PowerSumRegularizer

class PowerSumRegularizer(*, weight: float = 1.0, apply_only_once: bool = False, dim: int | None = -1, normalize: bool = False, p: float = 2.0, **kwargs)

Bases: Regularizer

A simple x^p based regularizer.

Has some nice properties, cf. e.g. https://github.com/pytorch/pytorch/issues/28119.

Initialize the regularizer.

Parameters:

• weight (Tensor) – The relative weight of the regularization

• apply_only_once (bool) – Should the regularization be applied more than once after reset?

• dim (int | None) – the dimension along which to calculate the Lp norm, cf. powersum_norm()

• normalize (bool) – whether to normalize the norm by the dimension, cf. powersum_norm()

• p (float) – the parameter \(p\) of the Lp norm, cf. powersum_norm()

• kwargs – additional keyword-based parameters passed to Regularizer.__init__()

Methods Summary

forward(x)

Compute the regularization term for one tensor.

Methods Documentation

forward(x: Tensor) → Tensor

Compute the regularization term for one tensor.

Parameters:

x (Tensor)

Return type:

Tensor