Sequential Multivariate Change Detection with Calibrated and Memoryless False Detection Rates (AISTATS 2022)

To run the experiments in Section 5.1 first install poetry if you do not already have it:

pip install poetry

Then instantiate an environment and install the required dependencies

poetry shell
poetry install

This doesn't install pytorch. Do this following the instructions at https://pytorch.org/.

To run experiments then run

cd experiments/toy_examples
python run.py

This obtains ARTs and ADDs for desired ERTs of 128, 256, 512, 1024. In the paper we perform additional experiments for batchmmdsim and lsddincsim with ERTs set equal to the ARTs obtained by bstat and lsddinc respectively. This can be performed by updating the ert parameters in gen_config.yaml files.

---

Running the experiments in Section 5.2 requires a couple of additional dependencies which are also better installed manually. First run

pip install torch-scatter -f https://data.pyg.org/whl/torch-${TORCH}+${CUDA}.html
pip install wilds==1.1.0

being careful to ensure ${TORCH} and ${CUDA} match your torch installation and CUDA version.

Then navigate into experiments/camelyon and run run.py. It will take a while to download the data the firts time this is run.