TX2之YOLOV3环境部署
参考：https://jkjung-avt.github.io/yolov3/
0 背景
YOLOV3作为目标检测现阶段性能最好的算法之一，具有很强的实用性，在TX2上部署YOLOV3可以解决很多现实的目标检测问题。
环境依赖：opencv3.4.0 + cuda9.0 + cudnn7.1.5
1 opencv3.4.0
网上介绍说opencv3.4.1运行yolo会有问题，因此先卸载旧版本opencv，然后安装3.4.0版本的，具体安装方法可参考《TX2查看OpenCV版本并卸载重装》，也可按照以下方法安装
1.1 安装依赖
### Remove all old opencv stuffs installed by JetPack (or OpenCV4Tegra)
$ sudo apt-get purge libopencv*
### I prefer using newer version of numpy (installed with pip), so
### I'd remove this python-numpy apt package as well
$ sudo apt-get purge python-numpy
### Remove other unused apt packages
$ sudo apt autoremove
### Upgrade all installed apt packages to the latest versions (optional)
$ sudo apt-get update
$ sudo apt-get dist-upgrade
### Update gcc apt package to the latest version (highly recommended)
$ sudo apt-get install --only-upgrade g++-5 cpp-5 gcc-5
### Install dependencies based on the Jetson Installing OpenCV Guide
$ sudo apt-get install build-essential make cmake cmake-curses-gui \
g++ libavformat-dev libavutil-dev \
libswscale-dev libv4l-dev libeigen3-dev \
libglew-dev libgtk2.0-dev
### Install dependencies for gstreamer stuffs
$ sudo apt-get install libdc1394-22-dev libxine2-dev \
libgstreamer1.0-dev \
libgstreamer-plugins-base1.0-dev
### Install additional dependencies according to the pyimageresearch
### article
$ sudo apt-get install libjpeg8-dev libjpeg-turbo8-dev libtiff5-dev \
libjasper-dev libpng12-dev libavcodec-dev
$ sudo apt-get install libxvidcore-dev libx264-dev libgtk-3-dev \
libatlas-base-dev gfortran
$ sudo apt-get install libopenblas-dev liblapack-dev liblapacke-dev
### Install Qt5 dependencies
$ sudo apt-get install qt5-default
### Install dependencies for python3
$ sudo apt-get install python3-dev python3-pip python3-tk
$ sudo pip3 install numpy
$ sudo pip3 install matplotlib
1.2 源码编译
### Download opencv-3.4.0 source code
$ mkdir -p ~/opencv3.4.0
$ cd ~/opencv3.4.0
$ wget https://github.com/opencv/opencv/archive/3.4.0.zip \
-O opencv-3.4.0.zip
$ unzip opencv-3.4.0.zip
### Build opencv (CUDA_ARCH_BIN="6.2" for TX2, or "5.3" for TX1)
$ cd ~/opencv3.4.0/opencv-3.4.0
$ mkdir build
$ cd build
$ cmake -D CMAKE_BUILD_TYPE=RELEASE -D CMAKE_INSTALL_PREFIX=/usr/local \
-D WITH_CUDA=ON -D CUDA_ARCH_BIN="6.2" -D CUDA_ARCH_PTX="" \
-D WITH_CUBLAS=ON -D ENABLE_FAST_MATH=ON -D CUDA_FAST_MATH=ON \
-D ENABLE_NEON=ON -D WITH_LIBV4L=ON -D BUILD_TESTS=OFF \
-D BUILD_PERF_TESTS=OFF -D BUILD_EXAMPLES=OFF \
-D WITH_QT=ON -D WITH_OPENGL=ON ..
$ make -j4
$ sudo make install配置信息
-- General configuration for OpenCV 3.4.0 =====================================
-- Version control: unknown
--
-- Platform:
-- Timestamp: 2019-08-06T03:14:54Z
-- Host: Linux 4.4.38-tegra aarch64
-- CMake: 3.5.1
-- CMake generator: Unix Makefiles
-- CMake build tool: /usr/bin/make
-- Configuration: RELEASE
--
-- CPU/HW features:
-- Baseline: NEON FP16
-- required: NEON
-- disabled: VFPV3
--
-- C/C++:
-- Built as dynamic libs?: YES
-- C++ Compiler: /usr/bin/c++ (ver 5.4.0)
-- C++ flags (Release): -fsigned-char -W -Wall -Werror=return-type -Werror=non-virtual-dtor -Werror=address -Werror=sequence-point -Wformat -Werror=format-security -Wmissing-declarations -Wundef -Winit-self -Wpointer-arith -Wshadow -Wsign-promo -Wuninitialized -Winit-self -Wno-narrowing -Wno-delete-non-virtual-dtor -Wno-comment -fdiagnostics-show-option -pthread -fomit-frame-pointer -ffunction-sections -fdata-sections -fvisibility=hidden -fvisibility-inlines-hidden -O3 -DNDEBUG -DNDEBUG
-- C++ flags (Debug): -fsigned-char -W -Wall -Werror=return-type -Werror=non-virtual-dtor -Werror=address -Werror=sequence-point -Wformat -Werror=format-security -Wmissing-declarations -Wundef -Winit-self -Wpointer-arith -Wshadow -Wsign-promo -Wuninitialized -Winit-self -Wno-narrowing -Wno-delete-non-virtual-dtor -Wno-comment -fdiagnostics-show-option -pthread -fomit-frame-pointer -ffunction-sections -fdata-sections -fvisibility=hidden -fvisibility-inlines-hidden -g -O0 -DDEBUG -D_DEBUG
-- C Compiler: /usr/bin/cc
-- C flags (Release): -fsigned-char -W -Wall -Werror=return-type -Werror=non-virtual-dtor -Werror=address -Werror=sequence-point -Wformat -Werror=format-security -Wmissing-declarations -Wmissing-prototypes -Wstrict-prototypes -Wundef -Winit-self -Wpointer-arith -Wshadow -Wuninitialized -Winit-self -Wno-narrowing -Wno-comment -fdiagnostics-show-option -pthread -fomit-frame-pointer -ffunction-sections -fdata-sections -fvisibility=hidden -O3 -DNDEBUG -DNDEBUG
-- C flags (Debug): -fsigned-char -W -Wall -Werror=return-type -Werror=non-virtual-dtor -Werror=address -Werror=sequence-point -Wformat -Werror=format-security -Wmissing-declarations -Wmissing-prototypes -Wstrict-prototypes -Wundef -Winit-self -Wpointer-arith -Wshadow -Wuninitialized -Winit-self -Wno-narrowing -Wno-comment -fdiagnostics-show-option -pthread -fomit-frame-pointer -ffunction-sections -fdata-sections -fvisibility=hidden -g -O0 -DDEBUG -D_DEBUG
-- Linker flags (Release):
-- Linker flags (Debug):
-- ccache: NO
-- Precompiled headers: YES
-- Extra dependencies: dl m pthread rt /usr/lib/aarch64-linux-gnu/libGLU.so /usr/lib/aarch64-linux-gnu/libGL.so cudart nppc nppial nppicc nppicom nppidei nppif nppig nppim nppist nppisu nppitc npps cublas cufft -L/usr/local/cuda-9.0/lib64
-- 3rdparty dependencies:
--
-- OpenCV modules:
-- To be built: calib3d core cudaarithm cudabgsegm cudacodec cudafeatures2d cudafilters cudaimgproc cudalegacy cudaobjdetect cudaoptflow cudastereo cudawarping cudev dnn features2d flann highgui imgcodecs imgproc ml objdetect photo python3 python_bindings_generator shape stitching superres video videoio videostab
-- Disabled: js world
-- Disabled by dependency: -
-- Unavailable: java python2 ts viz
-- Applications: apps
-- Documentation: NO
-- Non-free algorithms: NO
--
-- GUI:
-- QT: YES (ver 5.5.1)
-- QT OpenGL support: YES (Qt5::OpenGL 5.5.1)
-- GTK+: NO
-- OpenGL support: YES (/usr/lib/aarch64-linux-gnu/libGLU.so /usr/lib/aarch64-linux-gnu/libGL.so)
-- VTK support: NO
--
-- Media I/O:
-- ZLib: /usr/lib/aarch64-linux-gnu/libz.so (ver 1.2.8)
-- JPEG: /usr/lib/aarch64-linux-gnu/libjpeg.so (ver )
-- WEBP: build (ver encoder: 0x020e)
-- PNG: /usr/lib/aarch64-linux-gnu/libpng.so (ver 1.2.54)
-- TIFF: /usr/lib/aarch64-linux-gnu/libtiff.so (ver 42 / 4.0.6)
-- JPEG 2000: /usr/lib/aarch64-linux-gnu/libjasper.so (ver 1.900.1)
-- OpenEXR: build (ver 1.7.1)
--
-- Video I/O:
-- DC1394: YES (ver 2.2.4)
-- FFMPEG: YES
-- avcodec: YES (ver 56.60.100)
-- avformat: YES (ver 56.40.101)
-- avutil: YES (ver 54.31.100)
-- swscale: YES (ver 3.1.101)
-- avresample: NO
-- GStreamer:
-- base: YES (ver 1.8.3)
-- video: YES (ver 1.8.3)
-- app: YES (ver 1.8.3)
-- riff: YES (ver 1.8.3)
-- pbutils: YES (ver 1.8.3)
-- libv4l/libv4l2: 1.10.0 / 1.10.0
-- v4l/v4l2: linux/videodev2.h
-- gPhoto2: NO
--
-- Parallel framework: pthreads
--
-- Trace: YES (built-in)
--
-- Other third-party libraries:
-- Lapack: YES (/usr/lib/liblapack.so /usr/lib/libcblas.so /usr/lib/libatlas.so)
-- Eigen: YES (ver 3.2.92)
-- Custom HAL: YES (carotene (ver 0.0.1))
--
-- NVIDIA CUDA: YES (ver 9.0, CUFFT CUBLAS FAST_MATH)
-- NVIDIA GPU arch: 62
-- NVIDIA PTX archs:
--
-- OpenCL: YES (no extra features)
-- Include path: /home/nvidia/opencv3.4.0/opencv-3.4.0/3rdparty/include/opencl/1.2
-- Link libraries: Dynamic load
--
-- Python 3:
-- Interpreter: /usr/bin/python3 (ver 3.5.2)
-- Libraries: /usr/lib/aarch64-linux-gnu/libpython3.5m.so (ver 3.5.2)
-- numpy: /usr/local/lib/python3.5/dist-packages/numpy/core/include (ver 1.15.4)
-- packages path: lib/python3.5/dist-packages
--
-- Python (for build): /usr/bin/python2.7
--
-- Java:
-- ant: NO
-- JNI: NO
-- Java wrappers: NO
-- Java tests: NO
--
-- Matlab: NO
--
-- Install to: /usr/local
-- -----------------------------------------------------------------
安装完之后验证一下
$ python3 -c 'import cv2; print(cv2.__version__)'
3.4.02 yolov3
首先下载源码
$ cd ~/project
$ git clone https://github.com/pjreddie/darknet yolov3
$ cd yolov3修改Makefile
将前三个设置为1，并将ARCH改为如下
GPU=1
CUDNN=1
OPENCV=1
......
# Note that CUDA architecture of TX2 is “62”, while TX1 “53”.
ARCH= -gencode arch=compute_53,code=[sm_53,compute_53] \
-gencode arch=compute_62,code=[sm_62,compute_62]然后运行
make -j4即可编译生成darket可执行文件
3 图片测试
首先下载预训练模型
wget https://pjreddie.com/media/files/yolov3.weights
wget https://pjreddie.com/media/files/yolov3-tiny.weights
修改配置文件cfg/yolov3.cfg，将前两行打开，关闭training中的batch和subdivisions，否则会报内存错误
[net]
# Testing
batch=1
subdivisions=1
# Training
# batch=64
# subdivisions=16测试
./darknet detect cfg/yolov3.cfg yolov3.weights data/dog.jpg
./darknet detect cfg/yolov3-tiny.cfg yolov3-tiny.weights data/dog.jpg测试完成会出现如下信息，并弹出检测图片
layer filters size input output
0 conv 32 3 x 3 / 1 416 x 416 x 3 -> 416 x 416 x 32 0.299 BFLOPs
1 conv 64 3 x 3 / 2 416 x 416 x 32 -> 208 x 208 x 64 1.595 BFLOPs
.......
105 conv 255 1 x 1 / 1 52 x 52 x 256 -> 52 x 52 x 255 0.353 BFLOPs
106 detection
truth_thresh: Using default '1.000000'
Loading weights from yolov3.weights...Done!
data/dog.jpg: Predicted in 0.029329 seconds.
dog: 99%
truck: 93%
bicycle: 99%
4 视频测试
下载一个测试视频traffic.mp4,
### set TX2 to max performance mode
$ sudo nvpmodel -m 0
$ sudo ~/jetson_clocks.sh
### download the pre-trained weights and run YOLOv3
$ wget https://pjreddie.com/media/files/yolov3.weights
$ ./darknet detector demo cfg/coco.data cfg/yolov3.cfg yolov3.weights traffic.mp4如果要读取USB摄像头测试，运行下边的代码，-c 1表示/dev/video0
$ ./darknet detector demo cfg/coco.data cfg/yolov3.cfg yolov3.weights -c 0实际测试时，使用yolov3.weights在TX2上测试时，推理速度只有3~3.3 FPS，还达不到实时的效果