Causal Invariant Bayesian Neural Network

Repository of the Causal Invariant Bayesian Neural Network (CIBNN).

Installation

Run the following command to install the required packages:

pip install -r requirements.txt

You can install the packages in a virtual environment by running the following commands:

python -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Download data

The data used in the experiments are not included in the repository due to their large memory footprint. They are either automatically downloaded when the dataset is used or they must be downloaded separately, as indicated in the next table. The data can be downloaded from the following links:

Dataset	Split	Requires download	Link	File ID
MNIST		auto	torchvision documentation
CIFAR10		auto	torchvision documentation
OFFICEHOME		auto	project homepage
ACRE	IID	yes	project homepage	1P0WBnnjWolGsrATUQtx4ictiYlOGc-OT
	Comp	yes		1-LZMt08a1v-KSuaQTS1lqD6BCEw47LEY
	Comp	yes		1Sn_tKbe6mMv7Tc_y6hJZnm7lSenjwIys
CONCEPTARC		auto	github page
RAVEN	IID	yes	project homepage	111swnEzAY2NfZgeyAhVwQujMjRUfeyuY
	OOD	no (requires IID)

Download the data using the following command:

gdown <file_id>

file_id is the ID of the zip file in google drive, indicated in the last column. The zip must be placed in the corresponding data/<dataset_name> folder and unzipped:

cd data/<dataset_name>
unzip <file_name>.zip

Test

To run the tests, use the following command:

pytest

Make sure the data is properly downloaded before running the tests. Running the tests should take a couple minutes depending on your config.

Usage

To run the experiments from a config file, use the following command:

python run.py --load_config 'config/<config_file>.yaml' '<model_name>'

You can also pass arguments directly to the script:

python run.py --<parameter> '<value>' '<model_name>' --<model_parameter> '<value>'

Use the -h flag to see the available options:

python run.py -h

To see the model options, use the following command:

python run.py '<model_name>' -h

Common arguments:

• --load_config: Load the configuration file.
• --save_config: Save the configuration file.
• --data: Dataset to use.
• --save: Path to save the model.
• --train, --test, --train_and_test: Train, test, or train and test the model. Default option is train_and_test.
• --max_epochs: Maximum number of epochs. You might want to set it as the default value of Pytorch Lightning is 1000.

To run the experiments from the paper, use the provided config files in the config folder.

Out-of-distribution runs

When running an o.o.d config, replace the save parameter in the config file with your own save path. Here are some additional dataset-specific parameters that you can tweak:

Dataset	Parameter	Description
CIFAR10	--perturbation	Level of image perturbation. Float between 0.0 and 1.0.
OFFICEHOME	--split	The split to use. IID Options: RealWorld, Product, Art, Clipart. In OOD settings, write the test split next to the train split, e.g. to test a model on Product after training on RealWorld, write realWorld_Product.
ACRE	--split	The split to use. Options: IID, Comp, Sys.
RAVEN	--split	The split to use. Options: IID, ĪID_SMALL, IID_TRANSFER, OOD, OOD_SMALL, OOD_TRANSFER.

Hyperparameter search

To run a hyperparameter search, use the following command:

python hyp_search.py

You can use the same arguments as in the run.py script. Please, be aware that the Ray Tune trials are run in a different working directory: the path arguments must be absolute and the logging files are saved in a different directory (check Ray Tune documentation for more information).

Architecture