Open AruniRC opened 1 month ago
Thank you for the great resource!
Can this be installed purely via pip, as python packages or pre-built wheels, without requiring conda? This makes local installs much less messy.
Yes, absolutely. Refer to the following code:
POINTCEPT_VERSION=1.5.0
TORCH_VERSION=2.0.1
CUDA_VERSION=11.7
CUDNN_VERSION=8
ARGS=`getopt -o v:t:c: -l version:,torch:,cuda:,cudnn: -n "$0" -- "$@"`
[ $? != 0 ] && exit 1
eval set -- "${ARGS}"
while true ; do
case "$1" in
-v | --version)
POINTCEPT_VERSION=$2
shift 2
;;
-t | --torch)
TORCH_VERSION=$2
shift 2
;;
-c | --cuda)
CUDA_VERSION=$2
shift 2
;;
--cudnn)
CUDNN_VERSION=$2
shift 2
;;
--)
break
;;
*)
echo "Invalid option: $1"
exit 1
;;
esac
done
CUDA_VERSION_NO_DOT=`echo ${CUDA_VERSION} | tr -d "."`
BASE_TORCH_TAG=${TORCH_VERSION}-cuda${CUDA_VERSION}-cudnn${CUDNN_VERSION}-devel
IMG_TAG=pointcept/pointcept:v${POINTCEPT_VERSION}-pytorch${BASE_TORCH_TAG}
echo "POINTCEPT VERSION: ${POINTCEPT_VERSION}"
echo "TORCH VERSION: ${TORCH_VERSION}"
echo "CUDA VERSION: ${CUDA_VERSION}"
echo "CUDNN VERSION: ${CUDNN_VERSION}"
cat > ./Dockerfile <<- EOM
FROM pytorch/pytorch:${BASE_TORCH_TAG}
# Fix nvidia-key error issue (NO_PUBKEY A4B469963BF863CC)
RUN rm /etc/apt/sources.list.d/*.list
# Installing apt packages
RUN export DEBIAN_FRONTEND=noninteractive \
&& apt -y update --no-install-recommends \
&& apt -y install --no-install-recommends \
git wget tmux vim zsh build-essential cmake ninja-build libopenblas-dev libsparsehash-dev \
&& apt autoremove -y \
&& apt clean -y \
&& export DEBIAN_FRONTEND=dialog
# Install Pointcept base environment
RUN pip install --upgrade pip
RUN pip install h5py pyyaml sharedarray tensorboard tensorboardx yapf addict einops scipy plyfile termcolor timm ftfy regex tqdm
RUN conda install pytorch-cluster pytorch-scatter pytorch-sparse -c pyg -y
RUN pip install torch-geometric
RUN pip install spconv-cu${CUDA_VERSION_NO_DOT}
RUN pip install git+https://github.com/openai/CLIP.git
RUN pip install -U --index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/Triton-Nightly/pypi/simple/ triton-nightly
# Build pointops
RUN git clone https://github.com/Pointcept/Pointcept.git
RUN TORCH_CUDA_ARCH_LIST="5.2 6.0 6.1 7.0+PTX 8.0" pip install Pointcept/libs/pointops -v
# Build pointgroup_ops
RUN TORCH_CUDA_ARCH_LIST="5.2 6.0 6.1 7.0+PTX 8.0" pip install Pointcept/libs/pointgroup_ops -v
# Build swin3d
RUN TORCH_CUDA_ARCH_LIST="6.0 6.1 7.0+PTX 8.0" pip install -U git+https://github.com/microsoft/Swin3D.git -v
## Build MinkowskiEngine
#RUN git clone https://github.com/NVIDIA/MinkowskiEngine.git
#WORKDIR /workspace/MinkowskiEngine
#RUN TORCH_CUDA_ARCH_LIST="5.2 6.0 6.1 7.0+PTX 8.0" python setup.py install --blas=openblas --force_cuda
#WORKDIR /workspace
EOM
#docker build . -f ./Dockerfile -t $IMG_TAG
docker build . --no-cache -f ./Dockerfile -t $IMG_TAG
Thank you for the great resource!
Can this be installed purely via pip, as python packages or pre-built wheels, without requiring conda? This makes local installs much less messy.