abopt (ABstract OPTimizer) - optimization of generic numerical models
The package contains two components:
- optimize: L-BFGS, TrustRegion, and a bunch of simpler optimizer like gradient descent.
- model: vmad (Virtual Machine Automated Differentiation), a differentiable state machine for forward modelling, moved to https://github.com/rainwoodman/vmad
This is the second iteration of the design. The current main interface is in abopt.abopt2.
The main difference between abopt and scipy's algorithm is that the inner product
and linear operators are supplied via a vectorspace object. The reason for
this is because on a distributed problem the inner product must do a global
reduction.
The usage involves defining a Problem, then use an optimizer to minimize it.
The test suite are a good source of examples.
Things have been put together in more or less of a haste. I have a feeling we may need a restructure at some point; the current way of dealing with meta-parameters (e.g. trust region radius) is mutable and thus awkward.