MNN量化流程

int main(int argc, const char* argv[]) {
    if (argc < 4) {
        DLOG(INFO) << "Usage: ./quantized.out src.mnn dst.mnn preTreatConfig.json\n";
        return 0;
    }
    const char* modelFile      = argv[1];
    const char* preTreatConfig = argv[3];
    const char* dstFile        = argv[2];
    DLOG(INFO) << ">>> modelFile: " << modelFile;
    DLOG(INFO) << ">>> preTreatConfig: " << preTreatConfig;
    DLOG(INFO) << ">>> dstFile: " << dstFile;
    std::unique_ptr<MNN::NetT> netT;
    // 读取MNN模型
    {
        std::ifstream input(modelFile);
        std::ostringstream outputOs;
        outputOs << input.rdbuf();
        netT = MNN::UnPackNet(outputOs.str().c_str());
    }
    // temp build net for inference
    flatbuffers::FlatBufferBuilder builder(1024);
    auto offset = MNN::Net::Pack(builder, netT.get());
    builder.Finish(offset);
    int size      = builder.GetSize();
    auto ocontent = builder.GetBufferPointer();
    // model buffer for creating mnn Interpreter
    // 构造两个buffer，一个负责进行推理统计，一个负责生成量化网络
    std::unique_ptr<uint8_t> modelForInference(new uint8_t[size]);
    memcpy(modelForInference.get(), ocontent, size);
    std::unique_ptr<uint8_t> modelOriginal(new uint8_t[size]);
    memcpy(modelOriginal.get(), ocontent, size);
    netT.reset();
    netT = MNN::UnPackNet(modelOriginal.get());
    // quantize model's weight
    DLOG(INFO) << "Calibrate the feature and quantize model...";
    // 构造校准类从而进行量化
    std::shared_ptr<Calibration> calibration(
        new Calibration(netT.get(), modelForInference.get(), size, preTreatConfig));
    calibration->runQuantizeModel();
    DLOG(INFO) << "Quantize model done!";
    flatbuffers::FlatBufferBuilder builderOutput(1024);
    builderOutput.ForceDefaults(true);
    auto len = MNN::Net::Pack(builderOutput, netT.get());
    builderOutput.Finish(len);
    {
        std::ofstream output(dstFile);
        output.write((const char*)builderOutput.GetBufferPointer(), builderOutput.GetSize());
    }
}

quantized.cpp 文件负责实现量化，流程如下所示：

• 通过flatbuffer读取模型内容，并生成两个模型buffer，其中一个负责进行推理统计激活值信息，另一个负责生成量化网络。
• 调用Calibration类进行量化，生成量化网络。
• 保存量化后的模型。