Function UpdateQuantizedWeights

tools/quantization/quantizeWeight.cpp:42–53

}

void UpdateQuantizedWeights(const float* weight, const int weightNum, const int kernelNum, float* alpha,
        const float weightClampValue, int8_t* quantizedWeight) {
    const int kernelDim = weightNum / kernelNum;
    const float eps = 1e-9f;
    float weightQuan;
    CHECK((int)weightClampValue >= 7) << "quantization bits less than 4 not supported yet.";

    for (int i = 0; i < weightNum; i++) {
        weightQuan = weight[i] / (alpha[i / kernelDim]+ eps);
        quantizedWeight[i] = std::min(weightClampValue, std::max(-weightClampValue, std::roundf(weightQuan)));
    }
}

void UpdateAlpha(const float* weight, const int weightNum, const int kernelNum, float* alpha, int8_t* quantizedWeight) {
    const int kernelDim = weightNum / kernelNum;