Function _HQQQuant

tools/converter/source/common/WeightQuantAndCoding.cpp:43–73

}

static MNN::Quantization::HQQQuantizer::QuantizationResult _HQQQuant(const std::vector<float>& weights, int weightQuantBits, int weightQuantBlock, bool asymmetricQuantFlag) {
    MNN::Quantization::HQQQuantizer::QuantizationConfig hqqConfig;
    hqqConfig.bits = weightQuantBits;
    hqqConfig.group_size = weightQuantBlock;
    MNN::Quantization::HQQQuantizer hqq(hqqConfig);
    auto res = hqq.quantize(weights);
#if 0
    auto dequantized_weights = hqq.dequantize(res);
    // 计算量化误差
    float mse = 0.0f;
    float max_abs_error = 0.0f;

    for (size_t i = 0; i < weights.size(); ++i) {
        float error = weights[i] - dequantized_weights->readMap<float>()[i];
        float abs_error = std::abs(error);

        mse += error * error;
        max_abs_error = std::max(max_abs_error, abs_error);
    }

    mse /= weights.size();
    float rmse = std::sqrt(mse);

    std::cout << "量化误差分析:" << std::endl;
    std::cout << "  均方误差 (MSE): " << mse << std::endl;
    std::cout << "  均方根误差 (RMSE): " << rmse << std::endl;
    std::cout << "  最大绝对误差: " << max_abs_error << std::endl;
#endif
    return res;

}

void WeightQuantAndCoding(std::unique_ptr<MNN::OpT>& op, const modelConfig& config, const PostTreatContext* context) {
    const auto opType = op->type;