inspire-group / multiclass_robust_lb

Characterizing the Optimal 0 − 1 Loss for Multi-class Classification with a Test-time Attacker

Compute optimal loss of existing datasets

The main document to compute optimal loss of a dataset is optimal_log_loss_lp_hyper.py. It is able to compute loss on dataset MNIST, CIFAR-10, CIFAR-100 (using --dataset_in argument) and 0-1 or cross-entropy loss (using --loss argument). We can compute up to 4-way hyperedges. Users can also select specific classes of a dataset for computation (using --classes argument). More arguments and flags can be found under main().

Compute full dataset

For example, to compute optimal 0-1 loss of full CIFAR-10 dataset at $\epsilon=3$ up to degree 3 hyperedges, run the following command (substituting directory path):

python3 optimal_log_loss_lp_hyper.py --remove_redundant --out_dir OUT_DIR --data_dir DATA_DIR --compute_hyper 3 --norm l2 --dataset_in CIFAR-10 --loss 0-1 --use_all_classes --eps 3 --run_generic --run_nonconvex --num_samples 8000 --mosek.

The optimal loss along with edge information will be saved in a file under cost_result folder. Note that in order to include all data, --num_samples $\ge \max$ {sample per class}.

Compute partial dataset

To compute the optimal 0-1 loss of MNIST class 1, 4 and 7 at $\epsilon=3$ up to degree 2 hyperedges with 1000 samples per class, run

python3 optimal_log_loss_lp_hyper.py --remove_redundant --out_dir OUT_DIR --data_dir DATA_DIR --norm l2 --dataset_in MNIST --loss 0-1 --classes 1 4 7 --eps 3 --run_generic --run_nonconvex --num_samples 1000 --mosek.