peal.sparse_dictionaries.batch_topk¶
BatchTopK sparse autoencoder as a PEAL sparse dictionary.
Sparse dictionaries decompose a generator’s semantic latent (or a classifier’s
feature space) into interpretable directions that adaptors such as DiDAE edit
along. This module wraps PEAL’s own BatchTopK network from
peal.sparse_dictionaries.batch_topk_network behind the SparseDictionary
interface: it centres activations with a stored mean mu, trains with
train_batch_topk_saes or a local loop with TensorBoard logging, and
saves/loads the encoder/decoder weights as a plain torch checkpoint.
Classes
Sparse dictionary backed by the vendored BatchTopK SAE. |
|
Config of |