Cassette-based automatic differentiation for the Julia language.
The code/examples in this package's git history were simply proof-of-concepts that were built on an earlier prototype of Cassette. Nothing here (besides this README) should be assumed to be representative of Capstan's planned direction/implementation.
Now that an initial version of Cassette has been released, work on a more serious implementation of Capstan has begun. Since Cassette's contextual tagging system (on which Capstan relies) depends on as-of-yet unimplemented compiler optimizations/bug fixes to obtain reasonable performance, Capstan will likely not see an initial release until several minor version bumps into the Julia 1.x release cycle.
It's likely that not all of these features will make it into an initial release, but planned features include:
- forward-, reverse-, and mixed-mode operation
- no cumbersome custom array/number types
- works even with concrete dispatch/structural type constraints
- official support for complex differentiation
- safe nested/higher-order differentiation
- API for custom perturbation/sensitivity seeding
- user-extensible scalar and tensor derivative definitions
- configurable dynamic and static execution modes
- support for both GPU and CPU
- higher-order sparsity exploitation (edge-pushing)
- per-region dynamism for subgraphs