PyTorch GPU Support for ShapeWorks
ShapeWorks deep learning tools, such as the
DeepSSMUtils package, requires PyTorch with GPU support.
This is installed with the rest of the ShapeWorks Anaconda environment using
It selects the most recent stable release of PyTorch which can be found at pytorch.org.
When the Anaconda enironment is created using
install_shapeworks, PyTorch with GPU support is installed if the system's current CUDA driver version is supported. Otherwise it selects the CPU version of PyTorch.
Checking if PyTorch installation has GPU support
To check if your
shapeworks environment has PyTorch with GPU support, run the following:
conda activate shapeworks python >>> import torch >>> print(torch.cuda.is_available()) >>> exit()
torch.cuda.is_available()is True then PyTorch has GPU support, otherwise the CPU version was installed. If
torchcannot be imported than PyTorch was not installed to the
Reinstalling the Correct Pytorch Version
If you find that your system requires an older version of PyTorch or
install_shapeworks did not correcty find your CUDA version,
the following steps explain how to install a different PyTorch version in the
shapeworks conda environment.
Deltailed instructions about the different ways to install PyTorch can be found here: PyTorch Getting Started.
- If the CPU version of PyTorch was installed, that first needs to be uninstalled. To uninstall run:
conda activate shapeworks conda uninstall pytorch
- Check which CUDA version is installed on your system using one of the methods explained here: How to check CUDA version
Install the correct PyTorch to
conda install pytorch torchvision cudatoolkit=VERSION -c pytorch
VERSIONis your CUDA version (such as 11.0).
Restart your system and check if
shapeworksnow has PyTorch with GPU support using the instructions above.