文章目录
- 一、模型转换 onnx2trt
- 二、配置环境变量
- 三、调用推理
- python示例代码
- C++ 代码示例
测试使用：【Win10+cuda11.0+cudnn8.2.1+TensorRT8.2.5.1】 关于安装
一、模型转换 onnx2trt
方法1：使用wang-xinyu/tensorrtx部署yolov5方法：https://wangsp.blog.csdn.net/article/details/121718501
方法2：使用tensorRT转成engine
方法3：使用C++ onnx_tensorrt将onnx转为trt 的推理engine 参考 【python 方法参考】
方法4：直接使用TensorRT部署onnx【参考】
- 使用TensorRT部署pytorch模型（c++推理）【参考】
- TensorRT-pytorch权重文件转engine【参考】
- pth->onnx->下载好TensorRT库, 进入~/samples/trtexec, 运行make，生成.engine->python run engine 【参考】 【参考2】
使用 trtexec工具转engine
使用 ./trtexec --help 查看命令：
#生成静态batchsize的engine
./trtexec --onnx=<onnx_file> \ #指定onnx模型文件
--explicitBatch \ #在构建引擎时使用显式批大小(默认=隐式)显示批处理
--saveEngine=<tensorRT_engine_file> \ #输出engine
--workspace=<size_in_megabytes> \ #设置工作空间大小单位是MB(默认为16MB)
--fp16 #除了fp32之外，还启用fp16精度(默认=禁用)
#生成动态batchsize的engine
./trtexec --onnx=<onnx_file> \ #指定onnx模型文件
--minShapes=input:<shape_of_min_batch> \ #最小的NCHW
--optShapes=input:<shape_of_opt_batch> \ #最佳输入维度，跟maxShapes一样就好
--maxShapes=input:<shape_of_max_batch> \ #最大输入维度
--workspace=<size_in_megabytes> \ #设置工作空间大小单位是MB(默认为16MB)
--saveEngine=<engine_file> \ #输出engine
--fp16 #除了fp32之外，还启用fp16精度(默认=禁用)
#小尺寸的图片可以多batchsize即8x3x416x416
/home/zxl/TensorRT-7.2.3.4/bin/trtexec --onnx=yolov4_-1_3_416_416_dynamic.onnx \
--minShapes=input:1x3x416x416 \
--optShapes=input:8x3x416x416 \
--maxShapes=input:8x3x416x416 \
--workspace=4096 \
--saveEngine=yolov4_-1_3_416_416_dynamic_b8_fp16.engine \
--fp16
#由于内存不够了所以改成4x3x608x608
/home/zxl/TensorRT-7.2.3.4/bin/trtexec --onnx=yolov4_-1_3_608_608_dynamic.onnx \
--minShapes=input:1x3x608x608 \
--optShapes=input:4x3x608x608 \
--maxShapes=input:4x3x608x608 \
--workspace=4096 \
--saveEngine=yolov4_-1_3_608_608_dynamic_b4_fp16.engine \
--fp16
测试，执行：
二、配置环境变量
################ TenorRT 包含目录 ######################
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.0\include;
D:\opencv_build\install\include;D:\opencv_build\install\include\opencv2;
D:\Downloads\cuda_cudnn_TensorRT8\TensorRT-8.2.5.1.Windows10.x86_64.cuda-11.4.cudnn8.2\TensorRT-8.2.5.1\samples\common
#################### TenorRT 库目录 ############################
D:\opencv_build\install\x64\vc16\lib\*.lib
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.0\lib\x64\*.lib
三、调用推理
使用pycuda【下载】地址。模型训练代码来自 https://github.com/bubbliiiing
安装pycuda 对应python的版本：pycuda-2020.1+cuda101-cp38-cp38-win_amd64.whl
安装tensorrt对应python的版本：tensorrt-8.2.5.1-cp38-none-win_amd64.whl（来自TensorRT-8.2.5.1.Windows10.x86_64.cuda-11.4.cudnn8.2\TensorRT-8.2.5.1\python目录下）
TensorRT调用步骤
- 创建IBuilder的指针builder
- 设置推理的显存大小
- 设置推理的模式，float或者int
- 利用builder创建ICudaEngine的实例engine
- 由engine创建上下文context
- 利用context进行推理，得到结果
- 释放显存空间
python示例代码
# --*-- coding:utf-8 --*--
import pycuda.autoinit
import pycuda.driver as cuda
import tensorrt as trt
import torch
import time
from PIL import Image
import cv2, os
import torchvision
import numpy as np
filename = '/home/img.png'
max_batch_size = 1
onnx_model_path = "./resnet18.onnx"
TRT_LOGGER = trt.Logger(trt.Logger.WARNING)
def get_img_np_nchw(filename):
image = cv2.imread(filename)
image_cv = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
image_cv = cv2.resize(image_cv, (224, 224))
miu = np.array([0.485, 0.456, 0.406]).reshape(3, 1, 1)
std = np.array([0.229, 0.224, 0.225]).reshape(3, 1, 1)
img_np = np.array(image_cv, dtype=np.float) / 255.
img_np = img_np.transpose((2, 0, 1))
img_np -= miu
img_np /= std
img_np_nchw = img_np[np.newaxis]
img_np_nchw = np.tile(img_np_nchw, (max_batch_size, 1, 1, 1))
return img_np_nchw
class HostDeviceMem(object):
def __init__(self, host_mem, device_mem):
"""
host_mem: cpu memory
device_mem: gpu memory
"""
self.host = host_mem
self.device = device_mem
def __str__(self):
return "Host:\n" + str(self.host) + "\nDevice:\n" + str(self.device)
def __repr__(self):
return self.__str__()
def allocate_buffers(engine):
inputs, outputs, bindings = [], [], []
stream = cuda.Stream()
for binding in engine:
# print(binding) # 绑定的输入输出
# print(engine.get_binding_shape(binding)) # get_binding_shape 是变量的大小
size = trt.volume(engine.get_binding_shape(binding)) * engine.max_batch_size
# volume 计算可迭代变量的空间，指元素个数
# size = trt.volume(engine.get_binding_shape(binding)) # 如果采用固定bs的onnx，则采用该句
dtype = trt.nptype(engine.get_binding_dtype(binding))
# get_binding_dtype 获得binding的数据类型
# nptype等价于numpy中的dtype，即数据类型
# allocate host and device buffers
host_mem = cuda.pagelocked_empty(size, dtype) # 创建锁业内存
device_mem = cuda.mem_alloc(host_mem.nbytes) # cuda分配空间
# print(int(device_mem)) # binding在计算图中的缓冲地址
bindings.append(int(device_mem))
# append to the appropriate list
if engine.binding_is_input(binding):
inputs.append(HostDeviceMem(host_mem, device_mem))
else:
outputs.append(HostDeviceMem(host_mem, device_mem))
return inputs, outputs, bindings, stream
def get_engine(max_batch_size=1, onnx_file_path="", engine_file_path="", fp16_mode=False, save_engine=False):
"""
params max_batch_size: 预先指定大小好分配显存
params onnx_file_path: onnx文件路径
params engine_file_path: 待保存的序列化的引擎文件路径
params fp16_mode: 是否采用FP16
params save_engine: 是否保存引擎
returns: ICudaEngine
"""
# 如果已经存在序列化之后的引擎，则直接反序列化得到cudaEngine
if os.path.exists(engine_file_path):
print("Reading engine from file: {}".format(engine_file_path))
with open(engine_file_path, 'rb') as f, \
trt.Runtime(TRT_LOGGER) as runtime:
return runtime.deserialize_cuda_engine(f.read()) # 反序列化
else: # 由onnx创建cudaEngine
# 使用logger创建一个builder
# builder创建一个计算图 INetworkDefinition
explicit_batch = 1 << (int)(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH)
# In TensorRT 7.0, the ONNX parser only supports full-dimensions mode, meaning that your network definition must be created with the explicitBatch flag set. For more information, see Working With Dynamic Shapes.
with trt.Builder(TRT_LOGGER) as builder, \
builder.create_network(explicit_batch) as network, \
trt.OnnxParser(network, TRT_LOGGER) as parser: # 使用onnx的解析器绑定计算图，后续将通过解析填充计算图
builder.max_workspace_size = 1 << 30 # 预先分配的工作空间大小,即ICudaEngine执行时GPU最大需要的空间
builder.max_batch_size = max_batch_size # 执行时最大可以使用的batchsize
builder.fp16_mode = fp16_mode
# 解析onnx文件，填充计算图
if not os.path.exists(onnx_file_path):
quit("ONNX file {} not found!".format(onnx_file_path))
print('loading onnx file from path {} ...'.format(onnx_file_path))
with open(onnx_file_path, 'rb') as model: # 二值化的网络结果和参数
print("Begining onnx file parsing")
parser.parse(model.read()) # 解析onnx文件
# parser.parse_from_file(onnx_file_path) # parser还有一个从文件解析onnx的方法
print("Completed parsing of onnx file")
# 填充计算图完成后，则使用builder从计算图中创建CudaEngine
print("Building an engine from file{}' this may take a while...".format(onnx_file_path))
#################
print(network.get_layer(network.num_layers - 1).get_output(0).shape)
# network.mark_output(network.get_layer(network.num_layers -1).get_output(0))
engine = builder.build_cuda_engine(network) # 注意，这里的network是INetworkDefinition类型，即填充后的计算图
print("Completed creating Engine")
if save_engine: # 保存engine供以后直接反序列化使用
with open(engine_file_path, 'wb') as f:
f.write(engine.serialize()) # 序列化
return engine
def do_inference(context, bindings, inputs, outputs, stream, batch_size=1):
# Transfer data from CPU to the GPU.
[cuda.memcpy_htod_async(inp.device, inp.host, stream) for inp in inputs]
# htod： host to device 将数据由cpu复制到gpu device
# Run inference.
context.execute_async_v2(bindings=bindings, stream_handle=stream.handle)
# 当创建network时显式指定了batchsize， 则使用execute_async_v2, 否则使用execute_async
# Transfer predictions back from the GPU.
[cuda.memcpy_dtoh_async(out.host, out.device, stream) for out in outputs]
# gpu to cpu
# Synchronize the stream
stream.synchronize()
# Return only the host outputs.
return [out.host for out in outputs]
def postprocess_the_outputs(h_outputs, shape_of_output):
h_outputs = h_outputs.reshape(*shape_of_output)
return h_outputs
img_np_nchw = get_img_np_nchw(filename).astype(np.float32)
# These two modes are depend on hardwares
fp16_mode = False
trt_engine_path = "./model_fp16_{}.trt".format(fp16_mode)
# Build an cudaEngine
engine = get_engine(max_batch_size, onnx_model_path, trt_engine_path, fp16_mode)
# 创建CudaEngine之后,需要将该引擎应用到不同的卡上配置执行环境
context = engine.create_execution_context()
inputs, outputs, bindings, stream = allocate_buffers(engine) # input, output: host # bindings
# Do inference
shape_of_output = (max_batch_size, 1000)
# Load data to the buffer
inputs[0].host = img_np_nchw.reshape(-1)
# inputs[1].host = ... for multiple input
t1 = time.time()
trt_outputs = do_inference(context, bindings=bindings, inputs=inputs, outputs=outputs, stream=stream) # numpy data
t2 = time.time()
feat = postprocess_the_outputs(trt_outputs[0], shape_of_output)
print('TensorRT ok')
model = torchvision.models.resnet18(pretrained=True).cuda()
resnet_model = model.eval()
input_for_torch = torch.from_numpy(img_np_nchw).cuda()
t3 = time.time()
feat_2 = resnet_model(input_for_torch)
t4 = time.time()
feat_2 = feat_2.cpu().data.numpy()
print('Pytorch ok!')
mse = np.mean((feat - feat_2) ** 2)
print("Inference time with the TensorRT engine: {}".format(t2 - t1))
print("Inference time with the PyTorch model: {}".format(t4 - t3))
print('MSE Error = {}'.format(mse))
print('All completed!')
C++ 代码示例
TensorRT 傻瓜式部署流程：参考
#include <string>
#include <algorithm>
#include <assert.h>
#include <cmath>
#include <cuda_runtime_api.h>
#include <fstream>
#include <iomanip>
#include <iostream>
#include <sstream>
#include <sys/stat.h>
#include <time.h>
#include <opencv2/opencv.hpp>
#include <io.h>
#include "NvInfer.h"
#include "NvOnnxParser.h"
#include "argsParser.h"
#include "logger.h"
#include "common.h"
#ifndef NOMINMAX
#ifndef max_idx
#define max_idx(a,b) (((a) > (b)) ? (0) : (1))
#endif
#endif /* NOMINMAX */
#define DebugP(x) std::cout << "Line" << __LINE__ << " " << #x << "=" << x << std::endl
using namespace nvinfer1;
samplesCommon::Args gArgs;
using namespace sample;
static const int INPUT_H = 480;
static const int INPUT_W = 480;
static const int INPUT_C = 3;
static constexpr int INPUT_SIZE = INPUT_H * INPUT_W * 3;
static constexpr int OUTPUT_SIZE = INPUT_H * INPUT_W * 2;
static const cv::Size newShape = cv::Size(INPUT_W, INPUT_H);
const std::string trtModelName = "D:\\xxx.engine";
const std::string onnxModeName = "D:\\xxx.onnx";
const std::string file_name = "D:\\xxx.jpg";
struct TensorRT {
IExecutionContext* context;
ICudaEngine* engine;
IRuntime* runtime;
};
void image_to_center(const cv::Mat& image, cv::Mat& outImage, cv::Mat& IM, const cv::Scalar& color)
{
cv::Size shape = image.size();
float scale_xy = std::min((float)newShape.height / (float)shape.height,
(float)newShape.width / (float)shape.width);
cv::Mat M = (cv::Mat_<float>(2, 3) <<
scale_xy, 0, -scale_xy * (float)shape.width * 0.5 + (float)newShape.width * 0.5,
0, scale_xy, -scale_xy * (float)shape.height * 0.5 + (float)newShape.height * 0.5);
cv::invertAffineTransform(M, IM);
cv::warpAffine(image, outImage, M, newShape, 1, 0, color);
}
void center_to_image(const cv::Mat& image, cv::Mat& outImage, cv::Mat& IM)
{
cv::warpAffine(image, outImage, IM, newShape);
}
void normal_image2blob(float* blob, cv::Mat& img) {
for (int c = 0; c < 3; ++c) {
for (int i = 0; i < img.rows; ++i) {
cv::Vec3b* p1 = img.ptr<cv::Vec3b>(i);
for (int j = 0; j < img.cols; ++j) {
blob[c * img.cols * img.rows + i * img.cols + j] = p1[j][c] * 0.00392156862745098;
}
}
}
}
bool onnxToTRTModel(const std::string& modelFile, // name of the onnx model
unsigned int maxBatchSize, // batch size - NB must be at least as large as the batch we want to run with
IHostMemory*& trtModelStream) // output buffer for the TensorRT model
{
// create the builder
IBuilder* builder = createInferBuilder(gLogger.getTRTLogger());
assert(builder != nullptr);
nvinfer1::INetworkDefinition* network = builder->createNetworkV2(maxBatchSize);
nvinfer1::IBuilderConfig* config = builder->createBuilderConfig();
config->setMaxWorkspaceSize(1 << 20);
// parser
auto parser = nvonnxparser::createParser(*network, gLogger.getTRTLogger());
if (!parser->parseFromFile(modelFile.c_str(), static_cast<int>(gLogger.getReportableSeverity())))
{
gLogError << "Failure while parsing ONNX file" << std::endl;
return false;
}
if (builder->platformHasFastFp16()) {
config->setFlag(nvinfer1::BuilderFlag::kFP16);
}
else {
std::cout << "This platform does not support fp16" << std::endl;
}
// Build the engine
ICudaEngine* engine = builder->buildEngineWithConfig(*network, *config);
assert(engine);
// serialize the engine, then close everything down
trtModelStream = engine->serialize();
parser->destroy();
engine->destroy();
network->destroy();
builder->destroy();
std::ofstream ofs(trtModelName.c_str(), std::ios::out | std::ios::binary);
ofs.write((char*)(trtModelStream->data()), trtModelStream->size());
ofs.close();
DebugP("Trt model save success!");
return true;
}
TensorRT* LoadNet(const char* trtFileName)
{
std::ifstream t(trtFileName, std::ios::in | std::ios::binary);
std::stringstream tempStream;
tempStream << t.rdbuf();
t.close();
DebugP("TRT File Loaded successfully!");
tempStream.seekg(0, std::ios::end);
const int modelSize = tempStream.tellg();
tempStream.seekg(0, std::ios::beg);
void* modelMem = malloc(modelSize);
tempStream.read((char*)modelMem, modelSize);
IRuntime* runtime = createInferRuntime(gLogger);
if (runtime == nullptr)
{
DebugP("Build Runtime Failure");
return 0;
}
if (gArgs.useDLACore >= 0)
{
runtime->setDLACore(gArgs.useDLACore);
}
ICudaEngine* engine = runtime->deserializeCudaEngine(modelMem, modelSize, nullptr);
if (engine == nullptr)
{
DebugP("Build Engine Failure");
return 0;
}
IExecutionContext* context = engine->createExecutionContext();
if (context == nullptr)
{
DebugP("Build Context Failure");
return 0;
}
TensorRT* trt = new TensorRT();
trt->context = context;
trt->engine = engine;
trt->runtime = runtime;
DebugP("Build trt Model Success!");
return trt;
}
void doInference(IExecutionContext& context, float* input, float* output, int batchSize)
{
const ICudaEngine& engine = context.getEngine();
assert(engine.getNbBindings() == 2);
void* buffers[2];
int inputIndex, outputIndex;
for (int b = 0; b < engine.getNbBindings(); ++b)
{
if (engine.bindingIsInput(b))
inputIndex = b;
else
outputIndex = b;
}
std::cout << "inputIndex=" << inputIndex << "\n";
std::cout << "outputIndex=" << outputIndex << "\n";
// create GPU buffers and a stream
CHECK(cudaMalloc(&buffers[inputIndex], batchSize * INPUT_C * INPUT_H * INPUT_W * sizeof(float)));
CHECK(cudaMalloc(&buffers[outputIndex], batchSize * OUTPUT_SIZE * sizeof(float)));
cudaStream_t stream;
CHECK(cudaStreamCreate(&stream));
// DMA the input to the GPU, execute the batch asynchronously, and DMA it back:
CHECK(cudaMemcpyAsync(buffers[inputIndex], input, batchSize * INPUT_C * INPUT_H * INPUT_W * sizeof(float), cudaMemcpyHostToDevice, stream));
context.enqueue(batchSize, buffers, stream, nullptr);
CHECK(cudaMemcpyAsync(output, buffers[outputIndex], batchSize * OUTPUT_SIZE * sizeof(float), cudaMemcpyDeviceToHost, stream));
cudaStreamSynchronize(stream);
// release the stream and the buffers
cudaStreamDestroy(stream);
CHECK(cudaFree(buffers[inputIndex]));
CHECK(cudaFree(buffers[outputIndex]));
}
void PostProcessing(float* out, cv::Mat& image_clone, cv::Mat& IM, cv::Size& rawShape)
{
uchar colors[2][3] = { {0,0,0},{128,0,0} };
constexpr int single_len = INPUT_W * INPUT_H;
cv::Mat mask_mat = cv::Mat::zeros(INPUT_W, INPUT_H, CV_8UC3);
float src[2] = { 0 };
uchar color_idx = 0;
for (size_t i = 0; i < INPUT_H; i++) {
uchar* mask_ptr = mask_mat.ptr<uchar>(i);
for (size_t j = 0; j < INPUT_W; j++) {
color_idx = max_idx(out[i * INPUT_W + j], out[single_len + i * INPUT_W + j]);
*mask_ptr++ = colors[color_idx][2];
*mask_ptr++ = colors[color_idx][1];
*mask_ptr++ = colors[color_idx][0];
}
}
//cv::imwrite("../mask_mat.png", mask_mat);
//cv::warpAffine(mask_mat, mask_mat, IM, rawShape);
//cv::addWeighted(image_clone, 0.6, mask_mat, 0.4, 0, image_clone);
//cv::imwrite("../image_clone.png", image_clone);
}
int main(int argc, char** argv)
{
IHostMemory* trtModelStream{ nullptr };
TensorRT* ptensor_rt;
IExecutionContext* context = nullptr;
IRuntime* runtime = nullptr;
ICudaEngine* engine = nullptr;
if (_access(trtModelName.c_str(), 0) != 1)
{
ptensor_rt = LoadNet(trtModelName.c_str());
context = ptensor_rt->context;
runtime = ptensor_rt->runtime;
engine = ptensor_rt->engine;
}
else
{
if (!onnxToTRTModel(onnxModeName, 1, trtModelStream))
return 1;
assert(trtModelStream != nullptr);
std::cout << "Successfully parsed ONNX file!!!!" << std::endl;
// deserialize the engine
runtime = createInferRuntime(gLogger);
assert(runtime != nullptr);
if (gArgs.useDLACore >= 0)
{
runtime->setDLACore(gArgs.useDLACore);
}
engine = runtime->deserializeCudaEngine(trtModelStream->data(), trtModelStream->size(), nullptr);
assert(engine != nullptr);
trtModelStream->destroy();
context = engine->createExecutionContext();
assert(context != nullptr);
}
// 输入预处理
std::cout << "Start reading the input image!!!!" << std::endl;
cv::Mat image = cv::imread(file_name, cv::IMREAD_COLOR);
cv::cvtColor(image, image, cv::COLOR_BGR2RGB);
cv::Mat image_clone = image.clone();
cv::Size rawShape = image.size();
// 图像转成blob
cv::Mat outImage, IM;
image_to_center(image, outImage, IM, cv::Scalar(128, 128, 128));
float* blob = new float[INPUT_SIZE] { 0 };
normal_image2blob(blob, outImage);
float* out = new float[OUTPUT_SIZE] { 0 };
// 推理计时
typedef std::chrono::high_resolution_clock Time;
typedef std::chrono::duration<double, std::ratio<1, 1000>> ms;
typedef std::chrono::duration<float> fsec;
double total = 0.0;
auto t0 = Time::now();
doInference(*context, blob, out, 1);
auto t1 = Time::now();
fsec fs = t1 - t0;
ms d = std::chrono::duration_cast<ms>(fs);
total += d.count();
// 网络输出的后处理
PostProcessing(out, image_clone,IM, rawShape);
// 释放缓存
context->destroy();
engine->destroy();
runtime->destroy();
if (blob){
delete[] blob;
}
if (out){
delete[] out;
}
std::cout << std::endl << "Running time of one image is:" << total << "ms" << std::endl;
return 0;
}
编译添加预处理：_CRT_SECURE_NO_WARNINGS
文章出处登录后可见！
已经登录？立即刷新
