tesnortt_c++_c++ tensorrt推理比python慢很多

tensorrt_c++_api 加速推理

上一篇写了tensorrt python 加速，这一篇是关于c++版本的加速，但是由于在预处理没有找到c++ PIL的实现，在精度上复现不了python版本，并且速度也没有python的快，但是将模型做成服务的话，由于python的并发没有c++版本的快，所以在部署时还是使用c++更合适

#pragma once
#include <iostream>
#include <fstream>
#include <sstream>
#include "opencv2/opencv.hpp"
#include "cuda_runtime.h"
#include "torch/script.h"
#include "torch/torch.h"
#include "torch/cuda.h"
#include "NvInfer.h"
#include "NvOnnxParser.h"
#include "filesystem"
#include <ctime>


#define INPUT_CHANNEL 3
#define IMAGE_WIDTH 224
#define IMAGE_HEIGHT 224

// 实例化记录器，用来捕捉警告信息，并且忽略信息留言
class Logger : public nvinfer1::ILogger{
    void log(nvinfer1::ILogger::Severity severity, const char* msg) noexcept override{
        //忽略以下级别信息
        if (severity <= nvinfer1::ILogger::Severity::kWARNING){
            std::cout << msg << std::endl;
        }
    }

} logger;

//遍历文件夹图片
void getImageFiles(std::vector<std::string>& fileLists,std::string testPath) {
    std::filesystem::path path(testPath);
    assert(std::filesystem::exists(path));
    std::filesystem::directory_iterator files(path);
    for (auto& file : files) {
        if (!cv::imread(file.path()).empty())
            fileLists.push_back(file.path());
    }
}

void prepareImage(cv::Mat &vec_img, float* inputData) {
    // std::vector<float> img_mean{0.485, 0.456, 0.406};
    // std::vector<float> img_std{0.229, 0.224, 0.225};
    std::vector<float> result;

    if (!vec_img.data)
        std::cout << "error" <<std::endl;
    cv::Mat rsz_img, flt_img;
    cv::cvtColor(vec_img,rsz_img,cv::COLOR_BGR2RGB);
    cv::resize(vec_img, rsz_img, cv::Size(IMAGE_WIDTH, IMAGE_HEIGHT));

    for(int i = 0; i < 224; ++i){
        for(int j = 0; j < 224; ++j){
            std::cout << rsz_img.at<cv::Vec3b>(i,j) << std::endl;
        }
    }


    torch::Tensor img_tensor = torch::from_blob(rsz_img.data, { rsz_img.rows, rsz_img.cols, 3 }, torch::kByte);
    img_tensor = img_tensor.permute({ 2, 0, 1 });
    img_tensor = img_tensor.to(torch::kF32);
    img_tensor = img_tensor.div(255);
    img_tensor = img_tensor.unsqueeze(0);
    img_tensor[0][0] = img_tensor[0][0].sub_(0.5).div_(0.5);
    img_tensor[0][1] = img_tensor[0][1].sub_(0.5).div_(0.5);
    img_tensor[0][2] = img_tensor[0][2].sub_(0.5).div_(0.5);


    auto imgTensor = img_tensor.accessor<float, 4>();
    // 将vector换成数组
    for(int channel = 0; channel < 3;++channel){
        for(int left = 0; left < 224; ++left){
            for(int right = 0; right < 224; ++right){
                // result.emplace_back(imgTensor[0][channel][left][right]);
                inputData[channel*224*224 + left*224 + right] = imgTensor[0][channel][left][right];
                // std::cout << imgTensor[0][channel][left][right] << std::endl;
            }   
        }
    }
}

int64_t volume(const nvinfer1::Dims& d)
{
    return std::accumulate(d.d, d.d + d.nbDims, 1, std::multiplies<int64_t>());
}

unsigned int getElementSize(nvinfer1::DataType t)
{
    switch (t)
    {
        case nvinfer1::DataType::kINT32: return 4;
        case nvinfer1::DataType::kFLOAT: return 4;
        case nvinfer1::DataType::kHALF: return 2;
        case nvinfer1::DataType::kBOOL:
        case nvinfer1::DataType::kINT8: return 1;
    }
    throw std::runtime_error("Invalid DataType.");
    return 0;
}

int returnMax(float a[]){
    int length = 10;
    float temp;
    int flag = 0;
    for(int i = 1; i < length;++i){
        if(temp < a[i]){
            temp = a[i];
            flag = i;
        }
    }
    return flag;
}

int getIndex(std::vector<std::string> class_, std::string str){
    std::vector<std::string>::iterator begin = class_.begin();
    for(int i = 0; i < class_.size(); ++i){
        if (class_[i] == str){
            return i;
        }
    }
}


int getResult(std::string image_path, nvinfer1::ICudaEngine* engine, nvinfer1::IExecutionContext* context){
    // 读取图片图片
    cv::Mat image = cv::imread(image_path);
    assert(!image.empty());
    // 图片预处理
    float a[3*224*224];
    prepareImage(image, a);


    // 预测图片
    void *buffers[2];
    std::vector<int64_t> bufferSize;
    int nbindings = engine->getNbBindings();
    bufferSize.resize(nbindings);

    for(int i = 0;i < nbindings; ++i){
        nvinfer1::Dims dims = engine->getBindingDimensions(i);
        nvinfer1::DataType dtype = engine->getBindingDataType(i);
        int64_t totalSize = volume(dims) * 1 * getElementSize(dtype);
        // std::cout << i << " : " << totalSize << std::endl;
        bufferSize[i] = totalSize;
        cudaMalloc(&buffers[i], totalSize);
    }


    cudaStream_t stream;
    cudaStreamCreate(&stream);

    int outSize = bufferSize[1] / sizeof(float);

    cudaMemcpyAsync(buffers[0],&a, bufferSize[0],cudaMemcpyHostToDevice,stream);
    //注意：下面方法线程不安全
    context->execute(1, buffers);

    float out[outSize];
    cudaMemcpyAsync(out, buffers[1], bufferSize[1], cudaMemcpyDeviceToHost, stream);
    cudaStreamSynchronize(stream);
    cudaFree(buffers[0]);
    cudaFree(buffers[1]);
    cudaStreamDestroy(stream);
    return returnMax(out);
}

int main(){
    // 模型路径
    std::string model_path = "/data/kile/other/Inception/mobile_net/onnx_/mobilev2_onnx2.trt";
    // 定义文件流
    std::ifstream inFile(model_path, std::ios_base::in|std::ios_base::binary);
    std::string cached_engine = "";
    while(inFile.peek() != EOF){
        std::stringstream buffer;
        buffer << inFile.rdbuf();
        cached_engine.append(buffer.str());
    }
    inFile.close();
    // 反序列化模型
    nvinfer1::IRuntime* runtime = nvinfer1::createInferRuntime(logger);
    // 从内存中加载模型获得引擎
    nvinfer1::ICudaEngine* engine = runtime->deserializeCudaEngine(cached_engine.data(), cached_engine.size(), nullptr);

    // 开始推理
    // 创建推理上下文
    nvinfer1::IExecutionContext* context = engine->createExecutionContext();
    // 图片路径
    // std::string image_path = "/data/kile/other/Inception/mobile_net/dataset/test_one/airplane/airplane_3.jpg";
    // class类别
    std::vector<std::string> class_ = {"airplane", "automobile", "bird", "cat", "deer", "dog", "frog", "horse", "ship", "truck"};
    int correct = 0;
    int total = 0;
    // 计时
    clock_t start = clock();
    
    
    for (auto str:class_){
        int classId = getIndex(class_, str);
        std::cout << classId << std::endl;
        // 遍历图片
        std::vector<std::string> fileLists;
        // getImageFiles(fileLists, "/data/kile/other/Inception/mobile_net/dataset/test_data/"+str);
        getImageFiles(fileLists, "/data/kile/other/Inception/mobile_net/dataset/test_one/bird");
        for(auto filePath:fileLists){
            int predictClassId = getResult(filePath, engine, context);
            total += 1;
            std::cout << str << " " << class_[predictClassId] << " " << correct << std::endl;
            if(str == class_[predictClassId]){
                correct += 1;
            }
        }
    }
    clock_t end = clock();
    std::cout << "correct ：" << (float)correct / total << "time" << (float)(end-start)  << std::endl; 
    return 0;    
}


1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219