赞
踩
本文记录笔者在 nvidia TX2 系统上搭建 yolov5 环境的过程。
注意说明的是,本文在文后的文章基础上进行实践,根据自己的经历进行描述和补充。由于能力有限,对本文涉及的知识和相关问题无法回答。
本文不涉及 yolo 深度学习方面的内容。
本节对本文进行技术小结。
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
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:
chmod +x Archiconda3-0.2.3-Linux-aarch64.sh
./Archiconda3-0.2.3-Linux-aarch64.sh
source ~/tx/.bashrc
使用如下命令创建名为 yolov5 的 conda 环境:
$ 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。
进入 conda 环境后,命令行前有环境名称 yolov5,如:
(yolov5) tx@tx-desktop:/opt/nvidia/deepstream/yolo$
切换分支:
cd /opt/nvidia/deepstream/yolo/yolov5
git checkout -b v4.0 v4.0
安装:
pip install scikit-build
pip install -r requirements.txt
执行后命令,提示:
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:
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。
到 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:
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:
ModuleNotFoundError: No module named 'torch'
安装之:
pip install torch
torchvision编译需要等待一段时间。
依赖包安装完成后提示:
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:
cd /opt/nvidia/deepstream/yolo/yolov5
下载模型文件:
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。
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:
python gen_wts.py
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。
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。
#define NET s // s m l x
编译:
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:
sudo ./yolov5 -s
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:
$sudo ./yolov5 -s
Loading weights: ../yolov5s.wts
Building engine, please wait for a while...
Build engine successfully!
测试:
mkdir ../samples
从网上下载两张 coco 数据集图片,放到 samples 上。
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。
cp yolov5x.engine ../Deepstream\ 5.0/
cp libmyplugins.so ../Deepstream\ 5.0/
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:
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(本文使用的是 /opt/nvidia/deepstream/deepstream-5.0/sources/objectDetector_Yolo/labels.txt
)。修改 deepstream_app_config_yoloV5.txt 文件的视频地址(本文修改为 uri=file:/opt/nvidia/deepstream/deepstream-5.0/samples/streams/sample_720p.mp4
)。使用官方 deepstream-app 程序测试,执行:
LD_PRELOAD=./libmyplugins.so deepstream-app -c deepstream_app_config_yoloV5.txt
可得到识别的结果。
DeepStream5.0系列之yolov5使用: https://blog.csdn.net/zong596568821xp/article/details/109444343
pytorch源码:https://github.com/pytorch/vision/releases
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
找不到 torchvision 模块提示如下:
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:
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:
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:
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:
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。
使用 yolov5s.wts 在 AGX Xavier 平台上出现的错误:
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):
$ 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:
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。
$ 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
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。
Dockerfile:
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:
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。
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:
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
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
Copyright © 2003-2013 www.wpsshop.cn 版权所有,并保留所有权利。