softmax.hpp

Go to the documentation of this file.

#pragma once
 
#include "NumCpp/Core/Internal/StaticAsserts.hpp"
#include "NumCpp/Core/Internal/StlAlgorithms.hpp"
#include "NumCpp/Core/Types.hpp"
#include "NumCpp/Functions/exp.hpp"
#include "NumCpp/NdArray.hpp"
 
namespace nc::special
{
    //============================================================================
    // Method Description:
    template<typename dtype>
    NdArray<double> softmax(const NdArray<dtype>& inArray, Axis inAxis = Axis::NONE)
    {
        STATIC_ASSERT_ARITHMETIC(dtype);
 
        switch (inAxis)
        {
            case Axis::NONE:
            {
                auto returnArray = exp(inArray).template astype<double>();
                returnArray /= static_cast<double>(returnArray.sum().item());
                return returnArray;
            }
            case Axis::COL:
            {
                auto returnArray = exp(inArray).template astype<double>();
                auto expSums     = returnArray.sum(inAxis);
 
                for (uint32 row = 0; row < returnArray.shape().rows; ++row)
                {
                    const auto rowExpSum = static_cast<double>(expSums[row]);
                    stl_algorithms::for_each(returnArray.begin(row),
                                             returnArray.end(row),
                                             [rowExpSum](double& value) { value /= rowExpSum; });
                }
 
                return returnArray;
            }
            case Axis::ROW:
            {
                auto returnArray = exp(inArray.transpose()).template astype<double>();
                auto expSums     = returnArray.sum(Axis::COL);
 
                for (uint32 row = 0; row < returnArray.shape().rows; ++row)
                {
                    const auto rowExpSum = static_cast<double>(expSums[row]);
                    stl_algorithms::for_each(returnArray.begin(row),
                                             returnArray.end(row),
                                             [rowExpSum](double& value) { value /= rowExpSum; });
                }
 
                return returnArray.transpose();
            }
            default:
            {
                THROW_INVALID_ARGUMENT_ERROR("Unimplemented axis type.");
                return {}; // get rid of compiler warning
            }
        }
    }
} // namespace nc::special