radekd91
Installation break
Hi,
Current installation for python3.8 seems to severely break. There are issues with mediapipe 0.8.10 (too old), pytorch3d 0.6.2 (too old), mmcv, and more. I am testing the installation on a fresh ubuntu 22.04 ec2 aws instance, but could not successfully make it work. Would greatly appreciate if you guys can give a check.
This is what has worked for me so far.
#make sure sure you are using Cuda 11.3 conda create -n work38 python=3.8 conda activate work38 conda install mamba -n base -c conda-forge bash pull_submodules.sh pip install Cython==0.29.14 mamba install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.3 -c pytorch #IMPORTANT - My ubuntu was too new to build the wheel so you have to do this: conda install -c conda-forge gcc_linux-64=10 gxx_linux-64=10 pip install git+https://github.com/facebookresearch/[email protected] #IMPORTANT - download mediapipe 8.10 from StableBuild or my own copy: https://drive.google.com/file/d/1N-nW3gJMst8vO2qRkF1VFoXCOEqhbos1/view?usp=sharing #Place it in your top git folder pip install mediapipe-0.8.10-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl #IMPORTANT - onnxruntime_gpu is also missing. Download from StableBuild or my own copy: https://drive.google.com/file/d/1QyrrLLoIlWx-o9rovfPNC84ixQugEwSh/view?usp=sharing #Place it in your top git folder pip install onnxruntime_gpu-1.9.0-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl shell install_38.sh
Hi there,
thanks a lot @alonsorobots for sharing your setup.
@zyang-genies, I have always tested what I put out on my setup. Unfortunately, the environment is complex (with all the cuda, pytorch, pytorch3d versions etc) and it just seems that it does not always work for everybody.
Sadly, I don't have the bandwidth to look into every case and so I would encourage to update whatever packages you might have to update (try to the next supported versions). Likely, you'll find that after doings so, the code will run and hopefully without efforts. Of course PRs will be welcome if you stumble upon something critical.
Lastly, I would encourage to try the newer docker-based setup, so that you have an easier time isolating cuda versions and similar from your underlying system.
This is what has worked for me so far.
#make sure sure you are using Cuda 11.3 conda create -n work38 python=3.8 conda activate work38 conda install mamba -n base -c conda-forge bash pull_submodules.sh pip install Cython==0.29.14 mamba install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.3 -c pytorch #IMPORTANT - My ubuntu was too new to build the wheel so you have to do this: conda install -c conda-forge gcc_linux-64=10 gxx_linux-64=10 pip install git+https://github.com/facebookresearch/[email protected] #IMPORTANT - download mediapipe 8.10 from StableBuild or my own copy: https://drive.google.com/file/d/1N-nW3gJMst8vO2qRkF1VFoXCOEqhbos1/view?usp=sharing #Place it in your top git folder pip install mediapipe-0.8.10-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl #IMPORTANT - onnxruntime_gpu is also missing. Download from StableBuild or my own copy: https://drive.google.com/file/d/1QyrrLLoIlWx-o9rovfPNC84ixQugEwSh/view?usp=sharing #Place it in your top git folder pip install onnxruntime_gpu-1.9.0-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl shell install_38.sh
I'm trying to follow this fix but the google drive links don't work and it's not clear to me where to find the old versions of mediapipe.
Apologies, I was running out of space and moved the. Here is the StableBuild links:
https://dashboard.stablebuild.com/pypi-deleted-packages/pkg/mediapipe/0.8.10 https://dashboard.stablebuild.com/pypi-deleted-packages/pkg/onnxruntime-gpu/1.9.0