This is the implementation of Towards a Theoretical Framework of Out-of-Distribution Generalization. Our code is inherited from DomainBed.

Dataset

Downloads VLCS dataset and link it to

/datasets/VLCS

The structure of dataset should be same as DomainBed. To run our full experiments, one should also download PACS and OfficeHome dataset in the same way.

Single Experiment

To run a single experiment, run the following command:

python -m main --data_dir domainbed/datasets \
--trial_seed 89500 --algorithm ERM --dataset VLCS --test_envs 1 \
--steps 5001 --output_dir logs/ERM_VLCS_test_env1/ \
--extract_feature ERM_exp1  \
--output_result_file result.csv \
--checkpoint_freq 500 \
--save_feature_every_checkpoint \
--hparams "{\"lr\": 0.0001, \"resnet18\": false}" \

The final result will be saved in logs/ERM_VLCS_test_env1/ERM_exp1

Model Selection

To reproduce our experiment, one can run the following code to generate different models for model selection

python my_launcher.py --command_launcher my_multi_gpu launch

After model generation, one should use the following codes to collect feature and metric from different models. The result will be record in renamed/*.csv

cd logs
python collect_feature.py

to use our model selection criterion, one can run

python model_selection.py

This command use the result in renamed/*.csv to select model according to different criterion and report their test evironment accuracy. train_acc means the selection criterion is the average training accuracy, train_mix is the selection criterion defined in our paper