This repository contains scripts to implement SparC-IB from the paper Sparsity-Inducing Categorical Prior Improves Robustness of the Information Bottleneck. The codes in this repository is heavily based on the convexIB implementation.
Please follow the below code blocks to run the models used in the experiments section. Note that, the below codes run the models on MNIST dataset. Please pass --dataset 'CIFAR10' and the correct hyperparameters from the Appendix to run the models on CIFAR-10.
To run SparC-IB with compound strategy please use the below code.
cd ./SparC-IB-Compound
python3 train_model.py --repl_n 1 --beta 1 --K 100 --a 2.0 --b 2.0 --verbose --u_func_name 'pow' --eval_rate 1 --same_batch --dataset 'mnist' --method 'variational_IB'
To run SparC-IB with categorical strategy please use the below code.
cd ./SparC-IB-Categorical
python3 train_model.py --repl_n 1 --beta 1 --K 100 --a 2.0 --b 2.0 --verbose --u_func_name 'pow' --eval_rate 1 --same_batch --dataset 'mnist'
Note that, the above codes run SparC-IB with the hyperparameters (a,b)=(2,2) for both compound and categorical strategy. For other values, pass different numbers (as floats) in the arguments --a and --b.
To run Drop-VIB model please use the below code.
cd ./SparC-IB-Compound
python3 train_model.py --repl_n 1 --beta 1 --K 100 --verbose --u_func_name 'pow' --eval_rate 1 --same_batch --dataset 'mnist' --method 'drop_VIB'
To run Intel-VIB model please use the below code.
cd ./SparC-IB-Compound
python3 train_model.py --repl_n 1 --beta 1 --K 100 --verbose --u_func_name 'pow' --eval_rate 1 --same_batch --dataset 'mnist' --method 'intel_VIB'
To run Fixed-K-VIB model please use the below code.
cd ./Fixed-K
python3 train_model.py --repl_n 1 --beta 1 --K 32 --verbose --u_func_name 'pow' --eval_rate 1 --same_batch --dataset 'mnist' --method 'variational_IB'
Note that, the above code will run the Fixed K:32 model. To run for different latent dimensions please change the --K argument in the above code. Furthermore, all the above code is for one seed used in the experiments in the paper. To run the models for different seeds change the --repl_n argument. In the paper, we have used --repl_n 1, 2, and 3 as three seeds for all the experiments.
Upon completion of all the above runs, the results will be stored in ./results directory.
After running the models, to get the out-of-distribution results from the paper, please run the following lines of code.
Black box attacks
cd ./Evaluation
python3 Black-box.py --dataset 'mnist'
White box attacks
cd ./Evaluation
python3 White-box.py --dataset 'mnist'
Rotation on MNIST
cd ./Evaluation
python3 Rotation_mnist.py --dataset 'mnist'
Running the above codes will save the output in the folder ./results/figures/mnist/ under the sub-folders PGD, Whitebox, and Rotation respectively.