CrossEntropyLoss.hpp

Go to the documentation of this file.

#ifndef CROSSENTROPYLOSS_HPP_
#define CROSSENTROPYLOSS_HPP_

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

namespace mic {
namespace neural_nets {
namespace loss {

template <typename dtype=float>
class CrossEntropyLoss : 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());

        // Calculate loss (negative log probability).
        dtype loss =0;
        dtype eps = 1e-15;
        for (size_t i=0; i <(size_t)predicted_y_->size(); i++) {
            // -t * log (y + eps!)
            loss -= (*target_y_)[i] * std::log2((*predicted_y_)[i] + eps);
        }
        // Return cross-entropy error (CE).
        // The average cross entropy error (ACE) is loss divided by the batch size.
        //assert(std::isfinite(loss));
        return 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++) {
            // y - t
            (*dy)[i] = (*predicted_y_)[i] - (*target_y_)[i];
        }
        return dy;
    }

};

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

#endif /* CROSSENTROPYLOSS_HPP_ */