How to use ExpectedData correctly in training?

[AuroraRAS]

I'm trying to understand ExpectedData class in training

LabeledData class and ExpectedData class, they using same parameters and structure in training.

output result of LabeledData look's good

LabeledData Test:
-1 xor -1 = 0.729217
-1 xor 1 = 0.302157
1 xor -1 = 0.26969
1 xor 1 = 0.728911

but result of ExpectedData is strange

ExpectedData Test:
-1 xor -1 = 0.505612
-1 xor 1 = 0.508897
1 xor -1 = 0.499579
1 xor 1 = 0.293857

This my XOR case:

#include <iostream>

#include "deepcl/DeepCL.h"
#include "deepcl/batch/BatchData.h"

int main(int argc, char *argv[])
{    
    int numExamples = 4;

    float data[] = { -1, -1,
                     -1, 1,
                     1, -1,
                     1, 1 };

    EasyCL *cl;
    NeuralNet *net;
    SGD *trainer;

    /*
     * LabeledData Test Case
     */
    cl= new EasyCL();
    net = new NeuralNet(cl);
    net->addLayer( InputLayerMaker::instance()->numPlanes(2)->imageSize(1) );
    net->addLayer( ConvolutionalMaker::instance()->numFilters(4)->filterSize(1)->padZeros()->biased(1) );
    net->addLayer( ActivationMaker::instance()->relu() );
    net->addLayer( FullyConnectedMaker::instance()->numPlanes(2)->imageSize(1) );
    net->addLayer( ActivationMaker::instance()->sigmoid() );
    net->addLayer( SoftMaxMaker::instance() );
    trainer = SGD::instance( cl, 0.1f, 0.0f );

    // LabeledData
    for (int var = 0; var < 3000; ++var) {
        TrainingContext tc(var, 0);

        int a =rand()%2;
        int b =rand()%2;

        float input[2];
        input[0] = a? 1: -1;
        input[1] = b? 1: -1;

        int label = a ^ b;
        LabeledData labeledData(&label);

        net->setBatchSize(1);
        trainer->trainNet(net, &tc, input, &labeledData);

        if(var%100 == 0)
            std::cout << var << "...";
    }
    std::cout << std::endl;

    std::cout << "LabeledData Test:" << std::endl;
    for (int var1 = 0; var1 < numExamples; ++var1) {
        net->forward(&(data[var1*2]));
        const float * output = net->getOutput();
        std::cout << data[var1*2] << " xor " << data[(var1*2)+1] << " = " << *output << std::endl;
    }

    delete trainer;
    delete net;
    delete cl;

    /*
     * ExpectedData Test Case
     */
    cl= new EasyCL();
    net = new NeuralNet(cl);
    net->addLayer( InputLayerMaker::instance()->numPlanes(2)->imageSize(1) );
    net->addLayer( ConvolutionalMaker::instance()->numFilters(4)->filterSize(1)->padZeros()->biased(1) );
    net->addLayer( ActivationMaker::instance()->relu() );
    net->addLayer( FullyConnectedMaker::instance()->numPlanes(2)->imageSize(1) );
    net->addLayer( ActivationMaker::instance()->sigmoid() );
    net->addLayer( SoftMaxMaker::instance() );
    trainer = SGD::instance( cl, 0.1f, 0.0f );

    // ExpectedData
    for (int var = 0; var < 3000; ++var) {
        TrainingContext tc(var, 0);

        int a =rand()%2;
        int b =rand()%2;

        float input[2];
        input[0] = a? 1: -1;
        input[1] = b? 1: -1;

        float expectedOutput = a ^ b;
        ExpectedData expectedData(1, &expectedOutput);

        net->setBatchSize(1);
        trainer->trainNet(net, &tc, input, &expectedData);

        if(var%100 == 0)
            std::cout << var << "...";
    }
    std::cout << std::endl;

    std::cout << "ExpectedData Test:" << std::endl;
    for (int var1 = 0; var1 < numExamples; ++var1) {
        net->forward(&(data[var1*2]));
        const float * output = net->getOutput();
        std::cout << data[var1*2] << " xor " << data[(var1*2)+1] << " = " << *output << std::endl;
    }

    delete trainer;
    delete net;
    delete cl;

    return 0;
}

[AuroraRAS]

I learned how to use it

#include <iostream>

#include "deepcl/DeepCL.h"
#include "deepcl/batch/BatchData.h"

int main(int argc, char *argv[])
{    
    int numExamples = 4;

    float data[] = { -1, -1,
                     -1, 1,
                     1, -1,
                     1, 1 };

    EasyCL *cl;
    NeuralNet *net;
    Trainer *trainer;

    /*
     * ExpectedData Test Case
     */
    cl= new EasyCL();
    net = new NeuralNet(cl);
    net->addLayer( InputLayerMaker::instance()->numPlanes(2)->imageSize(1) );
    net->addLayer( ConvolutionalMaker::instance()->numFilters(4)->filterSize(1)->padZeros()->biased(1) );
    net->addLayer( ActivationMaker::instance()->relu() );
    net->addLayer( FullyConnectedMaker::instance()->numPlanes(2)->imageSize(1) );
    net->addLayer( ActivationMaker::instance()->sigmoid() );
//    net->addLayer( SoftMaxMaker::instance() );
    // SquareLoss looks better in ExpectedData
    net->addLayer( SquareLossMaker::instance() );
    // bigger learning rate
    trainer = SGD::instance( cl, 0.2f, 0.0f );

    // ExpectedData
    // more epoch is seems necessary
    for (int var = 0; var < 10000; ++var) {
        TrainingContext tc(var, 0);

        int a =rand()%2;
        int b =rand()%2;

        float input[2];
        input[0] = a? 1: -1;
        input[1] = b? 1: -1;

        float expectedOutput = a ^ b;
        //it's a part of Trainer
//        ExpectedData expectedData(1, &expectedOutput);

        net->setBatchSize(1);
        trainer->trainNet(net, &tc, input, &expectedOutput);

        if(var%100 == 0)
            std::cout << var << "..." << std::endl;
    }
    std::cout << std::endl;

    std::cout << "ExpectedData Test:" << std::endl;
    for (int var1 = 0; var1 < numExamples; ++var1) {
        net->forward(&(data[var1*2]));
        const float * output = net->getOutput();
        std::cout << data[var1*2] << " xor " << data[(var1*2)+1] << " = " << *output << std::endl;
    }

    delete trainer;
    delete net;
    delete cl;

[hughperkins]

Cool. PR for docs/examples welcome :)