tfp.glm.convergence_criteria_small_relative_norm_weights_change

Returns Python callable which indicates fitting procedure has converged.

tfp.glm.convergence_criteria_small_relative_norm_weights_change(
    tolerance=1e-05,
    norm_order=2
)

Defined in python/glm/fisher_scoring.py.

Writing old, new model_coefficients as w0, w1, this function defines convergence as,

relative_euclidean_norm = (tf.norm(w0 - w1, ord=2, axis=-1) /
                           (1. + tf.norm(w0, ord=2, axis=-1)))
reduce_all(relative_euclidean_norm < tolerance)

where tf.norm(x, ord=2) denotes the Euclidean norm of x.

Args:

• tolerance: float-like Tensor indicating convergence, i.e., when max relative Euclidean norm weights difference < tolerance. Default value:1e-5`.
• norm_order: Order of the norm. Default value: 2 (i.e., "Euclidean norm".)

Returns:

• convergence_criteria_fn: Python callable which returns bool Tensor indicated fitting procedure has converged. (See inner function specification for argument signature.) Default value: 1e-5.