conda update -n base -c defaults conda
conda create -n taskshift python==3.10
conda activate taskshift
conda install pytorch==1.12.1 torchvision==0.13.1 cudatoolkit=11.3 -c pytorch
pip install -r requirements.txt
To see all commands and options, run
python shift.py -h
The most important option is --cov_type, which enables either isotropic, spiked, or polynomial decay covariance.
Usage example for a simulation with spiked covariance and 2-sparse signal:
python shift.py --cov_type spiked --sparse_inds 1 2 --sparse_vals 0.2 -0.1 --spiked_r 0.5 --spiked_q 0.6
If you are not using a GPU, set --cuda False. It is True by default. If the code runs too slow, you may want to try --solver gd, although sometimes gradient descent is slower than solving for the MNI directly.