Concepts¶
TorchEBM treats generative modeling as the composition of a small number of mathematical objects. Each has one subpackage, one base class, and one page here:
| Axis | Question it answers | Package | Page |
|---|---|---|---|
| Model | Which potential or field defines the dynamics? | torchebm.core, torchebm.models | The Energy-Based View for potentials, Learning Objectives for fields |
| Dynamics | How are samples drawn? | torchebm.samplers, torchebm.integrators | Sampling and Integration |
| Objective | How is the model fit to data, with or without sampling in the loop? | torchebm.losses | Learning Objectives |
| Transport | Which path and pairing connect the source and the data? | torchebm.interpolants, torchebm.couplings | Interpolants and Couplings |
Design and Scope states the unifying object, a learned dynamical system on data space, and the axes that place every method family; it separates simulation-free from simulation-based training and marks which families are supported and which are on the roadmap, with references.
Read in any order; each page links to the runnable examples that exercise it.