//layer2 plane0=0 "planes not both -1 and planes not both 1"
// weights = plane0*(-1) + plane1*(-1)
// plane1=1 "planes both -1 or planes both 1"
// weights = plane0*(1) + plane1*(1)
TEST( testlogicaloperators, Convolve_2layers_relu_Xor ) {
cout << "Xor, convolve" << endl;
// LogicalDataCreator ldc(new TanhActivation());
// ldc.applyXorGate();
// int imageSize = 1;
// int inPlanes = 2;
int numExamples = 4;
// int filterSize = 1;
float data[] = { -1, -1,
-1, 1,
1, -1,
1, 1 };
float layer1weights[] = { // going to preset these, to near an optimal solution,
// and at least show the network is stable, and gives the correct
-0.4f,-0.55f, // result...
0.52f, 0.53f,
};
float layer1bias[] = {
0.1f,
-0.1f
};
float layer2weights[] = {
1.1f, 0.9f,
-0.8f, -1.2f
};
float layer2bias[] = {
0.1f,
1.1
};
float expectedOutput[] = {
1, 0,
0, 1,
0, 1,
1, 0
};
int labels[] = {
0,
1,
1,
0
};
EasyCL *cl = EasyCL::createForFirstGpuOtherwiseCpu();
ClBlasInstance blasInstance;
NeuralNet *net = NeuralNet::maker(cl)->planes(2)->imageSize(1)->instance();
net->addLayer( ConvolutionalMaker::instance()->numFilters(2)->filterSize(1)->biased(1) );
net->addLayer( ActivationMaker::instance()->relu() );
net->addLayer( ConvolutionalMaker::instance()->numFilters(2)->filterSize(1)->biased(1) );
net->addLayer( ActivationMaker::instance()->relu() );
net->addLayer( SquareLossMaker::instance() );;
cout << "hand-setting weights..." << endl;
net->initWeights( 1, layer1weights, layer1bias );
net->initWeights( 3, layer2weights, layer2bias );
// net->printWeights();
// net->setBatchSize(4);
// net->forward( data );
// net->print();
SGD *sgd = SGD::instance( cl, 0.1f, 0 );
for( int epoch = 0; epoch < 200; epoch++ ) {
net->epochMaker(sgd)->batchSize(numExamples)->numExamples(numExamples)->inputData(data)
->expectedOutputs(expectedOutput)->run( epoch );
if( epoch % 5 == 0 ) cout << "Loss L " << net->calcLoss(expectedOutput) << endl;
}
net->print();
AccuracyHelper::printAccuracy( numExamples, 2, labels, net->getOutput() );
float loss = net->calcLoss(expectedOutput);
cout << "loss, E, " << loss << endl;
EXPECT_GE( 0.0000001f, loss );
delete sgd;
delete net;
delete cl;
}
TEST(testbackward, squareloss) {
// here's the plan:
// generate some input, randomly
// generate some expected output, randomly
// forward propagate
// calculate loss
// calculate gradInput
// change some of the inputs, forward prop, recalculate loss, check corresponds
// to the gradient
EasyCL *cl = EasyCL::createForFirstGpuOtherwiseCpu();
NeuralNet *net = new NeuralNet(cl, 3, 5);
net->addLayer(ForceBackpropLayerMaker::instance());
net->addLayer(SquareLossMaker::instance());
cout << net->asString() << endl;
int batchSize = 32;
net->setBatchSize(batchSize);
int inputCubeSize = net->getInputCubeSize();
int outputCubeSize = net->getOutputCubeSize();
int inputTotalSize = inputCubeSize * batchSize;
int outputTotalSize = outputCubeSize * batchSize;
cout << "inputtotalsize=" << inputTotalSize << " outputTotalSize=" << outputTotalSize << endl;
float *input = new float[inputTotalSize];
float *expectedOutput = new float[outputTotalSize];
WeightRandomizer::randomize(0, input, inputTotalSize, -2.0f, 2.0f);
WeightRandomizer::randomize(1, expectedOutput, outputTotalSize, -2.0f, 2.0f);
// now, forward prop
// net->input(input);
net->forward(input);
net->print();
// net->printOutput();
// calculate loss
float lossBefore = net->calcLoss(expectedOutput);
// calculate gradInput
net->backward(expectedOutput);
// modify input slightly
mt19937 random;
const int numSamples = 10;
for(int i = 0; i < numSamples; i++) {
int inputIndex;
WeightRandomizer::randomizeInts(i, &inputIndex, 1, 0, inputTotalSize);
// cout << "i=" << i << " index " << inputIndex << endl;
float oldValue = input[inputIndex];
// grad for this index is....
float grad = net->getLayer(2)->getGradInput()[inputIndex];
// cout << "grad=" << grad << endl;
// tweak slightly
float newValue = oldValue * 1.01f;
float inputDelta = newValue - oldValue;
float predictedLossChange = inputDelta * grad;
input[inputIndex] = newValue;
// cout << "oldvalue=" << oldValue << " newvalue=" << newValue << endl;
// forwardProp
net->forward(input);
input[inputIndex] = oldValue;
// net->printOutput();
float lossAfter = net->calcLoss(expectedOutput);
float lossChange = lossAfter - lossBefore;
cout << "idx=" << inputIndex << " predicted losschange=" << predictedLossChange << " actual=" << lossChange << endl;
}
delete[] expectedOutput;
delete[] input;
delete net;
delete cl;
}
void testNumerically(float learningRate, int batchSize, int imageSize, int filterSize, int numPlanes, ActivationFunction *fn, bool padZeros, int its = 20) {
EasyCL *cl = EasyCL::createForFirstGpuOtherwiseCpu();
ClBlasInstance clblasInstance;
NeuralNet *net = NeuralNet::maker(cl)->planes(numPlanes)->imageSize(imageSize)->instance();
net->addLayer(ConvolutionalMaker::instance()->numFilters(1)->filterSize(filterSize)->biased(0)->padZeros(padZeros));
net->addLayer(ActivationMaker::instance()->fn(fn));
net->addLayer(ConvolutionalMaker::instance()->numFilters(1)->filterSize(filterSize)->biased(0)->padZeros(padZeros));
net->addLayer(ActivationMaker::instance()->fn(fn));
net->addLayer(SquareLossMaker::instance());
net->setBatchSize(batchSize);
int inputNumElements = net->getLayer(0)->getOutputNumElements();
int outputNumElements = net->getLastLayer()->getOutputNumElements();
int weightsSize1 = net->getLayer(1)->getWeightsSize();
int weightsSize2 = net->getLayer(3)->getWeightsSize();
float *inputData = new float[std::max<int>(10000, inputNumElements)];
float *expectedOutput = new float[std::max<int>(10000, outputNumElements)];
memset(inputData, 0, sizeof(float) * std::max<int>(10000, inputNumElements));
memset(expectedOutput, 0, sizeof(float) * std::max<int>(10000, outputNumElements));
// int seed = 0;
std::mt19937 random = WeightRandomizer::randomize(inputData, std::max<int>(10000, inputNumElements), -2.0f, 2.0f);
WeightRandomizer::randomize(random, expectedOutput, std::max<int>(10000, outputNumElements), -2.0f, 2.0f);
WeightRandomizer::randomize(random, dynamic_cast<ConvolutionalLayer*>(net->getLayer(1))->weights, weightsSize1, -2.0f, 2.0f);
dynamic_cast<ConvolutionalLayer*>(net->getLayer(1))->weightsWrapper->copyToDevice();
WeightRandomizer::randomize(random, dynamic_cast<ConvolutionalLayer*>(net->getLayer(3))->weights, weightsSize2, -2.0f, 2.0f);
dynamic_cast<ConvolutionalLayer*>(net->getLayer(3))->weightsWrapper->copyToDevice();
SGD *sgd = SGD::instance(cl, learningRate, 0.0f);
for(int it = 0; it < its; it++) {
float *weightsBefore1 = new float[weightsSize1];
float *currentWeights = net->getLayer(1)->getWeights();
for(int i = 0; i < weightsSize1; i++) {
weightsBefore1[i] = currentWeights[i];
}
float *weightsBefore2 = new float[weightsSize2];
currentWeights = net->getLayer(3)->getWeights();
for(int i = 0; i < weightsSize2; i++) {
weightsBefore2[i] = currentWeights[i];
}
net->forward(inputData);
// net->print();
float loss = net->calcLoss(expectedOutput);
dynamic_cast<LossLayer*>(net->getLayer(5))->calcLoss(expectedOutput);
// net->backward(expectedOutput);
TrainingContext context(0, 0);
sgd->train(net, &context, inputData, expectedOutput);
dynamic_cast<ConvolutionalLayer*>(net->getLayer(1))->weightsWrapper->copyToHost();
// restore 2nd layer weights :-)
for(int i = 0; i < weightsSize2; i++) {
// dynamic_cast<ConvolutionalLayer*>(net->getLayer(2))->weights[i] = weightsBefore2[i];
}
dynamic_cast<ConvolutionalLayer*>(net->getLayer(3))->weightsWrapper->copyToDevice();
net->forward(inputData);
float loss2 = net->calcLoss(expectedOutput);
float lossChange = loss - loss2;
cout << " loss " << loss << " loss2 " << loss2 << " change: " << lossChange << endl;
float *newWeights = net->getLayer(1)->getWeights();
float sumWeightDiff = 0;
float sumWeightDiffSquared = 0;
for(int i = 0; i < weightsSize1; i++) {
float diff = newWeights[i] - weightsBefore1[i];
sumWeightDiff += diff;
sumWeightDiffSquared += diff * diff;
}
newWeights = net->getLayer(3)->getWeights();
for(int i = 0; i < weightsSize2; i++) {
float diff = newWeights[i] - weightsBefore2[i];
sumWeightDiff += diff;
sumWeightDiffSquared += diff * diff;
}
cout << "sumweightsdiff " << sumWeightDiff << endl;
// cout << "sumweightsdiff / learningrate " << (sumWeightDiff / learningRate) << endl;
// cout << "sum weightsdiffsquared " << (sumWeightDiffSquared/ learningRate / learningRate * imageSize) << endl;
float estimatedLossChangeFromW = sumWeightDiffSquared/ learningRate; // / filterSize;
cout << " loss change " << lossChange << endl;
cout << " estimatedLossChangeFromW " << estimatedLossChangeFromW << endl;
// cout << abs(estimatedLossChangeFromW - lossChange) / lossChange << endl;
// cout << abs(estimatedLossChangeFromW - lossChange) / estimatedLossChangeFromW << endl;
EXPECT_GT(0.01f * imageSize * imageSize, abs(estimatedLossChangeFromW - lossChange) / lossChange);
EXPECT_GT(0.01f * imageSize * imageSize, abs(estimatedLossChangeFromW - lossChange) / estimatedLossChangeFromW);
delete[] weightsBefore1;
delete[] weightsBefore2;
}
// delete[] weights1;
// delete[] errors;
// delete[] output;
delete sgd;
delete[] inputData;
delete[] expectedOutput;
delete net;
delete cl;
}
TEST(testbackward, softmaxloss) {
// here's the plan:
// generate some input, randomly
// generate some expected output, randomly
// forward propagate
// calculate loss
// calculate gradInput
// change some of the inputs, forward prop, recalculate loss, check corresponds
// to the gradient
EasyCL *cl = EasyCL::createForFirstGpuOtherwiseCpu();
NeuralNet *net = new NeuralNet(cl, 5, 1);
net->addLayer(ForceBackpropLayerMaker::instance());
net->addLayer(SoftMaxMaker::instance());
cout << net->asString() << endl;
const int batchSize = 2;
net->setBatchSize(batchSize);
const int outputPlanes = net->getOutputPlanes();
int inputCubeSize = net->getInputCubeSize();
int outputCubeSize = net->getOutputCubeSize();
int inputTotalSize = inputCubeSize * batchSize;
int outputTotalSize = outputCubeSize * batchSize;
cout << "inputtotalsize=" << inputTotalSize << " outputTotalSize=" << outputTotalSize << endl;
float *input = new float[inputTotalSize];
float *expectedOutput = new float[outputTotalSize];
WeightRandomizer::randomize(0, input, inputTotalSize, 0.0f, 1.0f);
WeightRandomizer::randomize(1, expectedOutput, outputTotalSize, 0.0f, 1.0f);
// we should make the input and output a probability distribution I think
// so: add up the input, and divide each by that. do same for expectedoutput (?)
// normalizeAsProbabilityDistribution(input, inputTotalSize);
normalizeAsProbabilityDistribution(outputPlanes, expectedOutput, outputTotalSize);
// set all to zero, and one to 1, ie like labelled data
// for(int i = 0; i < outputTotalSize; i++) {
// expectedOutput[i] = 0;
// }
// for(int n = 0; n < batchSize; n++) {
// int chosenLabel = 0;
// WeightRandomizer::randomizeInts(n, &chosenLabel, 1, 0, net->getOutputPlanes());
// expectedOutput[ n * outputPlanes + chosenLabel ] = 1;
// }
// for(int i = 0; i < outputTotalSize; i++) {
// cout << "expected[" << i << "]=" << expectedOutput[i] << endl;
// }
//
// now, forward prop
// net->input(input);
net->forward(input);
net->print();
// net->printOutput();
// calculate loss
float lossBefore = net->calcLoss(expectedOutput);
// calculate gradInput
net->backward(expectedOutput);
// modify input slightly
mt19937 random;
const int numSamples = 10;
for(int i = 0; i < numSamples; i++) {
int inputIndex;
WeightRandomizer::randomizeInts(i, &inputIndex, 1, 0, inputTotalSize);
// cout << "i=" << i << " index " << inputIndex << endl;
float oldValue = input[inputIndex];
// grad for this index is....
float grad = net->getLayer(2)->getGradInput()[inputIndex];
// cout << "grad=" << grad << endl;
// tweak slightly
float newValue = oldValue * 1.001f;
float inputDelta = newValue - oldValue;
float predictedLossChange = inputDelta * grad;
input[inputIndex] = newValue;
// cout << "oldvalue=" << oldValue << " newvalue=" << newValue << endl;
// forwardProp
net->forward(input);
input[inputIndex] = oldValue;
// net->printOutput();
float lossAfter = net->calcLoss(expectedOutput);
float lossChange = lossAfter - lossBefore;
cout << "idx=" << inputIndex << " predicted losschange=" << predictedLossChange << " actual=" << lossChange << endl;
}
delete[] expectedOutput;
delete[] input;
delete net;
delete cl;
}
int main()
{
// init variables
double error = 0.;
int truecnt = 0;
int times,timed;
// print useful info for reference
std::cout << "\n" << "hidden neurons: " << "\t \t" << HIDDEN << std::endl;
// init random number generator
srand((int)time(NULL));
// create network
std::cout << "initializing network..." << "\t \t";
NeuralNet DigitNet;
NeuralLayer * pHiddenLayer1 = new NeuralTanhLayer(INPUT,HIDDEN);
DigitNet.addLayer( pHiddenLayer1 );
NeuralLayer * pOutputLayer = new NeuralSoftmaxLayer(HIDDEN,OUTPUT);
DigitNet.addLayer( pOutputLayer );
// set output type:
// SCALAR = tanh or sigmoid output layer (use one output neuron)
// PROB = softmax output layer, 1-of-N output encoding (use two output neurons)
const unsigned int outType = PROB;
// set learning rate, momentum, decay rate
const double learningRate = 0.15;
const double momentum = 0.0;
const double decayRate = 0.0;
DigitNet.setParams(learningRate,momentum,decayRate,outType);
std::cout << "done" << std::endl;
// load training and test data
std::cout << "loading data..." << "\t \t \t";
std::vector< std::vector<double> > bigData( DATA_SIZE,std::vector<double>(INPUT+1,0.0) );
loadFromFile(bigData,"train.txt");
std::vector< std::vector<double> > trainData( TRAIN_SIZE,std::vector<double>(INPUT+1,0.0) );
std::vector< std::vector<double> > testData( TEST_SIZE,std::vector<double>(INPUT+1,0.0) );
buildData(bigData,trainData,TRAIN_SIZE,testData,TEST_SIZE);
std::cout << "done" << std::endl;
// loop over training data points and train net
// slice off first column of each row (example)
times=(int)time(NULL); // init time counter
std::cout << "\n" << "training examples: " << "\t \t" << TRAIN_SIZE << std::endl;
std::cout << "learning rate: " << "\t \t \t" << learningRate << std::endl;
std::cout << "momentum: " << "\t \t \t" << momentum << std::endl;
std::cout << "weight decay: " << "\t \t \t" << decayRate << std::endl;
std::cout << "training network..." << "\t \t";
for(int i=0;i<TRAIN_SIZE;++i)
{
std::vector<double> data = trainData[i]; // extract data point
double label = data[0]; // extract point label
data.erase(data.begin());
std::vector<double> nLabel = encode((int)label); // encode to 1-of-N
std::vector<double> outputs = DigitNet.runNet(data);
error = DigitNet.trainNet(data,nLabel,outType); // train net, return MSE
// decode output and compare to correct output
if( decode(outputs) == (int)label )
truecnt++;
}
// stop timer and print out useful info
timed=(int)time(NULL);
times=timed-times;
std::cout << "done" << std::endl;
std::cout << "training time: " << "\t \t \t" << times << " seconds " << std::endl;
std::cout << "training accuracy: " << "\t \t" << truecnt*100./TRAIN_SIZE << "%" << std::endl;
// test net on test data
times=(int)time(NULL); // init time counter
std::cout << "\n" << "test points: " << "\t \t \t" << TEST_SIZE << std::endl;
std::cout << "testing network..." << "\t \t";
truecnt = 0;
for(int i=0;i<TEST_SIZE;++i)
{
std::vector<double> data = testData[i]; // extract data point
double label = data[0]; // extract label
data.erase(data.begin());
std::vector<double> outputs = DigitNet.runNet(data); // run net
// decode output and compare to correct output
if( decode(outputs) == (int)label )
truecnt++;
}
// stop timer and print out useful info
timed=(int)time(NULL);
times=timed-times;
std::cout << "done" << std::endl;
std::cout << "testing time: " << "\t \t \t" << times << " seconds " << std::endl;
std::cout << "test accuracy: " << "\t \t \t" << truecnt*100./TEST_SIZE << "% " << std::endl;
// save weights to reuse net in the future
DigitNet.saveNet();
}
void go(Config config) {
Timer timer;
int Ntrain;
int Ntest;
int numPlanes;
int imageSize;
unsigned char *trainData = 0;
unsigned char *testData = 0;
int *trainLabels = 0;
int *testLabels = 0;
int trainAllocateN = 0;
int testAllocateN = 0;
// int totalLinearSize;
GenericLoader::getDimensions( config.dataDir + "/" + config.trainFile, &Ntrain, &numPlanes, &imageSize );
Ntrain = config.numTrain == -1 ? Ntrain : config.numTrain;
// long allocateSize = (long)Ntrain * numPlanes * imageSize * imageSize;
cout << "Ntrain " << Ntrain << " numPlanes " << numPlanes << " imageSize " << imageSize << endl;
if( config.loadOnDemand ) {
trainAllocateN = config.batchSize; // can improve this later
} else {
trainAllocateN = Ntrain;
}
trainData = new unsigned char[ (long)trainAllocateN * numPlanes * imageSize * imageSize ];
trainLabels = new int[trainAllocateN];
if( !config.loadOnDemand && Ntrain > 0 ) {
GenericLoader::load( config.dataDir + "/" + config.trainFile, trainData, trainLabels, 0, Ntrain );
}
GenericLoader::getDimensions( config.dataDir + "/" + config.validateFile, &Ntest, &numPlanes, &imageSize );
Ntest = config.numTest == -1 ? Ntest : config.numTest;
if( config.loadOnDemand ) {
testAllocateN = config.batchSize; // can improve this later
} else {
testAllocateN = Ntest;
}
testData = new unsigned char[ (long)testAllocateN * numPlanes * imageSize * imageSize ];
testLabels = new int[testAllocateN];
if( !config.loadOnDemand && Ntest > 0 ) {
GenericLoader::load( config.dataDir + "/" + config.validateFile, testData, testLabels, 0, Ntest );
}
cout << "Ntest " << Ntest << " Ntest" << endl;
timer.timeCheck("after load images");
const int inputCubeSize = numPlanes * imageSize * imageSize;
float translate;
float scale;
int normalizationExamples = config.normalizationExamples > Ntrain ? Ntrain : config.normalizationExamples;
if( !config.loadOnDemand ) {
if( config.normalization == "stddev" ) {
float mean, stdDev;
NormalizationHelper::getMeanAndStdDev( trainData, normalizationExamples * inputCubeSize, &mean, &stdDev );
cout << " image stats mean " << mean << " stdDev " << stdDev << endl;
translate = - mean;
scale = 1.0f / stdDev / config.normalizationNumStds;
} else if( config.normalization == "maxmin" ) {
float mean, stdDev;
NormalizationHelper::getMinMax( trainData, normalizationExamples * inputCubeSize, &mean, &stdDev );
translate = - mean;
scale = 1.0f / stdDev;
} else {
cout << "Error: Unknown normalization: " << config.normalization << endl;
return;
}
} else {
if( config.normalization == "stddev" ) {
float mean, stdDev;
NormalizeGetStdDev<unsigned char> normalizeGetStdDev( trainData, trainLabels );
BatchProcess::run<unsigned char>( config.dataDir + "/" + config.trainFile, 0, config.batchSize, normalizationExamples, inputCubeSize, &normalizeGetStdDev );
normalizeGetStdDev.calcMeanStdDev( &mean, &stdDev );
cout << " image stats mean " << mean << " stdDev " << stdDev << endl;
translate = - mean;
scale = 1.0f / stdDev / config.normalizationNumStds;
} else if( config.normalization == "maxmin" ) {
NormalizeGetMinMax<unsigned char> normalizeGetMinMax( trainData, trainLabels );
BatchProcess::run( config.dataDir + "/" + config.trainFile, 0, config.batchSize, normalizationExamples, inputCubeSize, &normalizeGetMinMax );
normalizeGetMinMax.calcMinMaxTransform( &translate, &scale );
} else {
cout << "Error: Unknown normalization: " << config.normalization << endl;
return;
}
}
cout << " image norm translate " << translate << " scale " << scale << endl;
timer.timeCheck("after getting stats");
// const int numToTrain = Ntrain;
// const int batchSize = config.batchSize;
NeuralNet *net = new NeuralNet();
// net->inputMaker<unsigned char>()->numPlanes(numPlanes)->imageSize(imageSize)->insert();
net->addLayer( InputLayerMaker<unsigned char>::instance()->numPlanes(numPlanes)->imageSize(imageSize) );
net->addLayer( NormalizationLayerMaker::instance()->translate(translate)->scale(scale) );
if( !NetdefToNet::createNetFromNetdef( net, config.netDef ) ) {
return;
}
net->print();
bool afterRestart = false;
int restartEpoch = 0;
int restartBatch = 0;
float restartAnnealedLearningRate = 0;
int restartNumRight = 0;
float restartLoss = 0;
if( config.loadWeights && config.weightsFile != "" ) {
afterRestart = WeightsPersister::loadWeights( config.weightsFile, config.getTrainingString(), net, &restartEpoch, &restartBatch, &restartAnnealedLearningRate, &restartNumRight, &restartLoss );
if( !afterRestart && FileHelper::exists( config.weightsFile ) ) {
cout << "Weights file " << config.weightsFile << " exists, but doesnt match training options provided => aborting" << endl;
cout << "Please either check the training options, or choose a weights file that doesnt exist yet" << endl;
return;
}
}
timer.timeCheck("before learning start");
if( config.dumpTimings ) {
StatefulTimer::dump( true );
}
StatefulTimer::timeCheck("START");
Trainable *trainable = net;
MultiNet *multiNet = 0;
if( config.multiNet > 1 ) {
multiNet = new MultiNet( config.multiNet, net );
trainable = multiNet;
}
if( config.loadOnDemand ) {
NetLearnerOnDemand<unsigned char> netLearner( trainable );
netLearner.setTrainingData( config.dataDir + "/" + config.trainFile, Ntrain );
netLearner.setTestingData( config.dataDir + "/" + config.validateFile, Ntest );
netLearner.setSchedule( config.numEpochs, afterRestart ? restartEpoch : 1 );
netLearner.setBatchSize( config.fileReadBatches, config.batchSize );
netLearner.setDumpTimings( config.dumpTimings );
WeightsWriter weightsWriter( net, &config );
if( config.weightsFile != "" ) {
netLearner.addPostEpochAction( &weightsWriter );
}
netLearner.learn( config.learningRate, config.annealLearningRate );
} else {
NetLearner<unsigned char> netLearner( trainable );
netLearner.setTrainingData( Ntrain, trainData, trainLabels );
netLearner.setTestingData( Ntest, testData, testLabels );
netLearner.setSchedule( config.numEpochs, afterRestart ? restartEpoch : 1 );
netLearner.setBatchSize( config.batchSize );
netLearner.setDumpTimings( config.dumpTimings );
WeightsWriter weightsWriter( net, &config );
if( config.weightsFile != "" ) {
netLearner.addPostEpochAction( &weightsWriter );
}
netLearner.learn( config.learningRate, config.annealLearningRate );
}
if( multiNet != 0 ) {
delete multiNet;
}
delete net;
if( trainData != 0 ) {
delete[] trainData;
}
if( testData != 0 ) {
delete[] testData;
}
if( testLabels != 0 ) {
delete[] testLabels;
}
if( trainLabels != 0 ) {
delete[] trainLabels;
}
}
void go(Config config) {
Timer timer;
int Ntrain;
int Ntest;
int numPlanes;
int imageSize;
float *trainData = 0;
float *testData = 0;
int *trainLabels = 0;
int *testLabels = 0;
int trainAllocateN = 0;
int testAllocateN = 0;
// int totalLinearSize;
GenericLoaderv2 trainLoader( config.dataDir + "/" + config.trainFile );
Ntrain = trainLoader.getN();
numPlanes = trainLoader.getPlanes();
imageSize = trainLoader.getImageSize();
// GenericLoader::getDimensions( , &Ntrain, &numPlanes, &imageSize );
Ntrain = config.numTrain == -1 ? Ntrain : config.numTrain;
// long allocateSize = (long)Ntrain * numPlanes * imageSize * imageSize;
cout << "Ntrain " << Ntrain << " numPlanes " << numPlanes << " imageSize " << imageSize << endl;
if( config.loadOnDemand ) {
trainAllocateN = config.batchSize; // can improve this later
} else {
trainAllocateN = Ntrain;
}
trainData = new float[ (long)trainAllocateN * numPlanes * imageSize * imageSize ];
trainLabels = new int[trainAllocateN];
if( !config.loadOnDemand && Ntrain > 0 ) {
trainLoader.load( trainData, trainLabels, 0, Ntrain );
}
GenericLoaderv2 testLoader( config.dataDir + "/" + config.validateFile );
Ntest = testLoader.getN();
numPlanes = testLoader.getPlanes();
imageSize = testLoader.getImageSize();
Ntest = config.numTest == -1 ? Ntest : config.numTest;
if( config.loadOnDemand ) {
testAllocateN = config.batchSize; // can improve this later
} else {
testAllocateN = Ntest;
}
testData = new float[ (long)testAllocateN * numPlanes * imageSize * imageSize ];
testLabels = new int[testAllocateN];
if( !config.loadOnDemand && Ntest > 0 ) {
testLoader.load( testData, testLabels, 0, Ntest );
}
cout << "Ntest " << Ntest << " Ntest" << endl;
timer.timeCheck("after load images");
const int inputCubeSize = numPlanes * imageSize * imageSize;
float translate;
float scale;
int normalizationExamples = config.normalizationExamples > Ntrain ? Ntrain : config.normalizationExamples;
if( !config.loadOnDemand ) {
if( config.normalization == "stddev" ) {
float mean, stdDev;
NormalizationHelper::getMeanAndStdDev( trainData, normalizationExamples * inputCubeSize, &mean, &stdDev );
cout << " image stats mean " << mean << " stdDev " << stdDev << endl;
translate = - mean;
scale = 1.0f / stdDev / config.normalizationNumStds;
} else if( config.normalization == "maxmin" ) {
float mean, stdDev;
NormalizationHelper::getMinMax( trainData, normalizationExamples * inputCubeSize, &mean, &stdDev );
translate = - mean;
scale = 1.0f / stdDev;
} else {
cout << "Error: Unknown normalization: " << config.normalization << endl;
return;
}
} else {
if( config.normalization == "stddev" ) {
float mean, stdDev;
NormalizeGetStdDev normalizeGetStdDev( trainData, trainLabels );
BatchProcessv2::run( &trainLoader, 0, config.batchSize, normalizationExamples, inputCubeSize, &normalizeGetStdDev );
normalizeGetStdDev.calcMeanStdDev( &mean, &stdDev );
cout << " image stats mean " << mean << " stdDev " << stdDev << endl;
translate = - mean;
scale = 1.0f / stdDev / config.normalizationNumStds;
} else if( config.normalization == "maxmin" ) {
NormalizeGetMinMax normalizeGetMinMax( trainData, trainLabels );
BatchProcessv2::run( &trainLoader, 0, config.batchSize, normalizationExamples, inputCubeSize, &normalizeGetMinMax );
normalizeGetMinMax.calcMinMaxTransform( &translate, &scale );
} else {
cout << "Error: Unknown normalization: " << config.normalization << endl;
return;
}
}
cout << " image norm translate " << translate << " scale " << scale << endl;
timer.timeCheck("after getting stats");
// const int numToTrain = Ntrain;
// const int batchSize = config.batchSize;
EasyCL *cl = 0;
if( config.gpuIndex >= 0 ) {
cl = EasyCL::createForIndexedGpu( config.gpuIndex );
} else {
cl = EasyCL::createForFirstGpuOtherwiseCpu();
}
NeuralNet *net;
net = new NeuralNet(cl);
WeightsInitializer *weightsInitializer = 0;
if( toLower( config.weightsInitializer ) == "original" ) {
weightsInitializer = new OriginalInitializer();
} else if( toLower( config.weightsInitializer ) == "uniform" ) {
weightsInitializer = new UniformInitializer( config.initialWeights );
} else {
cout << "Unknown weights initializer " << config.weightsInitializer << endl;
return;
}
// net->inputMaker<unsigned char>()->numPlanes(numPlanes)->imageSize(imageSize)->insert();
net->addLayer( InputLayerMaker::instance()->numPlanes(numPlanes)->imageSize(imageSize) );
net->addLayer( NormalizationLayerMaker::instance()->translate(translate)->scale(scale) );
if( !NetdefToNet::createNetFromNetdef( net, config.netDef, weightsInitializer ) ) {
return;
}
// apply the trainer
Trainer *trainer = 0;
if( toLower( config.trainer ) == "sgd" ) {
SGD *sgd = new SGD( cl );
sgd->setLearningRate( config.learningRate );
sgd->setMomentum( config.momentum );
sgd->setWeightDecay( config.weightDecay );
trainer = sgd;
} else if( toLower( config.trainer ) == "anneal" ) {
Annealer *annealer = new Annealer( cl );
annealer->setLearningRate( config.learningRate );
annealer->setAnneal( config.anneal );
trainer = annealer;
} else if( toLower( config.trainer ) == "nesterov" ) {
Nesterov *nesterov = new Nesterov( cl );
nesterov->setLearningRate( config.learningRate );
nesterov->setMomentum( config.momentum );
trainer = nesterov;
} else if( toLower( config.trainer ) == "adagrad" ) {
Adagrad *adagrad = new Adagrad( cl );
adagrad->setLearningRate( config.learningRate );
trainer = adagrad;
} else if( toLower( config.trainer ) == "rmsprop" ) {
Rmsprop *rmsprop = new Rmsprop( cl );
rmsprop->setLearningRate( config.learningRate );
trainer = rmsprop;
} else if( toLower( config.trainer ) == "adadelta" ) {
Adadelta *adadelta = new Adadelta( cl, config.rho );
trainer = adadelta;
} else {
cout << "trainer " << config.trainer << " unknown." << endl;
return;
}
cout << "Using trainer " << trainer->asString() << endl;
// trainer->bindTo( net );
// net->setTrainer( trainer );
net->setBatchSize( config.batchSize );
net->print();
bool afterRestart = false;
int restartEpoch = 0;
int restartBatch = 0;
float restartAnnealedLearningRate = 0;
int restartNumRight = 0;
float restartLoss = 0;
if( config.loadWeights && config.weightsFile != "" ) {
cout << "loadingweights" << endl;
afterRestart = WeightsPersister::loadWeights( config.weightsFile, config.getTrainingString(), net, &restartEpoch, &restartBatch, &restartAnnealedLearningRate, &restartNumRight, &restartLoss );
if( !afterRestart && FileHelper::exists( config.weightsFile ) ) {
// try old trainingstring
afterRestart = WeightsPersister::loadWeights( config.weightsFile, config.getOldTrainingString(), net, &restartEpoch, &restartBatch, &restartAnnealedLearningRate, &restartNumRight, &restartLoss );
}
if( !afterRestart && FileHelper::exists( config.weightsFile ) ) {
cout << "Weights file " << config.weightsFile << " exists, but doesnt match training options provided." << endl;
cout << "Continue loading anyway (might crash, or weights might be completely inappropriate)? (y/n)" << endl;
string response;
cin >> response;
if( response != "y" ) {
cout << "Please either check the training options, or choose a weights file that doesnt exist yet" << endl;
return;
}
}
void go(Config config) {
Timer timer;
int Ntrain;
int Ntest;
int numPlanes;
int imageSize;
float *trainData = 0;
float *testData = 0;
int *trainLabels = 0;
int *testLabels = 0;
int trainAllocateN = 0;
int testAllocateN = 0;
// int totalLinearSize;
GenericLoader::getDimensions((config.dataDir + "/" + config.trainFile).c_str(), &Ntrain, &numPlanes, &imageSize );
Ntrain = config.numTrain == -1 ? Ntrain : config.numTrain;
// long allocateSize = (long)Ntrain * numPlanes * imageSize * imageSize;
cout << "Ntrain " << Ntrain << " numPlanes " << numPlanes << " imageSize " << imageSize << endl;
trainAllocateN = Ntrain;
trainData = new float[ (long)trainAllocateN * numPlanes * imageSize * imageSize ];
trainLabels = new int[trainAllocateN];
if( Ntrain > 0 ) {
GenericLoader::load((config.dataDir + "/" + config.trainFile).c_str(), trainData, trainLabels, 0, Ntrain );
}
GenericLoader::getDimensions((config.dataDir + "/" + config.validateFile).c_str(), &Ntest, &numPlanes, &imageSize );
Ntest = config.numTest == -1 ? Ntest : config.numTest;
testAllocateN = Ntest;
testData = new float[ (long)testAllocateN * numPlanes * imageSize * imageSize ];
testLabels = new int[testAllocateN];
if( Ntest > 0 ) {
GenericLoader::load((config.dataDir + "/" + config.validateFile).c_str(), testData, testLabels, 0, Ntest );
}
timer.timeCheck("after load images");
const int inputCubeSize = numPlanes * imageSize * imageSize;
float translate;
float scale;
int normalizationExamples = config.normalizationExamples > Ntrain ? Ntrain : config.normalizationExamples;
if( config.normalization == "stddev" ) {
float mean, stdDev;
NormalizationHelper::getMeanAndStdDev( trainData, normalizationExamples * inputCubeSize, &mean, &stdDev );
cout << " image stats mean " << mean << " stdDev " << stdDev << endl;
translate = - mean;
scale = 1.0f / stdDev / config.normalizationNumStds;
} else if( config.normalization == "maxmin" ) {
float mean, stdDev;
NormalizationHelper::getMinMax( trainData, normalizationExamples * inputCubeSize, &mean, &stdDev );
translate = - mean;
scale = 1.0f / stdDev;
} else {
cout << "Error: Unknown normalization: " << config.normalization << endl;
return;
}
cout << " image norm translate " << translate << " scale " << scale << endl;
timer.timeCheck("after getting stats");
// const int numToTrain = Ntrain;
// const int batchSize = config.batchSize;
EasyCL *cl = new EasyCL();
NeuralNet *net = new NeuralNet( cl );
// net->inputMaker<unsigned char>()->numPlanes(numPlanes)->imageSize(imageSize)->insert();
net->addLayer( InputLayerMaker::instance()->numPlanes(numPlanes)->imageSize(imageSize) );
net->addLayer( NormalizationLayerMaker::instance()->translate(translate)->scale(scale) );
if( !NetdefToNet::createNetFromNetdef( net, config.netDef ) ) {
return;
}
net->print();
for( int i = 1; i < net->getNumLayers() - 1; i++ ) {
Layer *layer = net->getLayer(i);
FullyConnectedLayer *fc = dynamic_cast< FullyConnectedLayer * >(layer);
ConvolutionalLayer *conv = dynamic_cast< ConvolutionalLayer * >(layer);
if( fc != 0 ) {
conv = fc->convolutionalLayer;
}
if( conv == 0 ) {
continue;
}
initrand.seed(0);
int weightsSize = conv->getWeightsSize();
//int weightsSize = layer->getPersistSize();
if( weightsSize > 0 ) {
cout << "weightsSize " << weightsSize << endl;
float *weights = new float[weightsSize];
for( int j = 0; j < weightsSize; j++ ) {
int thisrand = (int)initrand();
float thisweight = ( thisrand % 100000 ) / 1000000.0f;
weights[j] = thisweight;
}
conv->initWeights( weights );
}
if( conv->dim.biased ) {
initrand.seed(0);
int biasedSize = conv->getBiasSize();
float *biasWeights = new float[biasedSize];
for( int j = 0; j < biasedSize; j++ ) {
int thisrand = (int)initrand();
float thisweight = ( thisrand % 100000 ) / 1000000.0f;
biasWeights[j] = thisweight;
//biasWeights[j] = 0;
}
conv->initBias( biasWeights );
}
}
cout << "weight samples before learning:" << endl;
sampleWeights(net);
bool afterRestart = false;
int restartEpoch = 0;
// int restartBatch = 0;
// float restartAnnealedLearningRate = 0;
// int restartNumRight = 0;
// float restartLoss = 0;
timer.timeCheck("before learning start");
if( config.dumpTimings ) {
StatefulTimer::dump( true );
}
StatefulTimer::timeCheck("START");
SGD *sgd = SGD::instance( cl, config.learningRate, 0.0f );
Trainable *trainable = net;
NetLearner netLearner(
sgd, trainable,
Ntrain, trainData, trainLabels,
Ntest, testData, testLabels,
config.batchSize );
netLearner.setSchedule( config.numEpochs, afterRestart ? restartEpoch : 1 );
// netLearner.setBatchSize( config.batchSize );
netLearner.setDumpTimings( config.dumpTimings );
// netLearner.learn( config.learningRate, 1.0f );
cout << "forward output" << endl;
for( int layerId = 0; layerId < net->getNumLayers(); layerId++ ) {
Layer *layer = net->getLayer(layerId);
FullyConnectedLayer *fc = dynamic_cast< FullyConnectedLayer * >( layer );
ConvolutionalLayer *conv = dynamic_cast< ConvolutionalLayer * >( layer );
PoolingLayer *pool = dynamic_cast< PoolingLayer * >( layer );
SoftMaxLayer *softMax = dynamic_cast< SoftMaxLayer * >( layer );
if( fc != 0 ) {
conv = fc->convolutionalLayer;
}
int planes = 0;
int imageSize = 0;
if( conv != 0 ) {
cout << "convolutional (or conv based, ie fc)" << endl;
planes = conv->dim.numFilters;
imageSize = conv->dim.outputSize;
// continue;
} else if( pool != 0 ) {
cout << "pooling" << endl;
planes = pool->numPlanes;
imageSize = pool->outputSize;
} else if( softMax != 0 ) {
cout << "softmax" << endl;
planes = softMax->numPlanes;
imageSize = softMax->imageSize;
} else {
continue;
}
cout << "layer " << layerId << endl;
// conv->getOutput();
float const*output = layer->getOutput();
// for( int i = 0; i < 3; i++ ) {
// cout << conv->getOutput()[i] << endl;
// }
initrand.seed(0);
// LayerDimensions &dim = conv->dim;
for( int i = 0; i < 10; i++ ) {
int thisrand = abs( (int)initrand() );
int seq = thisrand % ( planes * imageSize * imageSize );
int outPlane = seq / ( imageSize * imageSize );
int rowcol = seq % ( imageSize * imageSize );
int row = rowcol / imageSize;
int col = rowcol % imageSize;
cout << "out[" << outPlane << "," << row << "," << col << "]=" << output[ seq ] << endl;
}
}
cout << "weight samples after learning:" << endl;
sampleWeights(net);
cout << "backprop output" << endl;
for( int layerId = net->getNumLayers() - 1; layerId >= 0; layerId-- ) {
Layer *layer = net->getLayer(layerId);
FullyConnectedLayer *fc = dynamic_cast< FullyConnectedLayer * >( layer );
ConvolutionalLayer *conv = dynamic_cast< ConvolutionalLayer * >( layer );
if( fc != 0 ) {
conv = fc->convolutionalLayer;
}
if( conv == 0 ) {
continue;
}
cout << "layer " << layerId << endl;
float const*weights = conv->getWeights();
float const*biases = conv->getBias();
int weightsSize = conv->getWeightsSize() / conv->dim.numFilters;
for( int i = 0; i < weightsSize; i++ ) {
cout << " weight " << i << " " << weights[i] << endl;
}
for( int i = 0; i < 3; i++ ) {
cout << " bias " << i << " " << biases[i] << endl;
}
}
cout << "done" << endl;
delete sgd;
delete net;
delete cl;
if( trainData != 0 ) {
delete[] trainData;
}
if( testData != 0 ) {
delete[] testData;
}
if( testLabels != 0 ) {
delete[] testLabels;
}
if( trainLabels != 0 ) {
delete[] trainLabels;
}
}
