Currently, convergence conditions are determined by specifying the converge_when argument of the generate_counterfactual method. If this is set to :generator_conditions, then convergence is defined as "gradients approaching zero".
This keeps things "clean" in the sense that generators really only differ in terms of their objective function. In other words, specifying a different decision threshold for Wachter does not change the fact that it's still Wachter.
For usability though, it's nice to be able to specify "Wachter ($\gamma=0.5)$" and "Wachter ($\gamma=0.9)$", for example.
Currently, convergence conditions are determined by specifying the
converge_whenargument of thegenerate_counterfactualmethod. If this is set to:generator_conditions, then convergence is defined as "gradients approaching zero".This keeps things "clean" in the sense that generators really only differ in terms of their objective function. In other words, specifying a different decision threshold for Wachter does not change the fact that it's still Wachter.
For usability though, it's nice to be able to specify "Wachter ($\gamma=0.5)$ " and "Wachter ($\gamma=0.9)$ ", for example.