Tools for working with linear and linearized models.

class trickster.linear.LinearGridHeuristic(problem_ctx, cache_grad=True, weight_vec=None, grid_step=1)

Snaps values of a linear heuristic to values on a regular grid.

This is useful when the transformations in the transformation graph have fixed costs.

Parameters:grid_step – Regular grid step.

Call self as a function.

class trickster.linear.LinearHeuristic(problem_ctx, cache_grad=True, weight_vec=None)

$$L_p$$ distance to the decision boundary of a binary linear classifier.

  • cache_grad – Cache the forward gradient norm.
  • weight_vec – Feature-wise weights if the $$L_p$$ space is weighted.

Call self as a function.

trickster.linear.create_reduced_linear_classifier(clf, x, transformable_feature_idxs)

Construct a reduced-dimension classifier based on the original one for a given example.

The reduced-dimension classifier should behave the same way as the original one, but operate in a smaller feature space. This is done by fixing the score of the classifier on a static part of x, and integrating it into the bias parameter of the reduced classifier.

For example, let $$x = [1, 2, 3]$$, weights of the classifier $$w = [1, 1, 1]$$ and bias term $$b = 0$$, and the only transformable feature index is 0. Then the reduced classifier has weights $$w’ = [1]$$, and the bias term incorporates the non-transformable part of $$x$$: $$b’ = -1 cdot 2 + 1 cdot 3$$.

  • clf – Original logistic regression classifier
  • x – An example
  • transformable_feature_idxs – List of features that can be changed in the given example.
trickster.linear.dist_to_decision_boundary(clf, x, target_class, target_confidence, lp_space, inv_feature_weights=None, _grad_norm=None)

Compute distance to the decision boundary of a binary linear classifier.

trickster.linear.get_forward_grad(clf, x, target_class=None)

Get the forward gradient of a classifier.

  • clf – Classifier.
  • x – Input.
  • target_class – Currently not supported.