Expand description
Diffusion model training pipeline: fit from column statistics, persist, and evaluate.
The DiffusionTrainer fits a TrainedDiffusionModel from per-column statistics
(mean, std, min, max, type) and an optional correlation matrix. The trained model
can be serialized to JSON for persistence and later reloaded for generation.
Generation uses the same statistical diffusion approach as
StatisticalDiffusionBackend: start from
Gaussian noise, iteratively denoise toward the target distribution, then apply
correlation structure via Cholesky decomposition.
Structs§
- Column
Diffusion Params - Statistical parameters for a single column in the trained model.
- Diffusion
Trainer - Trainer that fits diffusion model parameters from column statistics.
- FitReport
- Report from evaluating a trained diffusion model against its target statistics.
- Model
Metadata - Metadata about a trained diffusion model.
- Trained
Diffusion Model - A trained diffusion model that can generate samples and be persisted to disk.
Enums§
- Column
Type - The type of a column in the dataset.