LogLikelihoodLoss.hpp

Go to the documentation of this file.

#ifndef LOGLIKELIHOODLOSS_HPP_
#define LOGLIKELIHOODLOSS_HPP_






#include <cmath>
#include <loss/Loss.hpp>

namespace mic {
namespace neural_nets {
namespace loss {

template <typename dtype=float>
class LogLikelihoodLoss : public Loss<dtype> {
public:
    dtype calculateLoss (mic::types::MatrixPtr<dtype> target_y_, mic::types::MatrixPtr<dtype> predicted_y_) {
        // Sizes must match.
        assert(predicted_y_->size() == target_y_->size());

        size_t ind;
        // Calculate loss.
        dtype loss =0;
        // For each column (sample from batch).
        for (size_t i=0; i <(size_t)predicted_y_->cols(); i++) {
            // Get index of max coefficient in given column.
            target_y_.col(i).maxCoeff(ind);

            // Add loss.
            loss -= std::log((*predicted_y_)[ind]);
        }//: for
        // Return sum of log-likelihood cost.
        return loss;
        // Divide it by the batch size in order to calculate the mean loss.
    }

    mic::types::MatrixPtr<dtype> calculateGradient (mic::types::MatrixPtr<dtype> target_y_, mic::types::MatrixPtr<dtype> predicted_y_) {
        // Sizes must match.
        assert(predicted_y_->size() == target_y_->size());

        // Calculate gradient.
        mic::types::MatrixPtr<dtype> dy = MAKE_MATRIX_PTR(dtype, predicted_y_->rows(), predicted_y_->cols());
        for (size_t i=0; i <(size_t)predicted_y_->size(); i++) {
            (*dy)[i] = 0.0;
        }
        return dy;
    }

};

} //: loss
} //: neural_nets
} //: mic

#endif /* LOGLIKELIHOODLOSS_HPP_ */