https://github.com/bccp/abopt

abopt (ABstract OPTimizer) - optimization of generic numerical models
https://github.com/bccp/abopt

Last synced: 10 months ago
JSON representation

abopt (ABstract OPTimizer) - optimization of generic numerical models

• Host: GitHub
• URL: https://github.com/bccp/abopt
• Owner: bccp
• License: gpl-3.0
• Created: 2016-11-28T02:59:26.000Z (over 9 years ago)
• Default Branch: master
• Last Pushed: 2021-02-25T17:03:50.000Z (over 5 years ago)
• Last Synced: 2023-08-07T05:40:12.972Z (almost 3 years ago)
• Language: Python
• Homepage:
• Size: 381 KB
• Stars: 1
• Watchers: 4
• Forks: 1
• Open Issues: 6
• Metadata Files:
  • Readme: README.rst
  • License: LICENSE

Awesome Lists containing this project

README

          
abopt

=====

abopt (ABstract OPTimizer) - optimization of generic numerical models

.. image:: https://travis-ci.org/bccp/abopt.svg?branch=master

    :target: https://travis-ci.org/bccp/abopt

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.