peal.generators

Generative models that render the counterfactuals the explainers need.

Every generator implements interfaces.Generator; invertible ones (InvertibleGenerator) add encode / decode and edit-capable ones (EditCapableGenerator) an edit that moves an image across a predictor’s decision boundary. Implementations: the DiffAE-based diffusion_autoencoder and its RAE twin, pixel / latent DDPMs (ddpm_generator, ddpm_pathldm), Stable Diffusion 3, Flux and the Stable Diffusion autoencoder, PathLDM for histopathology and a tabular DDPM. generator_factory.get_generator builds one from a GeneratorConfig yaml; deeplift_resnet provides DeepLift-safe ResNets used for attribution-guided generation.

Submodules

ddpm_generator	Pixel-space DDPM generator and the ACE/FastDiME counterfactual edit path.
deeplift_resnet	ResNet variants with one nn.ReLU instance per activation site.
generator_factory	Resolve a generator specification into a live generator object.
interfaces	Abstract interfaces every PEAL generator implements.
rae_pipeline	Two-stage RAEv2 training pipeline behind the RAEDiffusionAutoencoder generator.
tabular_ddpm	Minimal DDPM for low-dimensional tabular data with guided counterfactuals.