            numSyn = 1;
        }
        else {
            ++numSyn;
        }
    }

    // send the last batch of delayed spikes
    (*firstSyn)->GetPostNeuron()->InputSpike(firstSyn, simTime, numSyn);

    // clean up
    spikeBatchCount -= spikeBatch.size();
    spikeBatch.clear();
    IncrementDelayOffset();
}


void Network::InitializeDelayedSpikeQ()
{
    if (maxSpikeDelay) {
        spikeDelaysOn = true;
        Neuron::EnableSpikeBatching();
    }

    maxOffset = maxSpikeDelay/simStepSize;
    delayedSpikeQ.reserve(maxOffset+1);
    delayedSpikeQ.resize(maxOffset+1);

    for (unsigned int i=0; i<delayedSpikeQ.size(); ++i) {
        // TODO: Assuming 100 delayed spikes per offset for now.
        // A more intelligent algorithm to determine proper
        // sizing should be developed later on.
        delayedSpikeQ[i].reserve(100);
    }
}

inline void Network::IncrementDelayOffset()
{
    if (currSpikeDelayOffset >= maxOffset) {
        currSpikeDelayOffset = 0;
    }
    else {
        ++currSpikeDelayOffset;
    }
}

Layer * Network::GetLayer(AmIdInt layerId){
    return layers[layerId];
}