Settings

Stressful Settings.

Dependencies

  • Detectron2
    • python 3.7 >=
    • torch 1.8 >=
  • scikit-learn == 0.22
    • it takes a while to set
    • pip: Cython
    • apt-get: gcc, g++
  • official neural renderer is tensorflow version.

Errors

“ImportError: libGL.so.1: cannot open shared object file: No such file or directory”

  • apt-get update && apt-get install ffmpeg libsm6 libxext6 -y

ImportError: (‘Unable to load OpenGL library’, ‘OSMesa: cannot open shared object file: No such file or directory’, ‘OSMesa’, None)

  • install OSMesa

AttributeError: ‘PosixPath’ object has no attribute ‘path’

  • link
  • link2
  • PosixPath? When I print os.name, ‘posix’ returned.

Docker

How to delete <none> docker images?

docker rmi $(docker images -f “dangling=true” -q)
After I tried this, every dockers are deleted.


Installation

OpenPose

  • Install guide link, link2
  • I’m using Ubuntu server and don’t have GUI
  • git clone regressively.
  • …. I just download openpose docker image. docker pull d0ckaaa/openpose
  • cmake ..
  • make -j4nproc
  • how to use: https://viso.ai/deep-learning/openpose/

OSMesa

  • Try this. link
  • Before try this,
    • apt-get install wget
    • apt-get install pkg-config
    • apt-get install libexpat1-dev
  • i installed in /usr/local
  • MESA_HOME=/usr/local (+/lib?)
  • issues
    • https://redstarhong.tistory.com/98 GL? PyOpenGL?
    • update “/opt/conda/lib/python3.9/site-packages/OpenGL/platform/egl.py”
    • apt-get install libosmesa6-dev
    • if PyOpenGL is failed, clone git and reinstall PyOpenGL.

PHALP

I failed while using conda yml file because:

  • First, I installed conda inside of a docker container to keep experiments going on.
    • when I tried to create conda env, it takes more than an hour to solve dependencies, but it still doesn’t finished.
  • Then, I tried to install every commands on my own, neural renderer and cuda path problems occurred.
    • It was hard for me to find cuda path in conda env because someone said that each conda environment would only contain the parts of cuda
    • So I should export every paths of cuda related stuff, but when I searched cuda paths in my server computer, there are too many other settings, so I’m not sure what to add
  • These are the reasons why I didn’t use conda

I use these commands sequentially. After that, demo works well on my docker container. Official PHALP’s colab pip output: phalp_colab.txt

! pip install pyrender
! pip install opencv-python
! pip install joblib
! pip install cython
! pip install scikit-learn==0.22
! pip install scikit-image
! pip install git+https://github.com/facebookresearch/detectron2.git
! pip install git+https://github.com/brjathu/pytube.git
! pip install git+https://github.com/brjathu/NMR.git
! pip install chumpy
! pip install ipython
! pip install gdown
! pip install dill
! pip install scenedetect[opencv]

But if you use v1.1, additional packages are needed.

  • Install OSMesa
  • pip install smplx submitit rich pyrootutils colordict
  • pip install -e .
  • install hydra-core again (Posix error)
  • python scripts/demo.py video.source='“https://www.youtube.com/watch?v=xEH_5T9jMVU”'
    • I use osmesa, not egl image
Written on April 7, 2023