TensorRT Samples: MNIST(serialize TensorRT model)

关于TensorRT的介绍可以参考： http://blog.csdn.net/fengbingchun/article/details/78469551   

这里实现在构建阶段将TensorRT model序列化存到本地文件，然后在部署阶段直接load TensorRT model序列化的文件进行推理，mnist_infer.cpp文件内容如下：

#include <iostream>
#include <string>
#include <tuple>
#include <fstream>
#include <memory>

#include <cuda_runtime_api.h>
#include <NvInfer.h>
#include <NvCaffeParser.h>
#include <opencv2/opencv.hpp>

#include "common.hpp"

// 序列化TensorRT模型，然后load TensorRT模型进行推理

namespace {
typedef std::tuple<int, int, int, std::string, std::string> DATA_INFO; // intput width, input height, output size, input blob name, output blob name

int caffeToGIEModel(const std::string& deployFile,	// name for caffe prototxt
					 const std::string& modelFile,	// name for model 
					 const std::vector<std::string>& outputs, // network outputs
					 unsigned int maxBatchSize,	// batch size - NB must be at least as large as the batch we want to run with)
					 Logger logger, const std::string& engine_file) 
{
	// create the builder
	nvinfer1::IBuilder* builder = nvinfer1::createInferBuilder(logger);

	// parse the caffe model to populate the network, then set the outputs
	nvinfer1::INetworkDefinition* network = builder->createNetwork();
	nvcaffeparser1::ICaffeParser* parser = nvcaffeparser1::createCaffeParser();
	const nvcaffeparser1::IBlobNameToTensor* blobNameToTensor = parser->parse(deployFile.c_str(), modelFile.c_str(), *network, nvinfer1::DataType::kFLOAT);

	// specify which tensors are outputs
	for (auto& s : outputs)
		network->markOutput(*blobNameToTensor->find(s.c_str()));

	// Build the engine
	builder->setMaxBatchSize(maxBatchSize);
	builder->setMaxWorkspaceSize(1 << 20);

	nvinfer1::ICudaEngine* engine = builder->buildCudaEngine(*network);
	CHECK(engine != nullptr);

	// we don't need the network any more, and we can destroy the parser
	network->destroy();
	parser->destroy();

	// serialize the engine, then close everything down
	nvinfer1::IHostMemory* gieModelStream = engine->serialize(); // GIE model
	fprintf(stdout, "allocate memory size: %d bytes\n", gieModelStream->size());
	std::ofstream outfile(engine_file.c_str(), std::ios::out | std::ios::binary);
	if (!outfile.is_open()) {
		fprintf(stderr, "fail to open file to write: %s\n", engine_file.c_str());
        return -1;
    }
	unsigned char* p = (unsigned char*)gieModelStream->data();
	outfile.write((char*)p, gieModelStream->size());
	outfile.close();

	engine->destroy();
	builder->destroy();
	if (gieModelStream) gieModelStream->destroy();	
	nvcaffeparser1::shutdownProtobufLibrary();

	return 0;
}

int doInference(nvinfer1::IExecutionContext& context, const float* input, float* output, int batchSize, const DATA_INFO& info)
{
	const nvinfer1::ICudaEngine& engine = context.getEngine();
	// input and output buffer pointers that we pass to the engine - the engine requires exactly IEngine::getNbBindings(),
	// of these, but in this case we know that there is exactly one input and one output.
	CHECK(engine.getNbBindings() == 2);
	void* buffers[2];

	// In order to bind the buffers, we need to know the names of the input and output tensors.
	// note that indices are guaranteed to be less than IEngine::getNbBindings()
	int inputIndex = engine.getBindingIndex(std::get<3>(info).c_str()), 
		outputIndex = engine.getBindingIndex(std::get<4>(info).c_str());

	// create GPU buffers and a stream
	checkCudaErrors(cudaMalloc(&buffers[inputIndex], batchSize * std::get<1>(info) * std::get<0>(info) * sizeof(float)));
	checkCudaErrors(cudaMalloc(&buffers[outputIndex], batchSize * std::get<2>(info) * sizeof(float)));

	cudaStream_t stream;
	checkCudaErrors(cudaStreamCreate(&stream));

	// DMA the input to the GPU,  execute the batch asynchronously, and DMA it back:
	checkCudaErrors(cudaMemcpyAsync(buffers[inputIndex], input, batchSize * std::get<1>(info) * std::get<0>(info) * sizeof(float), cudaMemcpyHostToDevice, stream));
	context.enqueue(batchSize, buffers, stream, nullptr);
	checkCudaErrors(cudaMemcpyAsync(output, buffers[outputIndex], batchSize * std::get<2>(info) * sizeof(float), cudaMemcpyDeviceToHost, stream));
	cudaStreamSynchronize(stream);

	// release the stream and the buffers
	cudaStreamDestroy(stream);
	checkCudaErrors(cudaFree(buffers[inputIndex]));
	checkCudaErrors(cudaFree(buffers[outputIndex]));

	return 0;
}

} // namesapce

int test_mnist_infer()
{
	// 1. build phase
	// stuff we know about the network and the caffe input/output blobs
	const DATA_INFO info(28, 28, 10, "data", "prob");
	const std::string deploy_file { "models/mnist.prototxt" };
	const std::string model_file { "models/mnist.caffemodel" };
	const std::string mean_file { "models/mnist_mean.binaryproto" };
	const std::string engine_file { "tensorrt_mnist.model" };

	Logger logger; // multiple instances of IRuntime and/or IBuilder must all use the same logger

   	CHECK(caffeToGIEModel(deploy_file, model_file, std::vector<std::string>{std::get<4>(info)}, 1, logger, engine_file) == 0);

	// 2. deploy phase
	// parse the mean file and 	subtract it from the image
	nvcaffeparser1::ICaffeParser* parser = nvcaffeparser1::createCaffeParser();
	nvcaffeparser1::IBinaryProtoBlob* meanBlob = parser->parseBinaryProto(mean_file.c_str());
	parser->destroy();

	std::ifstream in_file(engine_file.c_str(), std::ios::in | std::ios::binary);
	if (!in_file.is_open()) {
		fprintf(stderr, "fail to open file to write: %s\n", engine_file.c_str());
        return -1;
    }

	std::streampos begin, end;
    begin = in_file.tellg();
    in_file.seekg(0, std::ios::end);
    end = in_file.tellg();
	std::size_t size = end - begin;
	fprintf(stdout, "engine file size: %d bytes\n", size);
	in_file.seekg(0, std::ios::beg);
	std::unique_ptr<unsigned char[]> engine_data(new unsigned char[size]);
	in_file.read((char*)engine_data.get(), size);
	in_file.close();

	// deserialize the engine 
	nvinfer1::IRuntime* runtime = nvinfer1::createInferRuntime(logger);	
	nvinfer1::ICudaEngine* engine = runtime->deserializeCudaEngine((const void*)engine_data.get(), size, nullptr);
	nvinfer1::IExecutionContext* context = engine->createExecutionContext();
	const float* meanData = reinterpret_cast<const float*>(meanBlob->getData());

	const std::string image_path{ "images/digit/" };
	for (int i = 0; i < 10; ++i) {
		const std::string image_name = image_path + std::to_string(i) + ".png";
		cv::Mat mat = cv::imread(image_name, 0);
		if (!mat.data) {
			fprintf(stderr, "read image fail: %s\n", image_name.c_str());
			return -1;
		}

		cv::resize(mat, mat, cv::Size(std::get<0>(info), std::get<1>(info)));
		mat.convertTo(mat, CV_32FC1);

		float data[std::get<1>(info)*std::get<0>(info)];
		const float* p = (float*)mat.data;
		for (int j = 0; j < std::get<1>(info)*std::get<0>(info); ++j) {
			data[j] = p[j] - meanData[j];
		}

		// run inference
		float prob[std::get<2>(info)];
		doInference(*context, data, prob, 1, info);

		float val{-1.f};
		int idx{-1};

		for (int t = 0; t < std::get<2>(info); ++t) {
			if (val < prob[t]) {
				val = prob[t];
				idx = t;
			}
		}

		fprintf(stdout, "expected value: %d, actual value: %d, probability: %f\n", i, idx, val);
	}

	meanBlob->destroy();
	// destroy the engine
	context->destroy();
	engine->destroy();
	runtime->destroy();

	return 0;
}

测试图像如下：

测试结果如下：与http://blog.csdn.net/fengbingchun/article/details/78552908 结果一致

测试代码编译步骤如下(ReadMe.txt)：

在Linux下通过CMake编译TensorRT_Test中的测试代码步骤：
1. 将终端定位到CUDA_Test/prj/linux_tensorrt_cmake，依次执行如下命令：
	$ mkdir build
	$ cd build
	$ cmake ..
	$ make (生成TensorRT_Test执行文件)
	$ ln -s ../../../test_data/models  ./ (将models目录软链接到build目录下)
	$ ln -s ../../../test_data/images  ./ (将images目录软链接到build目录下)
	$ ./TensorRT_Test
2. 对于有需要用OpenCV参与的读取图像的操作，需要先将对应文件中的图像路径修改为Linux支持的路径格式

GitHub：  https://github.com/fengbingchun/CUDA_Test