void Server::saveState()
{
GDom doc;
doc.setRoot(serializeState(&doc));
char szStoragePath[300];
getStatePath(szStoragePath);
doc.saveJson(szStoragePath);
char szTime[256];
GTime::asciiTime(szTime, 256, false);
cout << "Server state saved at: " << szTime << "\n";
}
void Train()
{
printf("Learning the seed policy...\n");
GRand prng(0);
GMatrix features(0, FEATURE_DIMS);
GMatrix labels(0, LABEL_DIMS);
GenerateSeedTrainingSet(&features, &labels, &prng);
GNeuralNet nn(prng);
nn.addLayer(8);
nn.setActivationFunction(new GActivationBiDir(), true);
nn.train(features, labels);
nn.clipWeights(36.0); // ensure that the network is still somewhat malleable
int weightCount = nn.countWeights();
GTEMPBUF(double, hintVec, weightCount);
nn.weights(hintVec);
printf("Initializing a population for the evolutionary search...\n");
PolicyCritic critic(&nn, &prng, hintVec);
//GAssert(critic.Critique(hintVec) < -0.5, "bad critique");
//printf("HintVec Score: %lg\n", critic.Critique(hintVec));
GEvolutionaryOptimizer search(&critic, 15/*population*/, &prng, .93/*moreFitSurvivalProbability*/);
printf("Evolving better policies...\n");
char szBuf[64];
char szTime[64];
double err;
for(int i = 0; i <= 25000; i++)
{
err = search.iterate();
if(i % 250 == 0)
{
GDom doc;
nn.setWeights(search.currentVector());
doc.setRoot(nn.serialize(&doc));
sprintf(szBuf, "policy%d.json", i);
printf("Saving %s (height=%lg) at %s\n", szBuf, -err, GTime::asciiTime(szTime, 64, false));
doc.saveJson(szBuf);
}
}
}
void Train(GArgReader &args)
{
// Load series from file
std::unique_ptr<GMatrix> hSeries, hFeatures;
LoadData(args, hSeries);
GMatrix *pSeries = hSeries.get();
// Split features/labels
if(pSeries->cols() == 2)
{
GMatrix *pFeatures = pSeries->cloneSub(0, 0, pSeries->rows(), 1);
GMatrix *pLabels = pSeries->cloneSub(0, 1, pSeries->rows(), 1);
hFeatures.reset(pFeatures);
hSeries.reset(pLabels);
pSeries = pLabels;
}
else if(pSeries->cols() > 2)
{
throw Ex("Too many columns!");
}
// Parse options
GNeuralDecomposition *nd = new GNeuralDecomposition();
while(args.next_is_flag())
{
if(args.if_pop("-regularization"))
nd->setRegularization(args.pop_double());
else if(args.if_pop("-learningRate"))
nd->setLearningRate(args.pop_double());
else if(args.if_pop("-linearUnits"))
nd->setLinearUnits(args.pop_uint());
else if(args.if_pop("-softplusUnits"))
nd->setSoftplusUnits(args.pop_uint());
else if(args.if_pop("-sigmoidUnits"))
nd->setSigmoidUnits(args.pop_uint());
else if(args.if_pop("-epochs"))
nd->setEpochs(args.pop_uint());
else if(args.if_pop("-features"))
LoadData(args, hFeatures);
else if(args.if_pop("-filterLogarithm"))
nd->setFilterLogarithm(true);
else
throw Ex("Invalid option: ", args.peek());
}
if(hFeatures.get() == NULL)
{
// Generate features
GMatrix *pFeatures = new GMatrix(pSeries->rows(), 1);
for(size_t i = 0; i < pSeries->rows(); i++)
{
pFeatures->row(i)[0] = i / (double) pSeries->rows();
}
hFeatures.reset(pFeatures);
}
// Train
GMatrix *pFeatures = hFeatures.get();
nd->train(*pFeatures, *pSeries);
// Output the trained model
GDom doc;
doc.setRoot(nd->serialize(&doc));
doc.writeJson(cout);
}
void selfOrganizingMap(GArgReader& args){
// Load the file
GMatrix* pData = loadData(args.pop_string());
Holder<GMatrix> hData(pData);
// Parse arguments
std::vector<double> netDims;
unsigned numNodes = 1;
while(args.next_is_uint()){
unsigned dim = args.pop_uint();
netDims.push_back(dim);
numNodes *= dim;
}
if(netDims.size() < 1){
throw Ex("No dimensions specified for self organizing map. ",
"A map must be at least 1 dimensional.");
}
Holder<SOM::ReporterChain> reporters(new SOM::ReporterChain);
Holder<SOM::TrainingAlgorithm> alg(NULL);
Holder<GDistanceMetric> weightDist(new GRowDistance);
Holder<GDistanceMetric> nodeDist(new GRowDistance);
Holder<SOM::NodeLocationInitialization> topology(new SOM::GridTopology);
Holder<SOM::NodeWeightInitialization> weightInit
(new SOM::NodeWeightInitializationTrainingSetSample(NULL));
Holder<SOM::NeighborhoodWindowFunction>
windowFunc(new SOM::GaussianWindowFunction());
//Loading and saving
string loadFrom = "";
string saveTo = "";
//Parameters for different training algorithms
string algoName = "batch";
double startWidth = -1;//Start width - set later if still negative
double endWidth = -1;//End width - set later if still negative
double startRate = -1;//Start learning rate
double endRate = -1;//End learning rate
unsigned numIter = 100;//Total iterations
unsigned numConverge = 1;//#steps for batch to converge
while(args.next_is_flag()){
if(args.if_pop("-tofile")){
saveTo = args.pop_string();
}else if(args.if_pop("-fromfile")){
loadFrom = args.pop_string();
}else if(args.if_pop("-seed")){
GRand::global().setSeed(args.pop_uint());
}else if(args.if_pop("-neighborhood")){
string name = args.pop_string();
if(name == "gaussian"){
windowFunc.reset(new SOM::GaussianWindowFunction());
}else if(name == "uniform"){
windowFunc.reset(new SOM::UniformWindowFunction());
}else{
throw Ex("Only gaussian and uniform are acceptible ",
"neighborhood types");
}
}else if(args.if_pop("-printMeshEvery")){
using namespace SOM;
unsigned interval = args.pop_uint();
string baseFilename = args.pop_string();
unsigned xDim = args.pop_uint();
unsigned yDim = args.pop_uint();
bool showTrain = false;
if(args.if_pop("showTrain") || args.if_pop("showtrain")){
showTrain = true;
}
smart_ptr<Reporter> weightReporter
(new SVG2DWeightReporter(baseFilename, xDim, yDim, showTrain));
Holder<IterationIntervalReporter> intervalReporter
(new IterationIntervalReporter(weightReporter, interval));
reporters->add(intervalReporter.release());
}else if(args.if_pop("-batchTrain")){
algoName = "batch";
startWidth = args.pop_double();
endWidth = args.pop_double();
numIter = args.pop_uint();
numConverge = args.pop_uint();
}else if(args.if_pop("-stdTrain")){
algoName = "standard";
startWidth = args.pop_double();
endWidth = args.pop_double();
startRate = args.pop_double();
endRate = args.pop_double();
numIter = args.pop_uint();
}else{
throw Ex("Invalid option: ", args.peek());
}
}
//Create the training algorithm
Holder<SOM::TrainingAlgorithm> algo;
if(algoName == "batch"){
double netRadius = *std::max_element(netDims.begin(), netDims.end());
if(startWidth < 0){ startWidth = 2*netRadius; }
if(endWidth < 0){ endWidth = 1; }
algo.reset( new SOM::BatchTraining
(startWidth, endWidth, numIter, numConverge,
weightInit.release(), windowFunc.release(),
reporters.release()));
}else if(algoName == "standard"){
algo.reset( new SOM::TraditionalTraining
(startWidth, endWidth, startRate, endRate, numIter,
weightInit.release(), windowFunc.release(),
reporters.release()));
}else{
throw Ex("Unknown type of training algorithm: \"",
algoName, "\"");
}
//Create the network & transform the data
Holder<GSelfOrganizingMap> som;
Holder<GMatrix> out;
if(loadFrom == ""){
//Create map from arguments given
som.reset(new GSelfOrganizingMap
(netDims, numNodes, topology.release(), algo.release(),
weightDist.release(), nodeDist.release()));
//Train the network and transform the data in place
out.reset(som->doit(*pData));
}else{
//Create map from file
GDom source;
source.loadJson(loadFrom.c_str());
som.reset(new GSelfOrganizingMap(source.root()));
//Transform using the loaded network
out.reset(som->transformBatch(*pData));
}
//Save the trained network
if(saveTo != ""){
GDom serialized;
GDomNode* root = som->serialize(&serialized);
serialized.setRoot(root);
serialized.saveJson(saveTo.c_str());
}
//Print the result
out->print(cout);
}
void principalComponentAnalysis(GArgReader& args)
{
// Load the file
GMatrix* pData = loadData(args.pop_string());
Holder<GMatrix> hData(pData);
int nTargetDims = args.pop_uint();
// Parse options
string roundTrip;
unsigned int seed = getpid() * (unsigned int)time(NULL);
string eigenvalues;
string components;
string modelIn;
string modelOut;
bool aboutOrigin = false;
while(args.next_is_flag())
{
if(args.if_pop("-seed"))
seed = args.pop_uint();
else if(args.if_pop("-roundtrip"))
roundTrip = args.pop_string();
else if(args.if_pop("-eigenvalues"))
eigenvalues = args.pop_string();
else if(args.if_pop("-components"))
components = args.pop_string();
else if(args.if_pop("-aboutorigin"))
aboutOrigin = true;
else if(args.if_pop("-modelin"))
modelIn = args.pop_string();
else if(args.if_pop("-modelout"))
modelOut = args.pop_string();
else
throw Ex("Invalid option: ", args.peek());
}
// Transform the data
GRand prng(seed);
GPCA* pTransform = NULL;
if(modelIn.length() > 0)
{
GDom doc;
doc.loadJson(modelIn.c_str());
GLearnerLoader ll(prng);
pTransform = new GPCA(doc.root(), ll);
}
else
{
pTransform = new GPCA(nTargetDims, &prng);
if(aboutOrigin)
pTransform->aboutOrigin();
if(eigenvalues.length() > 0)
pTransform->computeEigVals();
pTransform->train(*pData);
}
Holder<GPCA> hTransform(pTransform);
GMatrix* pDataAfter = pTransform->transformBatch(*pData);
Holder<GMatrix> hDataAfter(pDataAfter);
// Save the eigenvalues
if(eigenvalues.length() > 0)
{
GArffRelation* pRelation = new GArffRelation();
pRelation->addAttribute("eigenvalues", 0, NULL);
sp_relation pRel = pRelation;
GMatrix dataEigenvalues(pRel);
dataEigenvalues.newRows(nTargetDims);
double* pEigVals = pTransform->eigVals();
for(int i = 0; i < nTargetDims; i++)
dataEigenvalues[i][0] = pEigVals[i];
dataEigenvalues.saveArff(eigenvalues.c_str());
}
// Save the components
if(components.length() > 0)
pTransform->components()->saveArff(components.c_str());
// Do the round-trip
if(roundTrip.size() > 0)
{
GMatrix roundTripped(pData->rows(), pData->cols());
for(size_t i = 0; i < pData->rows(); i++)
pTransform->untransform(pDataAfter->row(i), roundTripped.row(i));
roundTripped.saveArff(roundTrip.c_str());
}
if(modelOut.length() > 0)
{
GDom doc;
doc.setRoot(pTransform->serialize(&doc));
doc.saveJson(modelOut.c_str());
}
pDataAfter->print(cout);
}
void unsupervisedBackProp(GArgReader& args)
{
// Load the file and params
GMatrix* pData = loadData(args.pop_string());
Holder<GMatrix> hData(pData);
int targetDims = args.pop_uint();
// Parse Options
unsigned int nSeed = getpid() * (unsigned int)time(NULL);
GRand prng(nSeed);
GUnsupervisedBackProp* pUBP = new GUnsupervisedBackProp(targetDims, &prng);
Holder<GUnsupervisedBackProp> hUBP(pUBP);
vector<size_t> paramRanges;
string sModelOut;
string sProgress;
bool inputBias = true;
while(args.size() > 0)
{
if(args.if_pop("-seed"))
prng.setSeed(args.pop_uint());
else if(args.if_pop("-addlayer"))
pUBP->neuralNet()->addLayer(args.pop_uint());
else if(args.if_pop("-params"))
{
if(pUBP->jitterer())
throw Ex("You can't change the params after you add an image jitterer");
size_t paramDims = args.pop_uint();
for(size_t i = 0; i < paramDims; i++)
paramRanges.push_back(args.pop_uint());
}
else if(args.if_pop("-modelin"))
{
GDom doc;
doc.loadJson(args.pop_string());
GLearnerLoader ll(prng);
pUBP = new GUnsupervisedBackProp(doc.root(), ll);
hUBP.reset(pUBP);
}
else if(args.if_pop("-modelout"))
sModelOut = args.pop_string();
else if(args.if_pop("-intrinsicin"))
{
GMatrix* pInt = new GMatrix();
pInt->loadArff(args.pop_string());
pUBP->setIntrinsic(pInt);
}
else if(args.if_pop("-jitter"))
{
if(paramRanges.size() != 2)
throw Ex("The params must be set to 2 before a tweaker is set");
size_t channels = args.pop_uint();
double rot = args.pop_double();
double trans = args.pop_double();
double zoom = args.pop_double();
GImageJitterer* pJitterer = new GImageJitterer(paramRanges[0], paramRanges[1], channels, rot, trans, zoom);
pUBP->setJitterer(pJitterer);
}
else if(args.if_pop("-noinputbias"))
inputBias = false;
else if(args.if_pop("-progress"))
{
sProgress = args.pop_string();
pUBP->trackProgress();
}
else if(args.if_pop("-onepass"))
pUBP->onePass();
else
throw Ex("Invalid option: ", args.peek());
}
pUBP->setParams(paramRanges);
pUBP->setUseInputBias(inputBias);
// Transform the data
GMatrix* pDataAfter = pUBP->doit(*pData);
Holder<GMatrix> hDataAfter(pDataAfter);
pDataAfter->print(cout);
// Save the model (if requested)
if(sModelOut.length() > 0)
{
GDom doc;
doc.setRoot(pUBP->serialize(&doc));
doc.saveJson(sModelOut.c_str());
}
if(sProgress.length() > 0)
pUBP->progress().saveArff(sProgress.c_str());
}