NATTEN is a PyTorch CUDA extension, therefore requires PyTorch. The following are the recommended versions of these libraries and are strongly encouraged for reproducibility and speed:
torch==1.11.0+cu113
torchvision==0.12.0+cu113
ninja==1.10.2.3PyTorch version 1.11 is strongly recommended as the CUDA extension runs faster due to an updated
version of the atomic add operator we use for backpropagation.
The recommended way of installing it is through ninja.
By having ninja installed, you don't need to install anything manually.
The extension will compile when it's called first.
To compile and ensure that the extension is functioning correctly, please run:
python3 gradcheck.py
If you want to build without JIT, simply build with setup:
python setup.py installAfter it builds, please run the following to ensure that the extension is functioning correctly:
python3 gradcheck.py