peal.sparse_dictionaries.splice_decomposition.SpLICEDecompositionConfig¶

class peal.sparse_dictionaries.splice_decomposition.SpLICEDecompositionConfig[source]¶

Configuration for SpLICE-based sparse decomposition.

Uses a pretrained CLIP model and SpLICE vocabulary to decompose image embeddings into sparse concept weights. No training required.

Parameters:

clip_model (str) – SpLICE model name of the CLIP backbone ("clip:ViT-L/14"); it must be the encoder the generator conditions on.
vocabulary (str) – Name of the SpLICE concept vocabulary ("laion").
vocabulary_size (int) – Number of vocabulary entries to keep; -1 uses the full vocabulary.
l1_penalty (float) – LASSO sparsity penalty of the SpLICE solver.
solver (str) – "skl" (scikit-learn) or "admm".
ending (str) – File extension of the saved state.
n_components (int or None) – Set from the vocabulary size once the SpLICE model is loaded.
component_strings (list or None) – Concept names DiDAE should target when editing.
opposite_component_strings (list or None) – Opposite concepts used for DDPM-inversion benchmarking.
benchmark_ddpm_inversion (bool) – If True the explainer edits with text prompts instead of z_sem.

Methods

Attributes