We suggest to setup a conda environment:
conda create -n msc python=3.7
pip install torch==1.9 torchvision ujson
The code is written mainly for torch. The original code for training robust networks was written in Jax however. The folder loss_implementation_test
contains code for making sure the implementation in torch is correct. For this Jax is required:
pip install --upgrade pip
pip install --upgrade "jax[cuda111]" -f https://storage.googleapis.com/jax-releases/jax_releases.html
Please not that this donwloads Jax for the GPU (supported for linux only). If you need the CPU version, just do pip install jax
.