환경 구축

[jyh11224]

ssh로 서버 원격 접속 및 vscode 연동

[m0v3m1n]

DinoV2 모델 Git clone 및 Dockerfile 생성.

FROM nvidia/cuda:11.8.0-devel-ubuntu22.04

RUN apt-get update -y && apt-get -y install sudo git software-properties-common python3.9 python3-pip libgl1-mesa-glx

RUN mkdir src
WORKDIR /src
RUN git clone https://github.com/Media-4-Machine-Laboratory/DinoV2.git

WORKDIR /src/DinoV2
RUN pip3 install -r requirements.txt
RUN pip3 install -e .

[jyh11224]

1. requirements.txt

   • fire==0.2.1
   • ipython==6.5.0
   • matplotlib==3.0.3
   • numpy==1.17.2
   • parse==1.12.1
   • pip==19.3.1
   • plac==0.9.6
   • py==1.6.0
   • scipy==1.3.1
   • scikit-image==0.16.2
   • scikit-learn==0.22.1
   • torch==1.2.0
   • torchvision==0.4.0
   • tqdm==4.36.1
   • visdom==0.1.8.9
   • pathlib==1.0.1
   • pyyaml==5.3.1
   • pandas
   • seaborn
   • statsmodels

2. Jetson-inference 모델 실험 환경 생성 중 오류 발생

   • github에 private하게 접근하기 위해 키를 설정하는데 이 부분에서 문제가 발생
   • miniconda 설치 부분에서 에러 발생

[jyh11224]

github에 접근 불가하여 dockerhub에서 jetson-inference 검색 후 TAG 22.06 이미지 받기

[jyh11224]

• Docker 컨테이너 생성 sudo docker run --gpus all -it --shm-size=8g --ulimit memlock=-1 --ulimit stack=67108864 -e NVIDIA_DRIVER_CAPABILITIES=all dustynv/jetson-inference:22.06

• pretrained model 다운 bash download-models.sh

• 실행 명령어 # C++ $ ./depthnet "images/room_*.jpg" images/test/depth_room_%i.jpg # Python ./depthnet.py "images/room_*.jpg" images/test/depth_room_%i.jpg

[jyh11224]

• 추가 옵션 optional --network flag which changes the depth model being used (recommended default is fcn-mobilenet). optional --visualize flag which can be comma-separated combinations of input, depth

  • The default is --visualize=input,depth which displays the input and depth images side-by-side
  • To view only the depth image, use --visualize=depth optional --depth-size value which scales the size of the depth map relative to the input (the default is 1.0) optional --filter-mode flag which selects point or linear filtering used for upsampling (the default is linear) optional --colormap flag which sets the color mapping to use during visualization (the default is viridis_inverted)

• 도커 컨테이너 사용시 it's recommended to save the output images to the images/test mounted directory. These images will then be easily viewable from your host device under jetson-inference/data/images/test (for more info, see [Mounted Data Volumes].