search

TuringLang / Bijectors.jl

Implementation of normalising flows and constrained random variable transformations
https://turinglang.org/Bijectors.jl/
MIT License
200 stars 33 forks source link

Rewrite: what do we want #211

Open torfjelde opened 2 years ago

torfjelde commented 2 years ago

I'm just putting this here so that we know what to keep in mind when working on #183.

General

Cross-post of comment from @csherrer:

If there's refactoring in the works, here's my wishlist:

  • Not too Distributions-specific (though a dependency is no problem)
  • No requirement that dimensionality is constant (impossible in many cases)
  • Support for one-sided inverses (injections / surjections)
  • LADJ optional (e.g. doesn't apply for discrete measures)
  • No hard-coded type parameters. Real prevents symbolics, Int prevents static sizes
  • Callable transforms - no transform required for mapping points
  • Flattening should be optional, or maybe an intermediate form

I've also been thinking of using StrideArrays for intermediate representations, since allocations can have a lot of overhead. This is still in the pondering stage, and I guess for a lightweight dependency is a bit much. But maybe there can be an argument giving the array type? Most important IMO is not to disallow it by design.

Originally posted by @cscherrer in https://github.com/TuringLang/Bijectors.jl/issues/199#issuecomment-988028984

Handling batches/collections of inputs

The current idea is to generalize the approach taken by RowVecs and ColVecs from KernelFunctions.jl to handle more general storage, as discussed extensively in https://github.com/TuringLang/Bijectors.jl/discussions/178, by introducing a Batch and then providing general implementations which performs a naive map or something of the sort, while simultaneously allowing implementers to specialize batch-implementations for improved efficiency when it makes sense.