Code for "Invariant and Transportable Representations for Anti-Causal Domain Shifts"

Synthetic Dataset

python syn_running_script.py

Results will be saved as csv files in folder log_syn_classification. You can use csv_reading_script.py to get summary statistics by changing dir.

To get disentanglement plot disentangle_syn.png, run

python main.py --model_name=adp_invar_anti_causal --lr 1e-2 --classification --causal_dir_syn=anti --disentagnle_plot

Synthetic Dataset -- Counterexample

python syn_running_script.py --dataset=counter-example

Results will be saved as csv files in folder log_syn_classification. You can use csv_reading_script.py to get summary statistics by changing dir.

Color MNIST

python color_mnist_running_script.py

Results will be saved as csv files in folder log_color_mnist. You can use csv_reading_script.py to get summary statistics by changing dir.

To get disentanglement plot disentangle_color_mnist.png, run

python main.py --model_name=adp_invar_anti_causal --classification --dataset=color_mnist --disentagnle_plot --lr=0.001 --n_outer_loop=15 --fine_tune_lr 1e-2 --reg_lambda=2.1544346900318834 --reg_lambda_2=46.41588833612777 --phi_odim 8

Camlyon17

bash run_cam.sh

Models and tensors will be saved in ./saved_model and one can run run_adaptive.sh to run adaptive tests. For Camlyon17, please run it with mutiple random seeds as suggested by WILDS. We can change random seed by change RANDOMSEED in the run_cam.sh file.

PACS

Use dataset/DomianBed/download.py --dataset=pacs to download dataset and change the data directory in run_pacs.sh.

bash run_pacs.sh

Models and tensors will be saved in ./saved_model and one can run run_adaptive.sh to run adaptive tests. You can run it with different random seed

VLCS

Use dataset/DomianBed/download.py --dataset=vlcs to download dataset and change the data directory in run_vlcs.sh.

bash run_vlcs.sh

Models and tensors will be saved in ./saved_model and one can run run_adaptive.sh to run adaptive tests. You can run it with different random seed

Hyper-Parameter

See hyper_param_tuning.py for details on running hyper parameter search. Feel free to run your own hyperparameter search.

Others

Set --nb_workers=0 if you see this warning Warning: Cannot set number of intraop threads after parallel work has started or after set_num_threads call when using native parallel backend