This repository contains code for SBFW Algorithm presented in Consistent Plugin Classifier for Complex Objectives and Constraints.
To replicate the experiments presented in the paper run
python run_experiments.pyWe cover the following loss functions:
- H-mean loss ("hmean")
- Q-mean loss ("qmean")
- G-mean Loss ("gmean")
- Linear 0-1 Loss ("linear") -- For this you may want to provide a alternate loss implying the complex loss value when linear loss is minimized.
and following constraints:
- Demographic Parity ("DP")
- Coverage ("COV")
- Equal Oppotunity ("EOpp")
- KL-divergence ("KLD")
We provide a function SBFW that can be imported from algorithm.SBFW. It takes in the following Inputs
- X_train, y_train
- X_test, y_test
- Loss Name
- Constraint Name
- Lambda (Hyperparameter - Default value works fine)
- Epsilon (Slack depends on how precisely do we want the constraints to be satisfied - Lower values take longer iterations)
- Eta_t_arr (Can be a length 1 array)
- Total iterations (T)
loss_train, loss_test, constraint_value_train, constraint_value_test = SBFW(X_train, y_train, X_test, y_test, "gmean", "EOpp", lambda_val, epsilon, eta_t_array, T)We also provide the codes for various baseline methods that we compared with in the paper. For more details about the baseline methods please refer to their respective Repositories. Three Player COCO