README.md

Experiments: crc on multiple datasets

As an example, here is how to run one of our experiments:

1. Conformalize the predictor: compute optimal $\hat{\lambda}$ on calibration data $D_{\text{calib}}$.
   • Multiple runs: for better empirical evaluation, you can repeat with multiple random seeds (see config files)
2. Evaluate conformalization on $D_{\text{test}}$: we expect empirical risk to be close to nominal risk $\alpha$
   • For each random split in conformalization (point (1)), we test on the remaining data.

For details, see directly the python scripts and the configuration files called within the .sh scripts below.

(1) Conformalize the predictor

Example: Cityscapes, conformalize miscoverage loss with CRC on the second gpu (cuda:1)

You will need to specify the gpu to run the experiment on.

• If only 1 gpu: cuda:0
• if 2 or more gpus: cuda:0 or cuda:1 etc.
• if NO gpus: cpu

$ cd /path/to/cose
$ source .venv/bin/activate
(.venv) $ sh experiments/run_crc_miscov_loss_cityscapes.sh cuda:1

Post-process the output of conformalization:

(.venv) $ python ./cose/experiments/postprocess_expes.py --input-dir experiments/outputs/Cityscapes/miscoverage_loss/

Or directly post-processing all datasets and losses in experiments/outputs:

sh experiments/process_all_opt_lambdas.sh

(2) Evaluation on test split

Example: Cityscapes.

$ cd /path/to/output
$ source .venv/bin/activate
(.venv) $ sh experiments/[...].sh