MLP With OpenCV 3

The problem

This ANN works with 3-input XOR case. This is a classic non-linear problem that can't be solved with linear cassifiers. We have 3 input (A, B, C) and 2 output (0 or 1). Opencv uses symmetrical sigmoid (between -1 and 1) so we have "+1" for our true response and "-1" for false response.

XOR3 truth table:

Topology

• Input layer size: 3
• Hidden layer size: 10
• outnput layer size: 2

Show me the code!!

//
//  main.cpp
//  MLPCV
//
//  Created by Thomaz Maia on 12/08/16.
//  Copyright © 2016 Thomaz Maia . All rights reserved.
//

#include <iostream>
#include <opencv2/opencv.hpp>

int main(int argc, const char * argv[]) {
    // Database config
    const int numInstances = 8;
    const int numAttributes = 3;
    const int numHiddenNeurons = 10;
    const int numClasses = 2;

    // input XOR
    float data[] = {0,0,0,  0,0,1,  0,1,0,  0,1,1,  1,0,0,  1,0,1,  1,1,0,  1,1,1};
    cv::Mat input = cv::Mat(numInstances, numAttributes, CV_32FC1, data);
    std::cout << "Input layer:" << std::endl << input << std::endl;
    
    // output XOR
    float labels[] = {1,-1,  -1,1,  -1,1,  1,-1,  -1,1,  1,-1,  1,-1,  -1,1};
    cv::Mat output = cv::Mat(numInstances, numClasses, CV_32FC1, labels);
    std::cout << "Output layer:" << std::endl << output << std::endl;
    
    // Configurate topology
    float layer[] = {numAttributes, numHiddenNeurons, numClasses};
    cv::Mat layerSizes = cv::Mat(1, 3, CV_32FC1, layer);
    std::cout << "MLP topology: " << layerSizes << std::endl;
    
    // Create ANN
    cv::Ptr mlp = cv::ml::ANN_MLP::create();
    mlp->setLayerSizes(layerSizes);
    mlp->setActivationFunction(cv::ml::ANN_MLP::SIGMOID_SYM);
    mlp->setTrainMethod(cv::ml::ANN_MLP::RPROP);
    mlp->setTermCriteria(cv::TermCriteria(cv::TermCriteria::MAX_ITER + cv::TermCriteria::EPS, 1000, 0.0001));
    
    // Training...
    mlp->train(input, cv::ml::ROW_SAMPLE, output);
    
    // Prediction
    cv::Mat p_output;
    mlp->predict(input, p_output);
    std::cout << "Predicted output:" << std::endl << p_output << std::endl;
    
    return 0;
}