Function QuantizeDepthwiseConv

tools/quantization/quantizeWeight.cpp:194–227

}

int QuantizeDepthwiseConv(const float* weight, const int size, const float* bias, int8_t* quantizedWeight,
                          int32_t* quantizedBias, float* scale, const float inputScale, const float outputScale,
                          const int inputChannel, const int outputChannel, std::string method, float weightClampValue, bool mergeChannel) {
    DCHECK(inputChannel == outputChannel) << "Input Data Size Error!";

    std::vector<float> quantizedWeightScale(inputChannel);
    if (method == "MAX_ABS") {
        SymmetricQuantizeWeight(weight, size, quantizedWeight, quantizedWeightScale.data(), inputChannel, weightClampValue);
    }
    else if (method == "ADMM") {
        QuantizeWeightADMM(weight, size, quantizedWeight, quantizedWeightScale.data(), inputChannel, weightClampValue);
    }

    for (int c = 0; c < inputChannel; ++c) {
        const int index = c;
        if (fabs(outputScale) <= 1e-6) {
            scale[index] = 0.0f;
        } else {
            scale[index] = inputScale * quantizedWeightScale[c] / outputScale;
        }
    }

    if (bias) {
        for (int i = 0; i < outputChannel; ++i) {
            if (fabs(inputScale) <= 1e-6 || fabs(quantizedWeightScale[i]) <= 1e-6) {
                quantizedBias[i] = 0;
            } else {
                quantizedBias[i] = static_cast<int32_t>(bias[i] / (inputScale * quantizedWeightScale[i]));
            }
        }
    }

    return 0;
}