peal.generators.ddpm_generator¶
Pixel-space DDPM generator and the ACE/FastDiME counterfactual edit path.
This module wraps the guided-diffusion UNet vendored under
peal.dependencies.ace as a PEAL EditCapableGenerator and
InvertibleGenerator: it builds model and diffusion from a
DDPMConfig, trains it with the guided-diffusion TrainLoop, and
offers (DDIM or noising) encode, decode, repaint and sampling in pixel space.
DDPM.edit produces counterfactuals by handing the model, the diffusion and
the (optionally distilled) classifier to the vendored ACE or FastDiME main
functions, sweeping their attack hyper-parameters over several attempts.
Functions
Load a PyTorch file without redundant fetches across MPI ranks. |
Classes
Pixel-space denoising diffusion model used as a PEAL generator. |
|
TODO actually implement this class properly This class defines the config of a DDPM. |