MNIST::CatHinge Class Reference

Categorical hinge loss. Use if weights are restricted to be >0. More...

#include <mnist_mlp.hpp>

Static Public Member Functions
static Real	calc_loss (const uint16_t label, const std::vector< Real > &output)
static std::vector< Real >	calc_error (const uint16_t label, const std::vector< Real > &output)

Detailed Description

Categorical hinge loss. Use if weights are restricted to be >0.

Definition at line 71 of file mnist_mlp.hpp.

Member Function Documentation

static std::vector<Real> MNIST::CatHinge::calc_error ( const uint16_t  label, const std::vector< Real > &  output  )

inlinestatic

Definition at line 87 of file mnist_mlp.hpp.

    {
        // Real pos = output[label];
        std::vector<Real> vec = output;
        vec[label] = -0.0;
        auto neg_elem = std::max_element(vec.begin(), vec.end());
        auto index = std::distance(vec.begin(), neg_elem);
        auto res = std::vector<Real>(output.size(), 0.0);

        // Require that label neuron and the next most active neuron have at
        // least a difference of 1
        if ((*neg_elem) - output[label] + 1 >= 0.0) {
            res[label] = -1.0;
            if (label != index) {
                res[index] = +1;
            }
        }
        return res;
    }

static Real MNIST::CatHinge::calc_loss ( const uint16_t  label, const std::vector< Real > &  output  )

inlinestatic

Definition at line 73 of file mnist_mlp.hpp.

    {
        Real res = 0.0;
        for (size_t neuron = 0; neuron < output.size(); neuron++) {
            if (label == neuron) {
                res += std::max(0.0, 1.0 - Real(output[neuron]));
            }
            else {
                res += std::max(0.0, 1.0 + Real(output[neuron]));
            }
        }
        return res;
    }

---

The documentation for this class was generated from the following file:

• source/SNABs/mnist/mnist_mlp.hpp