kkthn.projection

class kkthn.projection.ProjectionSettings(fb_eps: 'float' = 1e-08, gn_max_iters: 'int' = 30, gn_tol: 'float' = 1e-06, gn_reg: 'float' = 0.001, newton_step_length: 'float' = 0.5, armijo_alpha: 'float' = 0.0001, armijo_beta: 'float' = 0.5, max_backtrack_iter: 'int' = 10, armijo_max_steps: 'int' = 10, backward_reg: 'float' = 1e-08)

Bases: object

fb_eps: float = 1e-08

gn_max_iters: int = 30

gn_tol: float = 1e-06

gn_reg: float = 0.001

newton_step_length: float = 0.5

armijo_alpha: float = 0.0001

armijo_beta: float = 0.5

max_backtrack_iter: int = 10

armijo_max_steps: int = 10

backward_reg: float = 1e-08

class kkthn.projection.ProjectionLayer(project: 'Callable', project_single: 'Callable', solve_single: 'Callable', eq_constraints: 'Callable', ineq_constraints: 'Callable', dims: 'dict[str, int]', settings: 'ProjectionSettings')

Bases: object

project: Callable

project_single: Callable

solve_single: Callable

eq_constraints: Callable

ineq_constraints: Callable

dims: dict[str, int]

settings: ProjectionSettings

kkthn.projection.make_projection_layer(model, *, param_name: str = 'x', settings: ProjectionSettings | None = None) → ProjectionLayer

Create the KKT-HardNet projection layer from a JaxNLPModel.

The layer solves, per sample,

min_y 0.5 ||y - y_hat||^2 s.t. h(y, x) = 0, g(y, x) <= 0, lb(x) <= y <= ub(x)

Inequalities and bounds are folded into one slack/complementarity block.