eegproc.model_explainability.model_agnostic.optimizer module

Gradient counterfactuals over the small CounterfactualAdapter contract.

class eegproc.model_explainability.model_agnostic.optimizer.ModelAgnosticCounterfactualOptimizer(adapter, *, target_probability=0.8, learning_rate=0.01, max_steps=200, gradient_clip_norm=5.0, target_weight=1.0, state_weight=None, signal_weight=None, constraint_weights=None, report_constraints=(), stop_on_success=False)

Bases: object

Optimize adapter state while leaving all model parameters unchanged.

Parameters:

• adapter (CounterfactualAdapter)

• target_probability (float, default: 0.8)

• learning_rate (float, default: 0.01)

• max_steps (int, default: 200)

• gradient_clip_norm (float | None, default: 5.0)

• target_weight (float, default: 1.0)

• state_weight (float | None, default: None)

• signal_weight (float | None, default: None)

• constraint_weights (Mapping[str, float] | None, default: None)

• report_constraints (Iterable[str], default: ())

• stop_on_success (bool, default: False)

optimize(inputs, *, target_class=None, progress=None)

Parameters:

• target_class (int | None, default: None)

• progress (Callable[[dict], None] | None, default: None)

Return type:

dict