本文记录笔者在 nvidia TX2 系统上搭建 yolov5 环境的过程。
一、要点 本节对本文进行技术小结。
TX2 刷机后，PYthon版本为 3.7.1。
保持T X2 联网，因为需要下载。
conda 创建环境，命令行前会有环境名称的提示。
文后的尝试及记录，建议看看。
二、过程 准备工程源码 1 2 3 4 5 mkdir -p /opt/nvidia/deepstream/yolo cd /opt/nvidia/deepstream/yolo git clone https://github.com/DanaHan/Yolov5-in-Deepstream-5.0.git git clone https://github.com/wang-xinyu/tensorrtx.git git clone https://github.com/ultralytics/yolov5.git
安装 conda 如无 conda，则安装之。到 https://github.com/Archiconda/build-tools/releases 下载。本文所用安装脚本为 Archiconda3-0.2.3-Linux-aarch64.sh 。下载到系统，再安装：
1 2 3 4 5 chmod +x Archiconda3-0.2.3-Linux-aarch64.sh ./Archiconda3-0.2.3-Linux-aarch64.sh source ~/tx/.bashrc
创建 conda 环境 使用如下命令创建名为 yolov5 的 conda 环境：
1 $ conda create -n yolov5 python
此过程需要手动输入 y 以安装软件包。安装时长取决于联网速度。
成功后，输入conda activate yolov5激活，使用conda deactivate退出。过程日志信息详见文后附录。
1 (yolov5) tx@tx-desktop:/opt/nvidia/deepstream/yolo$
安装 yolo 依赖库 切换分支：
1 2 cd /opt/nvidia/deepstream/yolo/yolov5 git checkout -b v4.0 v4.0
安装：
1 2 3 pip install scikit-build pip install -r requirements.txt
执行后命令，提示：
1 2 ERROR: Could not find a version that satisfies the requirement torchvision>=0.8.1 ERROR: No matching distribution found for torchvision>=0.8.1
安装 torchvision：
1 conda install torchvision -c pytorch
在实践中发现该命令无法从官方途径安装。故从源码安装。https://github.com/pytorch/vision/releases 下载源码压缩包，本文所用版本 v0.9.0，下载，并安装：
1 2 3 tar xf vision-0.9.0.tar.gz cd vision-0.9.0 python setup.py install
安装过程可能提示没有 torch 模块：
1 ModuleNotFoundError: No module named 'torch'
安装之：
torchvision编译需要等待一段时间。
依赖包安装完成后提示：
1 Successfully installed Cython-0.29.22 PyYAML-5.4.1 absl-py-0.12.0 cachetools-4.2.1 chardet-4.0.0 cycler-0.10.0 google-auth-1.27.1 google-auth-oauthlib-0.4.3 grpcio-1.36.1 idna-2.10 kiwisolver-1.3.1 markdown-3.3.4 matplotlib-3.3.4 oauthlib-3.1.0 opencv-python-4.5.1.48 pandas-1.2.3 protobuf-3.15.6 pyasn1-0.4.8 pyasn1-modules-0.2.8 pycocotools-2.0.2 python-dateutil-2.8.1 pytz-2021.1 requests-2.25.1 requests-oauthlib-1.3.0 rsa-4.7.2 scipy-1.6.1 seaborn-0.11.1 six-1.15.0 tensorboard-2.4.1 tensorboard-plugin-wit-1.8.0 thop-0.0.31.post2005241907 tqdm-4.59.0 urllib3-1.26.3 werkzeug-1.0.1
在yolov5目录下：
1 cd /opt/nvidia/deepstream/yolo/yolov5
下载模型文件：
1 bash weights/download_weights.sh
默认下载到yolov5目录，有文件：yolov5l.pt yolov5m.pt yolov5s.pt yolov5x.pt。可根据实际情况选，本文使用默认值，即yolov5s.pt文件。拷贝到 weights 目录。
继续在 yolov5 目录下，拷贝 gen_wts.py。
1 cp ../tensorrtx/yolov5/gen_wts.py ./
注意，可修改 gen_wts.py 文件，主要修改'weights/yolov5x.pt' 'yolov5x.wts'名称。
生成wts文件：
最终生成文件：yolov5x.wts。将其拷贝到 Yolov5-in-Deepstream-5.0 目录，并切换到该目录。
1 2 cp yolov5s.wts ../Yolov5-in-Deepstream-5.0 cd ../Yolov5-in-Deepstream-5.0
根据实际情况，修改 yolov5.cpp 文件，将 NET 宏改成自己对应的模型。本文使用默认值 s，无须改动。
1 #define NET s // s m l x
编译：
1 2 3 4 mkdir build cd build cmake .. make
生成 yolov5 文件，后面将用该文件生成模型和测试。
生成engine文件：
注意：程序使用了../yolov5x.wts，所以 yolov5x.wts 要拷贝 Yolov5-in-Deepstream-5.0 目录。成功输出如下信息：
1 2 3 4 $sudo ./yolov5 -s Loading weights: ../yolov5s.wts Building engine, please wait for a while... Build engine successfully!
测试：
从网上下载两张 coco 数据集图片，放到 samples 上。
1 sudo ./yolov5 -d ../samples
在build可看到有图片文件生成，且标注了，说明成功。
将 libmyplugins.so yolov5s.engine 拷贝到 Deepstream 5.0 目录。
1 2 cp yolov5x.engine ../Deepstream\ 5.0/ cp libmyplugins.so ../Deepstream\ 5.0/
Deepstream 测试 在 Yolov5-in-Deepstream-5.0\Deepstream 5.0\nvdsinfer_custom_impl_Yolo 目录执行make，生成libnvdsinfer_custom_impl_Yolo.so文件。进入 Yolov5-in-Deepstream-5.0\Deepstream 5.0 目录中，修改 config_infer_primary_yoloV5.tx 文件。主要修改：
1 2 3 model-engine-file=yolov5s.engine --> model-engine-file=yolov5x.engine custom-lib-path=objectDetector_Yolo_V5/nvdsinfer_custom_impl_Yolo/libnvdsinfer_custom_impl_Yolo.so --> custom-lib-path=nvdsinfer_custom_impl_Yolo/libnvdsinfer_custom_impl_Yolo.so
将 DeepStream 的 label.txt 拷贝到当前目录。修改 deepstream_app_config_yoloV5.txt 文件的视频地址。使用官方 deepstream-app 程序测试，执行：
1 LD_PRELOAD=./libmyplugins.so deepstream-app -c deepstream_app_config_yoloV5.txt
可得到识别的结果。
三、参考资源 DeepStream5.0系列之yolov5使用： https://blog.csdn.net/zong596568821xp/article/details/109444343 https://github.com/pytorch/vision/releases
附1：尝试 一般错误 1 2 3 4 5 6 7 ModuleNotFoundError: No module named 'Cython' --> python3.7 -m pip install cython Could not find a version that satisfies the requirement cmake --> python3.7 -m pip install cmake
torchvision 找不到 torchvision 模块提示如下：
1 2 ERROR: Could not find a version that satisfies the requirement torchvision>=0.8.1 ERROR: No matching distribution found for torchvision>=0.8.1
选按通用安装方式：
1 python3.7 -m pip install torchvision
但错误提示依旧。按网上说法使用如下命令安装：
1 pip3 install torchvision==0.8.1 -f https://download.pytorch.org/whl/torch_stable.html
但依然失败。再用官方 提供的命令：
1 conda install torchvision -c pytorch
还是失败。究其原因，可能是没有 aarch64 版本的缘故。所以用源码安装。
生成 egine 文件时提示：
1 2 3 4 5 6 Loading weights: ../yolov5x.wts [03/10/2021-01:00:59] [E] [TRT] (Unnamed Layer* 14) [Convolution]: kernel weights has count 0 but 12800 was expected [03/10/2021-01:00:59] [E] [TRT] (Unnamed Layer* 14) [Convolution]: count of 0 weights in kernel, but kernel dimensions (1,1) with 160 input channels, 80 output channels and 1 groups were specified. Expected Weights count is 160 * 1*1 * 80 / 1 = 12800 [03/10/2021-01:00:59] [E] [TRT] Parameter check failed at: ../builder/Network.cpp::addScale::482, condition: shift.count > 0 ? (shift.values != nullptr) : (shift.values == nullptr) yolov5: /opt/nvidia/deepstream/yolo-in-Deepstream-5.0/common.hpp:190: nvinfer1::IScaleLayer* addBatchNorm2d(nvinfer1::INetworkDefinition*, std::map<std::__cxx11::basic_string<char>, nvinfer1::Weights>&, nvinfer1::ITensor&, std::__cxx11::string, float): Assertion `scale_1' failed. Aborted
原因是使用 yolov5x 版本，改为默认的 yolov5s 不再出现。
1 2 3 4 5 6 7 sudo ./yolov5 -s Loading weights: ../yolov5s.wts [03/12/2021-16:06:07] [E] [TRT] (Unnamed Layer* 14) [Convolution]: kernel weights has count 0 but 2048 was expected [03/12/2021-16:06:07] [E] [TRT] (Unnamed Layer* 14) [Convolution]: count of 0 weights in kernel, but kernel dimensions (1,1) with 64 input channels, 32 output channels and 1 groups were specified. Expected Weights count is 64 * 1*1 * 32 / 1 = 2048 [03/12/2021-16:06:07] [E] [TRT] Parameter check failed at: ../builder/Network.cpp::addScale::482, condition: shift.count > 0 ? (shift.values != nullptr) : (shift.values == nullptr) yolov5: /opt/nvidia/deepstream/yolo/Yolov5-in-Deepstream-5.0/common.hpp:190: nvinfer1::IScaleLayer* addBatchNorm2d(nvinfer1::INetworkDefinition*, std::map<std::__cxx11::basic_string<char>, nvinfer1::Weights>&, nvinfer1::ITensor&, std::__cxx11::string, float): Assertion `scale_1' failed. Aborted
使用不同平台的 engine 文件的错误(TX2生成，AGX Xavier运行）：
1 2 3 4 5 6 7 $ sudo ./yolov5 -d ../samples [03/12/2021-16:12:13] [E] [TRT] INVALID_CONFIG: The engine plan file is generated on an incompatible device, expecting compute 7.2 got compute 6.2, please rebuild. [03/12/2021-16:12:13] [E] [TRT] engine.cpp (1546) - Serialization Error in deserialize: 0 (Core engine deserialization failure) [03/12/2021-16:12:13] [E] [TRT] INVALID_STATE: std::exception [03/12/2021-16:12:13] [E] [TRT] INVALID_CONFIG: Deserialize the cuda engine failed. yolov5: /opt/nvidia/deepstream/yolo/Yolov5-in-Deepstream-5.0/yolov5.cpp:534: int main(int, char**): Assertion `engine != nullptr' failed. Aborted
测试时，如果 samples 目录存在非图片时，会提示：
1 2 3 740ms terminate called after throwing an instance of 'cv::Exception' what(): OpenCV(4.1.1) /home/nvidia/host/build_opencv/nv_opencv/modules/imgcodecs/src/loadsave.cpp:662: error: (-2:Unspecified error) could not find a writer for the specified extension in function 'imwrite_'
解决：图片后缀需正确，如jpg、png。
创建 yolo 环境输出日志 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 $ conda create -n yolov5 python Solving environment: done ## Package Plan ## environment location: /home/tx/archiconda3/envs/yolov5 added / updated specs: - python The following packages will be downloaded: package | build ---------------------------|----------------- zlib-1.2.11 | h7b6447c_2 118 KB c4aarch64 openssl-1.0.2p | h7b6447c_0 3.1 MB c4aarch64 sqlite-3.25.2 | h7ce4240_0 2.2 MB c4aarch64 readline-7.0 | h7ce4240_5 440 KB c4aarch64 pip-10.0.1 | py37_0 1.7 MB c4aarch64 wheel-0.32.1 | py37_0 34 KB c4aarch64 libgcc-ng-7.3.0 | h5c90dd9_0 5.9 MB c4aarch64 libstdcxx-ng-7.3.0 | h5c90dd9_0 2.5 MB c4aarch64 setuptools-40.4.3 | py37_0 601 KB c4aarch64 xz-5.2.4 | h7ce4240_4 345 KB c4aarch64 certifi-2018.10.15 | py37_0 137 KB c4aarch64 ncurses-6.1 | h71b71f5_0 1.0 MB c4aarch64 tk-8.6.8 | hbc83047_0 3.2 MB c4aarch64 ca-certificates-2018.03.07 | 0 123 KB c4aarch64 python-3.7.2 | he90a169_0 36.1 MB c4aarch64 libffi-3.2.1 | h71b71f5_5 51 KB c4aarch64 libedit-3.1.20170329 | hc058e9b_2 188 KB c4aarch64 ------------------------------------------------------------ Total: 57.8 MB The following NEW packages will be INSTALLED: ca-certificates: 2018.03.07-0 c4aarch64 certifi: 2018.10.15-py37_0 c4aarch64 libedit: 3.1.20170329-hc058e9b_2 c4aarch64 libffi: 3.2.1-h71b71f5_5 c4aarch64 libgcc-ng: 7.3.0-h5c90dd9_0 c4aarch64 libstdcxx-ng: 7.3.0-h5c90dd9_0 c4aarch64 ncurses: 6.1-h71b71f5_0 c4aarch64 openssl: 1.0.2p-h7b6447c_0 c4aarch64 pip: 10.0.1-py37_0 c4aarch64 python: 3.7.2-he90a169_0 c4aarch64 readline: 7.0-h7ce4240_5 c4aarch64 setuptools: 40.4.3-py37_0 c4aarch64 sqlite: 3.25.2-h7ce4240_0 c4aarch64 tk: 8.6.8-hbc83047_0 c4aarch64 wheel: 0.32.1-py37_0 c4aarch64 xz: 5.2.4-h7ce4240_4 c4aarch64 zlib: 1.2.11-h7b6447c_2 c4aarch64 Proceed ([y]/n)? y Preparing transaction: done Verifying transaction: done Executing transaction: done # # To activate this environment, use # # $ conda activate yolov5 # # To deactivate an active environment, use # # $ conda deactivate
附2：成果 1、文中涉及的软件包只在 TX2 测试通过，已整合为 aarch64 系统的 docker 镜像，并上传至阿里云仓库备份。
1 2 3 4 5 FROM busybox RUN mkdir /yolotools COPY vision-0.9.0.tar.gz /yolotools COPY yolo.tar.bz2 /yolotools COPY Archiconda3-0.2.3-Linux-aarch64.sh /yolotools
创建并上传：
1 2 docker build -t registry.cn-hangzhou.aliyuncs.com/latelee/yoloenv:aarch64 . docker push registry.cn-hangzhou.aliyuncs.com/latelee/yoloenv:aarch64
镜像说明：yolo.tar.bz2包含了本文的工程源码。运行后查看容器的 /yolotools 目录。
1 2 docker run -itd --name foobar registry.cn-hangzhou.aliyuncs.com/latelee/yoloenv:aarch64 sh docker exec -it foobar ls /yolotools
在宿主机拷贝文件出来：
1 2 3 4 mkdir tools docker cp foobar:/yolotools/vision-0.9.0.tar.gz tools docker cp foobar:/yolotools/yolo.tar.bz2 tools docker cp foobar:/yolotools/Archiconda3-0.2.3-Linux-aarch64.sh tools
再附 一段时间来，github 和 gitlab 官方经常打不开，在访问网站上较耗时。不知何故。
李迟 2021.3.12 周五
