Implementation / Dockerfile

[J4nekT]

Hello everyone,

I would like to implement your code on my computer with a GTX3090 (driver version: 535.154.05). I would like to use a dockerfile to do this that looks like this:

FROM nvidia/cuda:11.3.1-devel-ubuntu20.04

RUN apt-get update && apt-get install wget -yq
RUN apt-get install build-essential g++ gcc -y
ENV DEBIAN_FRONTEND noninteractive
RUN apt-get install libgl1-mesa-glx libglib2.0-0 -y
RUN apt-get install openmpi-bin openmpi-common libopenmpi-dev libgtk2.0-dev git -y

ENV CONDA_DIR /opt/conda  
RUN wget --quiet https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O ~/miniconda.sh && \
     /bin/bash ~/miniconda.sh -b -p /opt/conda
# Put conda in path so we can use conda activate
ENV PATH=$CONDA_DIR/bin:$PATH
ENV TORCH_CUDA_ARCH_LIST="8.6"
RUN conda install python=3.8
RUN conda install pytorch==1.8.1 torchvision==0.9.1 torchaudio==0.8.1 cudatoolkit=11.3 -c pytorch -c conda-forge
# Install additional required packages
RUN pip install tensorboardX~=2.1 \
                setuptools~=39.0.1 \
                opencv-python~=4.5.1.48 \
                umsgpack~=0.1.0 \
                numpy~=1.21 \
                Pillow~=9.0.0 \
                configparser~=5.0.1 \
                kornia~=0.2.0 \
                git+https://github.com/mapillary/inplace_abn.git

The dockerfile creates a container with CUDA 11.3 and Python 3.8. Unfortunately, I was unable to create a container with CUDA 11.1 and Python 3.6. When I tried to run python3 setup.py develop, I got the following error:
ValueError: Unknown CUDA arch (8.6) or GPU not supported

I've identified two possible sources of the error, but I'm not sure if they are the real cause.

1. When installing Setuptool~=39.0.1, the _distutils_hack module disappears.
2. When installing kornia~=0.2.0, the previously installed PyTorch version 1.8.1 is reset to 1.4.

Unfortunately, I was not able to fix the problem myself. I hope to find a solution here, or possibly a working Dockerfile.

Best regards