HorizonNet在Jetson AGX Xavier部署

一、硬件环境

硬件：Nvidia Jetson Xavier AGX
NVIDIA SDK Manager Version: 1.2.0
CUDA版本：10.2.89
Jetpack版本：JP 4.4 DP
python版本：python2.7 python3.6

HorizonNet：https://github.com/sunset1995/HorizonNet

二、软件要求

• Python 3
• pytorch>=1.0.0
• numpy
• scipy
• sklearn
• Pillow
• tqdm
• tensorboardX
• opencv-python>=3.1 (for pre-processing) (also can’t be too new, the latest opencv removed a key algorithm due to patent, 3.1.0.5 works. )
• open3d>=0.7 (for layout 3D viewer)
• shapely
• torchvision

三、安装

1、系统自带的软件包

• python3: 系统自带版本为python3.6.9
• numpy: 1.13.3
• scipy: 0.19.1
• pillow: 7.2.0
• opencv: 4.1.1

1、pytorch，torchvision安装

https://blog.csdn.net/u013595260/article/details/107180928

2、opencv降级

https://blog.csdn.net/a_lien/article/details/103615619

opencv4对于HorizonNet来说，版本太高无法运行，降级为opencv3.4

sudo apt-get purge libopencv*
git clone https://github.com/jetsonhacks/buildOpenCVXavier
cd buildOpenCVXavier
git checkout v1.0
./buildOpenCV.sh

3、numpy、scipy版本不匹配

运行时报错：
ModuleNotFoundError: No module named 'numpy.testing.nosetester'

原因：
numpy、sklearn、scipy、joblib版本不兼容。

解决：

sudo apt install liblapack-dev
sudo apt install libblas-dev
sudo apt install gfortran

pip3 install --user --upgrade numpy 
pip3 install --user --upgrade scikit-learn
pip3 install --user --upgrade scipy
pip3 install --user --upgrade joblib

4、其他安装

sudo apt-get install python3-sklearn
pip3 install tqdm
pip3 install tensorboardX
pip3 install open3d-python		# 测试失败
sudo apt-get install python3-shapely

五、测试

python3 inference.py --pth ckpt/resnet50_rnn__mp3d.pth --img_glob assets/preprocessed/demo_aligned_rgb.png --output_dir assets/inferenced

程序启动时间：4s
模型加载时间：10s
计算时间：8s（1024x512 pixels）、12s（11520x5760 pixels）