tcri

Code and data for Olawale Salaudeen, Sanmi Koyejo, "Causally Inspired Regularization Enables Domain General Representations" International Conference on Artificial Intelligence and Statistics, 2024

https://arxiv.org/abs/2404.16277

Create Spurious PACS

Spurious PACS is a version of PACS that induces a stronger distribution shift by spuriously correlating original PACS domains with a binary label.

python create_spurious_PACS.py \
    --original_PACS_dir <> \
    --spurious_PACS_dir <> \
    --confound_strength 1.0

Run Experiments

Please see DomainBed/README.md for more detail.

1. python DomainBed/setup.py install

2. ColoredMNIST.

python -m domainbed.scripts.sweep launch \
    --command_launcher multi_gpu \
    --data_dir <> \
    --datasets ColoredMNIST \
    --algorithms ERM IRM GroupDRO VREx IB_ERM IB_IRM TCRI_HSIC \
    --n_trials 25 \
    --n_trials 5 \
    --output_dir <> \
    --single_test_envs

3. Spurious PACS

python -m domainbed.scripts.sweep launch \
    --command_launcher multi_gpu \
    --data_dir <> \
    --datasets SpuriousPACS \
    --algorithms ERM IRM GroupDRO VREx IB_ERM IB_IRM TCRI_HSIC \
    --n_trials 5 \
    --n_trials 3 \
    --output_dir <> \
    --single_test_envs

4. TerraIncognita

python -m domainbed.scripts.sweep launch \
    --command_launcher multi_gpu \
    --data_dir <> \
    --datasets TerraIncognita \
    --algorithms ERM IRM GroupDRO VREx IB_ERM IB_IRM TCRI_HSIC \
    --n_trials 5 \
    --n_trials 3 \
    --output_dir <> \
    --single_test_envs

Ablation

The following flags --ablat_TCRI_TCRI and --ablat_TCRI_TIC can be used to ablate the TCRI and TIC penalties, respectively.

Changes to DomainBed

We primarly run our experiments with the included version of DomainBed. The primary difference between this version and the original is the addition of the TCRI algorithm to Domainbed/domainbed/algorithms.py and SpuriousPACS to Domainbed/domainbed/datasets.py. Additionally, to facilitate model selection using the 'tcri' criterion, we store the 'tcri' and 'tic' values in 'results.jsonl' which is a record of model quality during training.

When the algorithm is not TCRI, its 'tcri' and 'tic' value is set to -1 by default. This can be found at DomainBed/domainbed/lib/misc.py.