Skip to main content

YOLOv4 Ubuntu 使用

本文将介绍 YOLOv4 官方 Darknet 实现,如何于 Ubuntu 18.04 编译,及使用 Python 接口。

主要内容有:

  • 准备基础环境: Nvidia Driver, CUDA, cuDNN, CMake, Python
  • 编译应用环境: OpenCV, Darknet
  • 用预训练模型进行推断: darknet 执行,或 python

而 YOLOv4 的介绍或训练,可见《YOLOv4 Docker 使用》。

准备基础环境

Nvidia Driver

推荐使用 graphics drivers PPA 安装 Nvidia 驱动:

sudo add-apt-repository ppa:graphics-drivers/ppa
sudo apt update

查看推荐的 Nvidia 显卡驱动:

ubuntu-drivers devices

安装 Nvidia 驱动:

apt-cache search nvidia | grep ^nvidia-driver
sudo apt install nvidia-driver-450

之后, sudo reboot 重启。运行 nvidia-smi 查看 Nvidia 驱动信息。

Nvidia CUDA Toolkit

获取地址:

 建议选择 CUDA 10.2 ,为目前 PyTorch 可支持的最新版本。

下载安装:

wget http://developer.download.nvidia.com/compute/cuda/10.2/Prod/local_installers/cuda_10.2.89_440.33.01_linux.run
sudo sh cuda_10.2.89_440.33.01_linux.run

注意:安装时,请手动取消驱动安装选项。

安装输出:

===========
= Summary =
===========

Driver: Not Selected
Toolkit: Installed in /usr/local/cuda-10.2/
Samples: Installed in /home/john/cuda-10.2/, but missing recommended libraries

Please make sure that
- PATH includes /usr/local/cuda-10.2/bin
- LD_LIBRARY_PATH includes /usr/local/cuda-10.2/lib64, or, add /usr/local/cuda-10.2/lib64 to /etc/ld.so.conf and run ldconfig as root

To uninstall the CUDA Toolkit, run cuda-uninstaller in /usr/local/cuda-10.2/bin

Please see CUDA_Installation_Guide_Linux.pdf in /usr/local/cuda-10.2/doc/pdf for detailed information on setting up CUDA.
***WARNING: Incomplete installation! This installation did not install the CUDA Driver. A driver of version at least 440.00 is required for CUDA 10.2 functionality to work.
To install the driver using this installer, run the following command, replacing <CudaInstaller> with the name of this run file:
sudo <CudaInstaller>.run --silent --driver

Logfile is /var/log/cuda-installer.log

添加环境变量:

$ vi ~/.bashrc
export CUDA_HOME=/usr/local/cuda
export PATH=$CUDA_HOME/bin:$PATH
export LD_LIBRARY_PATH=$CUDA_HOME/lib64:$LD_LIBRARY_PATH

重启终端后,检查:

$ nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2019 NVIDIA Corporation
Built on Wed_Oct_23_19:24:38_PDT_2019
Cuda compilation tools, release 10.2, V10.2.89

Nvida cuDNN

获取地址:

需选择 CUDA 10.2 对应的版本。

安装 deb 包:

sudo apt install ./libcudnn8_8.0.2.39-1+cuda10.2_amd64.deb
sudo apt install ./libcudnn8-dev_8.0.2.39-1+cuda10.2_amd64.deb
sudo apt install ./libcudnn8-doc_8.0.2.39-1+cuda10.2_amd64.deb

查看 deb 包:

dpkg -c libcudnn8_8.0.2.39-1+cuda10.2_amd64.deb

CMake

下载安装:

curl -O -L https://github.com/Kitware/CMake/releases/download/v3.18.2/cmake-3.18.2-Linux-x86_64.sh
sh cmake-*.sh --prefix=$HOME/Applications/

添加环境变量:

$ vi ~/.bashrc
export PATH=$HOME/Applications/cmake-3.18.2-Linux-x86_64/bin:$PATH

说明: apt 源的 cmake 太旧, darknet 编译不过。

Python

获取地址:

Python 建议用 Anaconda 发行版。

安装命令:

# bash Anaconda3-2020.07-Linux-x86_64.sh
bash Anaconda3-2019.10-Linux-x86_64.sh

编译应用环境

OpenCV 4.4.0

安装依赖:

apt install -y build-essential git libgtk-3-dev

编译命令:

conda deactivate

# export CONDA_HOME="/home/john/anaconda3/envs/clenv"
export CONDA_HOME=`conda info -s | grep -Po "sys.prefix:\s*\K[/\w]*"`

cd ~/Codes/

git clone -b 4.4.0 --depth 1 https://github.com/opencv/opencv.git
git clone -b 4.4.0 --depth 1 https://github.com/opencv/opencv_contrib.git

cd opencv/
mkdir _build && cd _build/

cmake -DCMAKE_BUILD_TYPE=Release \
-DCMAKE_INSTALL_PREFIX=$HOME/opencv-cuda-4.4.0 \
-DOPENCV_EXTRA_MODULES_PATH=$HOME/Codes/opencv_contrib/modules \
\
-DPYTHON_EXECUTABLE=$CONDA_HOME/bin/python3.7 \
-DPYTHON3_EXECUTABLE=$CONDA_HOME/bin/python3.7 \
-DPYTHON3_LIBRARY=$CONDA_HOME/lib/libpython3.7m.so \
-DPYTHON3_INCLUDE_DIR=$CONDA_HOME/include/python3.7m \
-DPYTHON3_NUMPY_INCLUDE_DIRS=$CONDA_HOME/lib/python3.7/site-packages/numpy/core/include \
-DBUILD_opencv_python2=OFF \
-DBUILD_opencv_python3=ON \
\
-DWITH_CUDA=ON \
\
-DBUILD_DOCS=OFF \
-DBUILD_EXAMPLES=OFF \
-DBUILD_TESTS=OFF \
..

make -j$(nproc)
make install

其中 Python 路径请对应自己安装的版本。

运行检查:

conda activate

export LD_LIBRARY_PATH=$HOME/opencv-cuda-4.4.0/lib:$LD_LIBRARY_PATH
export PYTHONPATH=$HOME/opencv-cuda-4.4.0/lib/python3.7/site-packages:$PYTHONPATH

python - <<EOF
import cv2
print(cv2.__version__)
EOF

问题: libfontconfig.so.1

Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/home/john/opencv-cuda-4.4.0/lib/python3.7/site-packages/cv2/__init__.py", line 96, in <module>
bootstrap()
File "/home/john/opencv-cuda-4.4.0/lib/python3.7/site-packages/cv2/__init__.py", line 86, in bootstrap
import cv2
ImportError: /home/john/anaconda3/bin/../lib/libfontconfig.so.1: undefined symbol: FT_Done_MM_Var

解决办法:

cd $HOME/anaconda3/lib/
mv libfontconfig.so.1 libfontconfig.so.1.bak
ln -s /usr/lib/x86_64-linux-gnu/libfontconfig.so.1 libfontconfig.so.1

问题: libpangoft2-1.0.so.0

ImportError: /home/john/anaconda3/bin/../lib/libpangoft2-1.0.so.0: undefined symbol: FcWeightToOpenTypeDouble

解决办法:

cd $HOME/anaconda3/lib/
mv libpangoft2-1.0.so.0 libpangoft2-1.0.so.0.bak
ln -s /usr/lib/x86_64-linux-gnu/libpangoft2-1.0.so.0 libpangoft2-1.0.so.0

Darknet

编译命令:

export OpenCV_DIR=$HOME/opencv-cuda-4.4.0/lib/cmake

cd ~/Codes/

git clone https://github.com/AlexeyAB/darknet.git

cd darknet/
./build.sh

运行检查:

$ export LD_LIBRARY_PATH=$HOME/opencv-cuda-4.4.0/lib:$LD_LIBRARY_PATH

$ ./darknet v
CUDA-version: 10020 (10020), cuDNN: 8.0.2, CUDNN_HALF=1, GPU count: 1
CUDNN_HALF=1
OpenCV version: 4.4.0
Not an option: v

用预训练模型进行推断

准备模型与数据

预训练模型 yolov4.weights ,下载地址 https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v3_optimal/yolov4.weights

可以准备 MS COCO 数据集,下载地址 http://cocodataset.org/#download 。或者自己找个图片。

darknet 执行

cd ~/Codes/darknet/

export LD_LIBRARY_PATH=$HOME/opencv-cuda-4.4.0/lib:$LD_LIBRARY_PATH

export MY_MODEL_DIR=~/Codes/devel/models/yolov4
export MY_COCO_DIR=~/Codes/devel/datasets/coco2017

./darknet detector test cfg/coco.data cfg/yolov4.cfg \
$MY_MODEL_DIR/yolov4.weights \
-thresh 0.25 -ext_output -show \
$MY_COCO_DIR/test2017/000000000001.jpg

推断结果:

python 执行

Darknet 于其根目录,提供有 Python 接口。如下执行:

cd ~/Codes/darknet/

export LD_LIBRARY_PATH=$HOME/opencv-cuda-4.4.0/lib:$LD_LIBRARY_PATH
export PYTHONPATH=$HOME/opencv-cuda-4.4.0/lib/python3.7/site-packages:$PYTHONPATH

python darknet_images.py -h

python darknet_images.py \
--batch_size 1 \
--thresh 0.1 \
--ext_output \
--config_file cfg/yolov4.cfg \
--data_file cfg/coco.data \
--weights $MY_MODEL_DIR/yolov4.weights \
--input $MY_COCO_DIR/test2017/000000000001.jpg

推断结果,如前一小节。