peal.sparse_dictionaries.probe_sparse_autoencoder.ProbeSAE

class peal.sparse_dictionaries.probe_sparse_autoencoder.ProbeSAE

Bases: SparseDictionary

Dictionary whose atoms are the rows of a classifier’s last linear layer.

Parameters:

config (ProbeSAEConfig) – Points to the checkpoint; fc.weight is cloned into self.W with shape (n_classes, feature_dim), and self.mu is a zero mean vector of length feature_dim so callers treating the dictionary as centered get an identity offset.

Methods

fit	No-op: the probe weights are fixed by the checkpoint.
fit_from_dataloaders	No-op: the probe weights are fixed by the checkpoint.
get_components	Return the atoms as columns, i.e. W.t() of shape (D, n_classes).
load_from_disk	Restore self.W from a file written by save_on_disk().
save_on_disk	Save {"W": self.W} with torch.save to path.

Attributes

config

__init__(config)

Load the checkpoint on CPU and copy its fc weight matrix.

fit(X)

No-op: the probe weights are fixed by the checkpoint.

fit_from_dataloaders(dataloaders, feature_extractor)

No-op: the probe weights are fixed by the checkpoint.

get_components()

Return the atoms as columns, i.e. W.t() of shape (D, n_classes).

save_on_disk(path)

Save {"W": self.W} with torch.save to path.

load_from_disk(path)

Restore self.W from a file written by save_on_disk().