[人工智能]【TensorRT】mnist的onnx模型转engine模型并调用

参考：使用TensorRT7.0.0.11工具trtexec onnx模型转engine_kangkjz的博客-CSDN博客_onnx转engine

打开./samples/trtexec/trtexec.sln，编译之后可以看到

执行命令行：

trtexec.exe --onnx=C:\Project\TensorRT-8.0.3.4\data\mnist\mnist.onnx --saveEngine=C:\Project\TensorRT-8.0.3.4\data\mnist\mnist.engine

?需要花一点点时间：

在SampleOnnxMNIST这个实例中，对代码简单修改，把build函数中主动去读取mnist.engine文件，实例化ICudaEngine。

bool SampleOnnxMNIST::build()
{
#if 0


    auto builder = SampleUniquePtr<nvinfer1::IBuilder>(nvinfer1::createInferBuilder(sample::gLogger.getTRTLogger()));
    
	if (!builder)
    {
        return false;
    }

    const auto explicitBatch = 1U << static_cast<uint32_t>(NetworkDefinitionCreationFlag::kEXPLICIT_BATCH);
    auto network = SampleUniquePtr<nvinfer1::INetworkDefinition>(builder->createNetworkV2(explicitBatch));
    if (!network)
    {
        return false;
    }

    auto config = SampleUniquePtr<nvinfer1::IBuilderConfig>(builder->createBuilderConfig());
    if (!config)
    {
        return false;
    }

    auto parser
        = SampleUniquePtr<nvonnxparser::IParser>(nvonnxparser::createParser(*network, sample::gLogger.getTRTLogger()));
    if (!parser)
    {
        return false;
    }

    auto constructed = constructNetwork(builder, network, config, parser);
    if (!constructed)
    {
        return false;
    }

    // CUDA stream used for profiling by the builder.
    auto profileStream = samplesCommon::makeCudaStream();
    if (!profileStream)
    {
        return false;
    }
    config->setProfileStream(*profileStream);

    SampleUniquePtr<IHostMemory> plan{builder->buildSerializedNetwork(*network, *config)};
    if (!plan)
    {
        return false;
    }

    SampleUniquePtr<IRuntime> runtime{createInferRuntime(sample::gLogger.getTRTLogger())};
    if (!runtime)
    {
        return false;
    }

    mEngine = std::shared_ptr<nvinfer1::ICudaEngine>(
        runtime->deserializeCudaEngine(plan->data(), plan->size()), samplesCommon::InferDeleter());
    if (!mEngine)
    {
        return false;
    }

    ASSERT(network->getNbInputs() == 1);
    mInputDims = network->getInput(0)->getDimensions();
    ASSERT(mInputDims.nbDims == 4);

	std::cout << "[jhq] mInputDims: " << mInputDims.d[0] << " , " << mInputDims.d[1] << " , " << mInputDims.d[2] << " , " << mInputDims.d[3] << std::endl;

    ASSERT(network->getNbOutputs() == 1);
    mOutputDims = network->getOutput(0)->getDimensions();
    ASSERT(mOutputDims.nbDims == 2);
	std::cout << "[jhq] mOutputDims: " << mInputDims.d[0] << " , " << mInputDims.d[1] << std::endl;



    return true;

#else

	mInputDims.d[0] = 1;
	mInputDims.d[1] = 1;
	mInputDims.d[2] = 28;
	mInputDims.d[3] = 28;

	mOutputDims.d[0] = 1;
	mOutputDims.d[1] = 1;

	std::string engine_name = "./data/mnist.engine";
	std::ifstream file(engine_name, std::ios::binary);
    if (!file.good()) {
        std::cerr << "read " << engine_name << " error!" << std::endl;
        return -1;
    }

    char *trtModelStream = nullptr;
    size_t size = 0;
    file.seekg(0, file.end);
    size = file.tellg();
    file.seekg(0, file.beg);
    trtModelStream = new char[size];
    assert(trtModelStream);
    file.read(trtModelStream, size);
    file.close();

    SampleUniquePtr<IRuntime> runtime{createInferRuntime(sample::gLogger.getTRTLogger())};
    if (!runtime)
    {
        return false;
    }

    mEngine = std::shared_ptr<nvinfer1::ICudaEngine>(
        runtime->deserializeCudaEngine(trtModelStream, size), samplesCommon::InferDeleter());
    if (!mEngine)
    {
        return false;
    }

	return true;

#endif
}

结果输出：

总体感觉engine模型是要比onnx模型要快一些，输出的log信息也相对少一些。

后续的话，Tensorflow的pb模型或者pytorch的onnx模型，都可以以某种方式（TensorRT模型装换 - 小小马进阶笔记 - 博客园）转换为TensorRT的engine模型，转换的过程应该是有对网络层进行优化，所以速度才提升了。但是也要留意精度的损失程度。

同时，在使用trtexec这个工具的时候，也可以考虑设置一些参数，比如int8，对模型进行量化

利用TensorRT实现INT8量化感知训练QAT_ZONGXP的博客-CSDN博客_int8量化训练

参考命令行如下：

trtexec.exe --onnx=C:\Project\TensorRT-8.0.3.4\data\mnist\mnist.onnx --saveEngine=C:\Project\TensorRT-8.0.3.4\data\mnist\mnist.engine --int8