TensorRT Samples: MNIST API

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

以下是参考TensorRT 2.1.2中的sampleMNISTAPI.cpp文件改写的实现对手写数字0-9识别的测试代码，各个文件内容如下：

common.hpp:

#ifndef FBC_TENSORRT_TEST_COMMON_HPP_
#define FBC_TENSORRT_TEST_COMMON_HPP_#include <cuda_runtime.h>
#include <device_launch_parameters.h>
#include <NvInfer.h>template< typename T >
static inline int check_Cuda(T result, const char * const func, const char * const file, const int line)
{if (result) {fprintf(stderr, "Error CUDA: at %s: %d, error code=%d, func: %s\n", file, line, static_cast<unsigned int>(result), func);cudaDeviceReset(); // Make sure we call CUDA Device Reset before exitingreturn -1;}
}template< typename T >
static inline int check(T result, const char * const func, const char * const file, const int line)
{if (result) {fprintf(stderr, "Error: at %s: %d, error code=%d, func: %s\n", file, line, static_cast<unsigned int>(result), func);return -1;}
}#define checkCudaErrors(val) check_Cuda((val), __FUNCTION__, __FILE__, __LINE__)
#define checkErrors(val) check((val), __FUNCTION__, __FILE__, __LINE__)#define CHECK(x) { \if (x) {} \else { fprintf(stderr, "Check Failed: %s, file: %s, line: %d\n", #x, __FILE__, __LINE__); return -1; } \
}// Logger for GIE info/warning/errors
class Logger : public nvinfer1::ILogger
{void log(Severity severity, const char* msg) override{// suppress info-level messagesif (severity != Severity::kINFO)std::cout << msg << std::endl;}
};#endif // FBC_TENSORRT_TEST_COMMON_HPP_

mnist_api.cpp:

#include <string>
#include <fstream>
#include <iostream>
#include <map>
#include <tuple>#include <NvInfer.h>
#include <NvCaffeParser.h>
#include <cuda_runtime_api.h>
#include <opencv2/opencv.hpp>#include "common.hpp"// reference: TensorRT-2.1.2/samples/sampleMNIST/sampleMNISTAPI.cpp// intput width, input height, output size, input blob name, output blob name, weight file, mean file
typedef std::tuple<int, int, int, std::string, std::string, std::string, std::string> DATA_INFO;// Our weight files are in a very simple space delimited format.
// [type] [size] <data x size in hex>
static std::map<std::string, nvinfer1::Weights> loadWeights(const std::string& file)
{std::map<std::string, nvinfer1::Weights> weightMap;std::ifstream input(file);if (!input.is_open()) {fprintf(stderr, "Unable to load weight file: %s\n", file.c_str());return  weightMap;}int32_t count;input >> count;if (count <= 0) {fprintf(stderr, "Invalid weight map file: %d\n", count);return weightMap;}while(count--) {nvinfer1:: Weights wt{nvinfer1::DataType::kFLOAT, nullptr, 0};uint32_t type, size;std::string name;input >> name >> std::dec >> type >> size;wt.type = static_cast<nvinfer1::DataType>(type);if (wt.type == nvinfer1::DataType::kFLOAT) {uint32_t *val = reinterpret_cast<uint32_t*>(malloc(sizeof(val) * size));for (uint32_t x = 0, y = size; x < y; ++x) {input >> std::hex >> val[x];}wt.values = val;} else if (wt.type == nvinfer1::DataType::kHALF) {uint16_t *val = reinterpret_cast<uint16_t*>(malloc(sizeof(val) * size));for (uint32_t x = 0, y = size; x < y; ++x) {input >> std::hex >> val[x];}wt.values = val;}wt.count = size;weightMap[name] = wt;}return weightMap;
}// Creat the Engine using only the API and not any parser.
static nvinfer1::ICudaEngine* createMNISTEngine(unsigned int maxBatchSize, nvinfer1::IBuilder* builder, nvinfer1::DataType dt, const DATA_INFO& info)
{nvinfer1::INetworkDefinition* network = builder->createNetwork();//  Create input of shape { 1, 1, 28, 28 } with name referenced by INPUT_BLOB_NAMEauto data = network->addInput(std::get<3>(info).c_str(), dt, nvinfer1::DimsCHW{ 1, std::get<1>(info), std::get<0>(info)});assert(data != nullptr);// Create a scale layer with default power/shift and specified scale parameter.float scale_param = 0.0125f;nvinfer1::Weights power{nvinfer1::DataType::kFLOAT, nullptr, 0};nvinfer1::Weights shift{nvinfer1::DataType::kFLOAT, nullptr, 0};nvinfer1::Weights scale{nvinfer1::DataType::kFLOAT, &scale_param, 1};auto scale_1 = network->addScale(*data,   nvinfer1::ScaleMode::kUNIFORM, shift, scale, power);assert(scale_1 != nullptr);// Add a convolution layer with 20 outputs and a 5x5 filter.std::map<std::string, nvinfer1::Weights> weightMap = loadWeights(std::get<5>(info));auto conv1 = network->addConvolution(*scale_1->getOutput(0), 20, nvinfer1::DimsHW{5, 5}, weightMap["conv1filter"], weightMap["conv1bias"]);assert(conv1 != nullptr);conv1->setStride(nvinfer1::DimsHW{1, 1});// Add a max pooling layer with stride of 2x2 and kernel size of 2x2.auto pool1 = network->addPooling(*conv1->getOutput(0), nvinfer1::PoolingType::kMAX, nvinfer1::DimsHW{2, 2});assert(pool1 != nullptr);pool1->setStride(nvinfer1::DimsHW{2, 2});// Add a second convolution layer with 50 outputs and a 5x5 filter.auto conv2 = network->addConvolution(*pool1->getOutput(0), 50, nvinfer1::DimsHW{5, 5}, weightMap["conv2filter"], weightMap["conv2bias"]);assert(conv2 != nullptr);conv2->setStride(nvinfer1::DimsHW{1, 1});// Add a second max pooling layer with stride of 2x2 and kernel size of 2x3>auto pool2 = network->addPooling(*conv2->getOutput(0), nvinfer1::PoolingType::kMAX, nvinfer1::DimsHW{2, 2});assert(pool2 != nullptr);pool2->setStride(nvinfer1::DimsHW{2, 2});// Add a fully connected layer with 500 outputs.auto ip1 = network->addFullyConnected(*pool2->getOutput(0), 500, weightMap["ip1filter"], weightMap["ip1bias"]);assert(ip1 != nullptr);// Add an activation layer using the ReLU algorithm.auto relu1 = network->addActivation(*ip1->getOutput(0), nvinfer1::ActivationType::kRELU);assert(relu1 != nullptr);// Add a second fully connected layer with 20 outputs.auto ip2 = network->addFullyConnected(*relu1->getOutput(0), std::get<2>(info), weightMap["ip2filter"], weightMap["ip2bias"]);assert(ip2 != nullptr);// Add a softmax layer to determine the probability.auto prob = network->addSoftMax(*ip2->getOutput(0));assert(prob != nullptr);prob->getOutput(0)->setName(std::get<4>(info).c_str());network->markOutput(*prob->getOutput(0));// Build the enginebuilder->setMaxBatchSize(maxBatchSize);builder->setMaxWorkspaceSize(1 << 20);auto engine = builder->buildCudaEngine(*network);// we don't need the network any morenetwork->destroy();// Once we have built the cuda engine, we can release all of our held memory.for (auto &mem : weightMap) {free((void*)(mem.second.values));}return engine;
}static int APIToModel(unsigned int maxBatchSize, // batch size - NB must be at least as large as the batch we want to run with)nvinfer1::IHostMemory** modelStream, Logger logger, const DATA_INFO& info)
{// create the buildernvinfer1::IBuilder* builder = nvinfer1::createInferBuilder(logger);// create the model to populate the network, then set the outputs and create an enginenvinfer1::ICudaEngine* engine = createMNISTEngine(maxBatchSize, builder, nvinfer1::DataType::kFLOAT, info);CHECK(engine != nullptr);// serialize the engine, then close everything down(*modelStream) = engine->serialize();engine->destroy();builder->destroy();return 0;
}static int doInference(nvinfer1::IExecutionContext& context, 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 streamcheckCudaErrors(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 bufferscudaStreamDestroy(stream);checkCudaErrors(cudaFree(buffers[inputIndex]));checkCudaErrors(cudaFree(buffers[outputIndex]));return 0;
}int test_mnist_api()
{Logger logger; // multiple instances of IRuntime and/or IBuilder must all use the same logger// stuff we know about the network and the caffe input/output blobsconst DATA_INFO info(28, 28, 10, "data", "prob", "models/mnistapi.wts", "models/mnist_mean.binaryproto");// create a model using the API directly and serialize it to a streamnvinfer1::IHostMemory* modelStream{ nullptr };APIToModel(1, &modelStream, logger, info);// parse the mean file produced by caffe and subtract it from the imagenvcaffeparser1::ICaffeParser* parser = nvcaffeparser1::createCaffeParser();nvcaffeparser1::IBinaryProtoBlob* meanBlob = parser->parseBinaryProto(std::get<6>(info).c_str());parser->destroy();const float* meanData = reinterpret_cast<const float*>(meanBlob->getData());nvinfer1::IRuntime* runtime = nvinfer1::createInferRuntime(logger);nvinfer1::ICudaEngine* engine = runtime->deserializeCudaEngine(modelStream->data(), modelStream->size(), nullptr);nvinfer1::IExecutionContext* context = engine->createExecutionContext();uint8_t fileData[std::get<1>(info) * std::get<0>(info)];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 inferencefloat 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();if (modelStream) modelStream->destroy();// destroy the enginecontext->destroy();engine->destroy();runtime->destroy();return 0;
}

测试图像如下：

执行结果如下：(与 http://blog.csdn.net/fengbingchun/article/details/78552908 中结果一致)

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

TensorRT Samples: MNIST API相关推荐

TensorRT Samples: MNIST(Plugin, add a custom layer)
关于TensorRT的介绍可以参考:http://blog.csdn.net/fengbingchun/article/details/78469551 以下是参考TensorRT 2.1.2中的sa ...
TensorRT Samples: MNIST
关于TensorRT的介绍可以参考: http://blog.csdn.net/fengbingchun/article/details/78469551 以下是参考TensorRT 2.1.2中的 ...
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 mod ...
TensorRT(3)-C++ API使用：mnist手写体识别
本节将介绍如何使用tensorRT C++ API 进行网络模型创建. 1 使用C++ API 进行 tensorRT 模型创建还是通过 tensorRT官方给的一个例程来学习. 还是mnist手写 ...
TensorRT/samples/common/argsParser.h源碼研讀
TensorRT/samples/common/argsParser.h源碼研讀 argsParser.h namespace struct的繼承 caffe特有參數 UFF格式不需要typedef ...
TensorRT Samples: CharRNN
关于TensorRT的介绍可以参考: http://blog.csdn.net/fengbingchun/article/details/78469551 以下是参考TensorRT 2.1.2中的 ...
TensorRT Samples: GoogleNet
关于TensorRT的介绍可以参考: http://blog.csdn.net/fengbingchun/article/details/78469551 以下是参考TensorRT 2.1.2中的 ...
TensorRT 下不同 API 推理时间的对比实验
文章目录项目简介实验结果环境构建宿主机基础环境基础镜像拉取安装其他库项目代码文件说明项目简介基于 TensorRT 8.2.4 版本,具体环境见下面的环境构建部分目标: 对比 p ...
TensorRT(6)-INT8 inference
这一节通过官方例程介绍 INT8 inference mode. 例程位于 /usr/src/tensorrt/samples/sampleINT8 ,是基于mnist的,大体流程是一致的. 流程同 ...

TensorRT Samples: MNIST API

TensorRT Samples: MNIST API相关推荐

最新文章

热门文章