Torchsde Versions Save

Two-and-a-half-years since the last release. How time flies. This is primarily a maintenance release.

Changes since v0.2.4

• @akx, @vladmandic, and @jphan32 fixed up the dependency list and fixed the CI. This resolves the PyPI errors that folks have been seeing. Credit to them for making this release possible.
• Added an SDE-GAN example, as introduced in Neural SDEs as Infinite-Dimensional GANs, Kidger et al. 2021.
• Added the reversible Heun method, as introduced in Efficient and Accurate Gradients for Neural SDEs, Kidger et al. 2021.

See also

Note that Diffrax now comes highly recommended. This is a JAX-based suite of numerical ODE+SDE solvers. For most applications it is more advanced than torchsde. (In particular, it is often several times faster.) Whilst torchsde will continue to exist, new projects should consider this alternative.

New Contributors

• @dhirschfeld made their first contribution in https://github.com/google-research/torchsde/pull/99
• @jphan32 made their first contribution in https://github.com/google-research/torchsde/pull/133
• @vladmandic made their first contribution in https://github.com/google-research/torchsde/pull/137
• @akx made their first contribution in https://github.com/google-research/torchsde/pull/140

Full Changelog: https://github.com/google-research/torchsde/compare/v0.2.4...v0.2.6

Efficiency improvements:

• add f_and_g and f_and_g_prod functions so that drift and diffusion can be computed together

Bug fixes:

• calling solver.integrate twice internally
• adaptive error estimation not recognizing singleton tensors

New features include

• BrownianInterval: A Brownian motion data structure that has constant memory storage and exact queries relying on LRU caches
• Basic solvers for Stratonovich SDEs
• Full adjoint support for Ito and Stratonovich SDEs for all noise types declared in the codebase
• Various Python performance enhancements

• Add new Brownian motion classes with faster query speed based on PyTorch C++ API.
  • The new Brownian motion classes have the same API as existing ones, so they serve as direct replacements.
  • Importing these classes is as simple as from torchsde.brownian_lib import BrownianPath, BrownianTree.
  • The old Brownian motion classes written in pure Python are not yet deprecated, and likely won't be deprecated in the near future.
• Add type hints for functions of the public API.

This codebase provides stochastic differential equation (SDE) solvers with GPU support and efficient sensitivity analysis.