Function QuantizeWeightADMM

tools/quantization/quantizeWeight.cpp:73–102

weight format is [co, ci, kh, kw]

// weight format is [co, ci, kh, kw]
int QuantizeWeightADMM(const float* weight, const int weightNum, int8_t* quantizedWeight, float* alpha,
                            const int kernelNum, const float weightClampValue) {
    // channels: co
    DCHECK((weightNum % kernelNum) == 0) << "weight size error!";
    const int kernelDim     = weightNum / kernelNum; // ci * kh * kw

    InitAlpha(weight, weightNum, kernelNum, alpha, weightClampValue);

    int iter = 0;
    float diffRate = 1;
    float preSum = 0;
    float curSum = 0;
    const int maxIter = 1000;

    for (int i = 0; i < weightNum; i++){
        preSum += std::fabs(weight[i]);
    }
    // update weights quan
    while(iter < maxIter) {
        UpdateQuantizedWeights(weight, weightNum, kernelNum, alpha, weightClampValue, quantizedWeight);
        UpdateAlpha(weight, weightNum, kernelNum, alpha, quantizedWeight);
        iter++;
    }

    for (int i = 0; i < weightNum; i++){
        curSum += std::fabs(quantizedWeight[i]*alpha[i/kernelDim]);
    }
    DLOG(INFO) << "iter: " << iter << " with diff "<< preSum-curSum;
    return 0;
}

// weight format is [co, ci, kh, kw]
int SymmetricQuantizeWeight(const float* weight, const int size, int8_t* quantizedWeight, float* scale,