准备环境

安装jetPack组件

Jetpack 是 Nvidia为 Jetson系列开发板开发的一款软件开发包，常用的开发工具基本都包括了，并在在安装 Jetpack的时候，会自动的将匹配版本的CUDA、cuDNN、TensorRT等。官方提供套件中默认已经安装，可以通过以下命令查看jetPack是否已经安装。

官网指导链接：How to Install JetPack :: NVIDIA JetPack Documentation

# 法一：需要已安装jtop
sudo jetson_release
# 法二：无需安装jtop 
sudo apt-cache show nvidia-jetack

配置cuda环境变量

# 配置cuda环境变量
sudo vim ~/.bashrc
# 在文本中末尾添加如下代码
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda/lib64
export PATH=$PATH:/usr/local/cuda/bin
export CUDA_HOME=$CUDA_HOME:/usr/local/cuda
​
#使该配置生效
sorce ~/.bashrc
​
# 查看cuda 版本
nvcc -V

配置cuDNN

虽然jetpack安装了cuDNN,但并没有将对应的头文件、库文件放到cuda目录，cuDNN的头文件在：/usr/include，库文件位于：/usr/lib/aarch64-linux-gnu。将头文件与库文件复制到cuda目录下

# 复制文件到cuda目录下
cd /usr/include && sudo cp cudnn* /usr/local/cuda/include
cd /usr/lib/aarch64-linux-gnu && sudo cp libcudnn* /usr/local/cuda/lib64
​
# 修改文件权限，修改复制完的头文件与库文件的权限，所有用户都可读，可写，可执行：
sudo chmod 777 /usr/local/cuda/include/cudnn.h 
sudo chmod 777 /usr/local/cuda/lib64/libcudnn*
​
# 重新软链接，这里的8.6.0和8对应安装的cudnn版本号和首数字
cd /usr/local/cuda/lib64
​
sudo ln -sf libcudnn.so.8.6.0 libcudnn.so.8
​
sudo ln -sf libcudnn_ops_train.so.8.6.0 libcudnn_ops_train.so.8
sudo ln -sf libcudnn_ops_infer.so.8.6.0 libcudnn_ops_infer.so.8
​
sudo ln -sf libcudnn_adv_train.so.8.6.0 libcudnn_adv_train.so.8
sudo ln -sf libcudnn_adv_infer.so.8.6.0 libcudnn_adv_infer.so.8
​
sudo ln -sf libcudnn_cnn_train.so.8.6.0 libcudnn_cnn_train.so.8
sudo ln -sf libcudnn_cnn_infer.so.8.6.0 libcudnn_cnn_infer.so.8
​
sudo ldconfig
​
#测试cudnn
sudo cp -r /usr/src/cudnn_samples_v8/ ~/
cd ~/cudnn_samples_v8/mnistCUDNN
sudo chmod 777 ~/cudnn_samples_v8
sudo make clean && sudo make
./mnistCUDNN
​

可能存在问题

1. 编译时freeImage.h文件不存在

# 执行安装命令
sudo apt-get install libfreeimage3 libfreeimage-dev

查看tensorrt版本

import tensorrt
print(tensort.__version__)

部署步骤

1. 拉取代码

#拉取 yolo-tensorRt
git clone https://github.com/Monday-Leo/YOLOv7_Tensorrt
# 拉取yolo 代码
git clone https://github.com/WongKinYiu/yolov7
#将**YOLOv7_Tensorrt**下的**EfficientNMS.py**和**export_onnx.py**复制到**yolov7**下，导出含有#EfficientNMS的onnx模型。
python3 export_onnx.py --weights ./weights/yolov7.pt
​

注：

如果遇到:EfficientNMS_TRT type is missing, so it may result in wrong shape inference for the exported graph.异常时，需要安装以下依赖

pip install onnx-simplifier pip install coremltools pip install nvidia-pyindex pip install onnx-graphsurgeon

2. 生成engine模型文件

trtexec --onnx=./yolov7.onnx --saveEngine=./yolov7_fp16.engine --fp16 --workspace=200

3. 运行代码

在YOLOv7_Tensorrt 工程项目中新增trt_model文件夹，并将转换好的yolov7_f16.engine 拷贝到 文件下，修改infer.py代码段

import cv2
import tensorrt as trt
import torch
import numpy as np
from collections import OrderedDict,namedtuple

class TRT_engine():
    def __init__(self, weight) -> None:
        self.imgsz = [640,640]
        self.weight = weight
        self.device = torch.device('cuda:0')
        self.init_engine()

    def init_engine(self):
        # Infer TensorRT Engine
        self.Binding = namedtuple('Binding', ('name', 'dtype', 'shape', 'data', 'ptr'))
        self.logger = trt.Logger(trt.Logger.INFO)
        trt.init_libnvinfer_plugins(self.logger, namespace="")
        with open(self.weight, 'rb') as self.f, trt.Runtime(self.logger) as self.runtime:
            self.model = self.runtime.deserialize_cuda_engine(self.f.read())
        self.bindings = OrderedDict()
        self.fp16 = False
        for index in range(self.model.num_bindings):
            self.name = self.model.get_binding_name(index)
            self.dtype = trt.nptype(self.model.get_binding_dtype(index))
            self.shape = tuple(self.model.get_binding_shape(index))
            self.data = torch.from_numpy(np.empty(self.shape, dtype=np.dtype(self.dtype))).to(self.device)
            self.bindings[self.name] = self.Binding(self.name, self.dtype, self.shape, self.data, int(self.data.data_ptr()))
            if self.model.binding_is_input(index) and self.dtype == np.float16:
                self.fp16 = True
        self.binding_addrs = OrderedDict((n, d.ptr) for n, d in self.bindings.items())
        self.context = self.model.create_execution_context()

    def letterbox(self,im,color=(114, 114, 114), auto=False, scaleup=True, stride=32):
        # Resize and pad image while meeting stride-multiple constraints
        shape = im.shape[:2]  # current shape [height, width]
        new_shape = self.imgsz
        if isinstance(new_shape, int):
            new_shape = (new_shape, new_shape)
        # Scale ratio (new / old)
        self.r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])
        if not scaleup:  # only scale down, do not scale up (for better val mAP)
            self.r = min(self.r, 1.0)
        # Compute padding
        new_unpad = int(round(shape[1] * self.r)), int(round(shape[0] * self.r))
        self.dw, self.dh = new_shape[1] - new_unpad[0], new_shape[0] - new_unpad[1]  # wh padding
        if auto:  # minimum rectangle
            self.dw, self.dh = np.mod(self.dw, stride), np.mod(self.dh, stride)  # wh padding
        self.dw /= 2  # divide padding into 2 sides
        self.dh /= 2
        if shape[::-1] != new_unpad:  # resize
            im = cv2.resize(im, new_unpad, interpolation=cv2.INTER_LINEAR)
        top, bottom = int(round(self.dh - 0.1)), int(round(self.dh + 0.1))
        left, right = int(round(self.dw - 0.1)), int(round(self.dw + 0.1))
        self.img = cv2.copyMakeBorder(im, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color)  # add border
        return self.img,self.r,self.dw,self.dh

    def preprocess(self,image):
        self.img,self.r,self.dw,self.dh = self.letterbox(image)
        self.img = self.img.transpose((2, 0, 1))
        self.img = np.expand_dims(self.img,0)
        self.img = np.ascontiguousarray(self.img)
        self.img = torch.from_numpy(self.img).to(self.device)
        self.img = self.img.float()
        return self.img

    def predict(self,img,threshold):
        img = self.preprocess(img)
        self.binding_addrs['images'] = int(img.data_ptr())
        self.context.execute_v2(list(self.binding_addrs.values()))
        nums = self.bindings['num_dets'].data[0].tolist()
        boxes = self.bindings['det_boxes'].data[0].tolist()
        scores =self.bindings['det_scores'].data[0].tolist()
        classes = self.bindings['det_classes'].data[0].tolist()
        num = int(nums[0])
        new_bboxes = []
        for i in range(num):
            if(scores[i] < threshold):
                continue
            xmin = (boxes[i][0] - self.dw)/self.r
            ymin = (boxes[i][1] - self.dh)/self.r
            xmax = (boxes[i][2] - self.dw)/self.r
            ymax = (boxes[i][3] - self.dh)/self.r
            new_bboxes.append([classes[i],scores[i],xmin,ymin,xmax,ymax])
        return new_bboxes

def visualize(img,bbox_array):
    for temp in bbox_array:
        xmin = int(temp[2])
        ymin = int(temp[3])
        xmax = int(temp[4])
        ymax = int(temp[5])
        clas = int(temp[0])
        score = temp[1]
        cv2.rectangle(img,(xmin,ymin),(xmax,ymax), (105, 237, 249), 2)
        img = cv2.putText(img, "class:"+str(clas)+" "+str(round(score,2)), (xmin,int(ymin)-5), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (105, 237, 249), 1)
    return img

def test(path):
    import os
    out = './runs'
    images = os.listdir(path)
    engine = TRT_engine("./trt_model/video_fp32.engine")
    for item in images:
        img = cv2.imread(f'{path}/{item}')
        results = engine.predict(img,threshold=0.25)
        img = visualize(img,results)
        cv2.imwrite(f'{out}/{item}',img)

test('./pictures')