seriesClassification.cpp

/**
 *  @file seriesClassification.cpp
 *  RapidLib
 *
 *  @author Michael Zbyszynski
 *  @date 08 Jun 2017
 *  @copyright Copyright © 2017 Goldsmiths. All rights reserved.
 */

#include <vector>
#include <cassert>
#include "seriesClassification.h"
#ifdef EMSCRIPTEN
#include "emscripten/seriesClassificationEmbindings.h"
#endif

#define SEARCH_RADIUS 1

template<typename T>
seriesClassificationTemplate<T>::seriesClassificationTemplate() : hopSize(1), counter(0) {};

template<typename T>
seriesClassificationTemplate<T>::~seriesClassificationTemplate() {};

template<typename T>
bool seriesClassificationTemplate<T>::train(const std::vector<trainingSeriesTemplate<T> > &seriesSet) {
    assert(seriesSet.size() > 0);
    reset();
    vectorLength = int(seriesSet[0].input[0].size()); //TODO: check that all vectors are the same size
    bool trained = true;
    allTrainingSeries = seriesSet;
    minLength = maxLength = int(allTrainingSeries[0].input.size());
    for (int i = 0; i < allTrainingSeries.size(); ++i) {
        //for (auto trainingSeries : allTrainingSeries)
        //Global
        int newLength = int(allTrainingSeries[i].input.size());
        if (newLength < minLength) {
            minLength = newLength;
        }
        if (newLength > maxLength) {
            maxLength = newLength;
        }
        //Per Label
        typename std::map<std::string, minMax<int> >::iterator it = lengthsPerLabel.find(allTrainingSeries[i].label);
        if (it != lengthsPerLabel.end()) {
            int newLength = int(allTrainingSeries[i].input.size());
            if (newLength < it->second.min) {
                it->second.min = newLength;
            }
            if (newLength > it->second.max) {
                it->second.max = newLength;
            }
        } else {
            minMax<int> tempLengths;
            tempLengths.min = tempLengths.max = int(allTrainingSeries[i].input.size());
            lengthsPerLabel[allTrainingSeries[i].label] = tempLengths;
        }
    }
    //TODO: make this size smarter?
    std::vector<T> zeroVector;
    for (int i = 0; i < vectorLength; ++i) {
        zeroVector.push_back(0.0);
    }
    for (int i = 0; i < minLength; ++i ) {
        seriesBuffer.push_back(zeroVector); //set size of continuous buffer
    }
    return trained;
};

template<typename T>
void seriesClassificationTemplate<T>::reset() {
    allCosts.clear();
    allTrainingSeries.clear();
    lengthsPerLabel.clear();
    minLength = -1;
    maxLength = -1;
}

template<typename T>
std::string seriesClassificationTemplate<T>::run(const std::vector<std::vector<T>> &inputSeries) {
    //TODO: Check to see if trained
    int closestSeries = 0;
    allCosts.clear();
    T lowestCost = fastDTW<T>::getCost(inputSeries, allTrainingSeries[0].input, SEARCH_RADIUS);
    allCosts.push_back(lowestCost);
    
    for (int i = 1; i < allTrainingSeries.size(); ++i) {
        T currentCost = fastDTW<T>::getCost(inputSeries, allTrainingSeries[i].input, SEARCH_RADIUS);
        allCosts.push_back(currentCost);
        if (currentCost < lowestCost) {
            lowestCost = currentCost;
            closestSeries = i;
        }
    }
    return allTrainingSeries[closestSeries].label;
};

template<typename T>
T seriesClassificationTemplate<T>::run(const std::vector<std::vector<T>> &inputSeries, std::string label) {
    //TODO: Check to see if trained
    int closestSeries = 0;
    allCosts.clear();
    T lowestCost = std::numeric_limits<T>::max();
    for (int i = 0; i < allTrainingSeries.size(); ++i) {
        if (allTrainingSeries[i].label == label) {
            T currentCost = fastDTW<T>::getCost(inputSeries, allTrainingSeries[i].input, SEARCH_RADIUS);
            allCosts.push_back(currentCost);
            if (currentCost < lowestCost) {
                lowestCost = currentCost;
                closestSeries = i;
            }
        }
    }
    return lowestCost;
};

template<typename T>
std::string seriesClassificationTemplate<T>::runContinuous(const std::vector<T> &inputVector) {
    seriesBuffer.erase(seriesBuffer.begin());
    seriesBuffer.push_back(inputVector);
    std::string returnString = "none";
    if ((counter % hopSize) == 0 ) {
        returnString = run(seriesBuffer);
        counter = 0;
    }
    ++counter;
    return returnString;
}

template<typename T>
std::vector<T> seriesClassificationTemplate<T>::getCosts() const{
    return allCosts;
}

template<typename T>
int seriesClassificationTemplate<T>::getMinLength() const{
    return minLength;
}

template<typename T>
int seriesClassificationTemplate<T>::getMinLength(std::string label) const {
    int labelMinLength = -1;
    typename std::map<std::string, minMax<int> >::const_iterator it = lengthsPerLabel.find(label);
    if (it != lengthsPerLabel.end()) {
        labelMinLength = it->second.min;
    }
    return labelMinLength;
}

template<typename T>
int seriesClassificationTemplate<T>::getMaxLength() const {
    return maxLength;
}

template<typename T>
int seriesClassificationTemplate<T>::getMaxLength(std::string label) const {
    int labelMaxLength = -1;
    typename std::map<std::string, minMax<int> >::const_iterator it = lengthsPerLabel.find(label);
    if (it != lengthsPerLabel.end()) {
        labelMaxLength = it->second.max;
    }
    return labelMaxLength;
}

template<typename T>
seriesClassificationTemplate<T>::minMax<T> seriesClassificationTemplate<T>::calculateCosts(std::string label) const {
    minMax<T> calculatedMinMax;
    bool foundSeries = false;
    std::vector<T> labelCosts;
    for (int i = 0; i < (allTrainingSeries.size() - 1); ++i) { //these loops are a little different than the two-label case
        if (allTrainingSeries[i].label == label) {
            foundSeries = true;
            for (int j = (i + 1); j < allTrainingSeries.size(); ++j) {
                if (allTrainingSeries[j].label == label) {
                    labelCosts.push_back(fastDTW<T>::getCost(allTrainingSeries[i].input, allTrainingSeries[j].input, SEARCH_RADIUS));
                }
            }
        }
    }
    if (foundSeries) {
        auto minmax_result = std::minmax_element(std::begin(labelCosts), std::end(labelCosts));
        calculatedMinMax.min = *minmax_result.first;
        calculatedMinMax.max = *minmax_result.second;
    } else {
        calculatedMinMax.min = calculatedMinMax.max = 0;
    }
    return calculatedMinMax;
}

template<typename T>
seriesClassificationTemplate<T>::minMax<T> seriesClassificationTemplate<T>::calculateCosts(std::string label1, std::string label2) const {
    minMax<T> calculatedMinMax;
    bool foundSeries = false;
    std::vector<T> labelCosts;
    for (int i = 0; i < (allTrainingSeries.size()); ++i) {
        if (allTrainingSeries[i].label == label1) {
            for (int j = 0; j < allTrainingSeries.size(); ++j) {
                if (allTrainingSeries[j].label == label2) {
                    foundSeries = true;
                    labelCosts.push_back(fastDTW<T>::getCost(allTrainingSeries[i].input, allTrainingSeries[j].input, SEARCH_RADIUS));
                }
            }
        }
    }
    if (foundSeries) {
        auto minmax_result = std::minmax_element(std::begin(labelCosts), std::end(labelCosts));
        calculatedMinMax.min = *minmax_result.first;
        calculatedMinMax.max = *minmax_result.second;
    } else {
        calculatedMinMax.min = calculatedMinMax.max = 0;
    }
    return calculatedMinMax;
}

//explicit instantiation
template class seriesClassificationTemplate<double>;
template class seriesClassificationTemplate<float>;


//
//std::vector<T> seriesClassification::getCosts(const std::vector<trainingExample> &trainingSet) {
//    run(trainingSet);
//    return allCosts;
//}