peal.sparse_dictionaries

Sparse dictionaries: concept bases over an encoder’s activation space.

A SparseDictionary (interfaces) is fitted on the activations of a feature extractor (the diffusion autoencoder’s semantic encoder, CLIP, DINO, …) and exposes its atoms through get_components so that DiDAE can sweep, rank and edit along them. Implementations: probe SAEs, BatchTopK, the MSAE (Matryoshka) and RA-SAE decompositions of published checkpoints, SpLiCE, plain and SAE-filtered SVD and the supervised orthogonal Procrustes dictionary. sparse_dictionary_factory.get_sparse_dictionary builds one from its config, activation_cache memoises extracted activations, sae_evaluation matches atoms to ground-truth attributes and utils holds shared plotting helpers.

Submodules

activation_cache	Optional on-disk cache for encoder activations.
batch_topk	BatchTopK sparse autoencoder as a PEAL sparse dictionary.
batch_topk_network	PEAL's own batch top-K sparse autoencoder, its trainer and an activation source.
interfaces	Base config and interface shared by every PEAL sparse dictionary.
msae_decomposition	Wrapper around the MSAE (Matryoshka Sparse Autoencoder) release for PEAL.
orthogonal_procrustes_dictionary	Supervised concept dictionary fitted by orthogonal Procrustes analysis.
probe_sparse_autoencoder	Sparse dictionary made of the weight rows of a pretrained linear probe.
ra_sae_decomposition	Pretrained RA-SAE (Relaxed Archetypal SAE) dictionaries of DINOv2 token space.
sae_evaluation	Ground-truth evaluation of sparse dictionaries (SAE / MSAE / SVD-filtered).
singular_value_decomposition	Truncated SVD (PCA) as the simplest PEAL sparse dictionary.
sparse_dictionary_factory	Factory that builds a SparseDictionary from a config, path or instance.
splice_decomposition	SpLICE concept dictionary over CLIP image embeddings.
svd_filtered_sae	SVD-prefiltered sparse autoencoders.
utils	Plotting helpers for evaluating sparse dictionaries against known attributes.