remove_spectral_norm

→Module

remove_spectral_norm(module: Module, name: str = 'weight')

source edit

Reverse spectral_norm and restore a plain leaf parameter.

Copies module.<name>_orig back into a fresh leaf parameter module.<name>, removes the <name>_u / <name>_v power- iteration buffers, and detaches the forward pre-hook. Used before exporting for inference when the normalisation overhead is no longer wanted, or before fine-tuning without the spectral cap.

Parameters

moduleModule

Module previously passed through spectral_norm.

namestr= 'weight'

Attribute name of the parameter to restore. Default "weight".

Returns

Module

The same module, with <name> as a plain lucid.nn.parameter.Parameter carrying the unnormalised weight that was being trained behind the parametrisation.

Raises

ValueError

If no spectral-norm registration exists on <name>.

Notes

The restored parameter is $\mathbf{W}_{\text{orig}}$ — not the rescaled $\mathbf{W}_{\text{orig}} / \sigma$ that was visible during training. If you want the normalised matrix in the final checkpoint, copy it out before calling this function.

Examples

>>> import lucid.nn as nn
>>> from lucid.nn.utils import spectral_norm, remove_spectral_norm
>>> disc = spectral_norm(nn.Linear(16, 1))
>>> _ = remove_spectral_norm(disc)
>>> sorted(name for name, _ in disc.named_parameters())
['bias', 'weight']

Used by 1

lucid.nn.utils

>>> import lucid.nn as nn >>> from lucid.nn.utils import spectral_norm, remove_spectral_norm >>> disc = spectral_norm(nn.Linear(16, 1)) >>> _ = remove_spectral_norm(disc) >>> sorted(name for name, _ in disc.named_parameters()) ['bias', 'weight']