cvxgrp / cvxpy / 2018

20%

push

travis-ci

Disciplined geometric programming (#643)

* Experimental API changes to make using reductions nicer.

* Prototype of GP support.

* Adds a notion of Disciplined Geometric Programming and log-log convexity.
* A handful of atoms now have log-log curvature, including *, +, /, power,
  maximum, sum_largest, hstack, reshape, and index; we can add more.
* Adds a reduction from DGPs to DCPs (Dgp2Dcp)
* Adds log-log atom canonicalizers for the DGP to DCP reduction.
* Adds methods to Reduction to simplify the API (in particular, reductions
  can now be used without having to pass around inverse_data).
* Adds a method to Problem that unpacks a Solution object and updates the
  problem state.
* Adds tests for DGP and for Dgp2Dcp, all of which are passing.
* Adds pretty printing to the Solution object.

TODO:
* Add tests for dual variables.
* Clean up logic related to constraints.
* Make problem.solve(gp=True) hook into Dgp2Dcp.
* Add a few more atoms (e.g,. norm, geo_mean, sqrt, cumsum, exp, log, and
  vstack+etc.)

* Add more GP atoms.

* Additions include some affine atoms (e.g. stacking) and geo_mean.
* Also clean up some log_log_convex logic.

* Introduce Equality and Inequality constraint classes.

* Adds Equality and Inequality constraint classes, which generalize
  the Zero and NonPos constraints.
* Motivation: Zero and NonPos constraints don't make sense for
  geometric programs, where the natural constraints are log-log affine == 1.
  and log-log convex <= 1.
* Equality and Inequality constraints are lowered to Zero and NonPos
  constraints by the MatrixStuffing constraint.

* Solve GPs via problem.solve(gp=True)

* With this change, problem.solve(gp=True) will parse the problem as a GP
  (and reduce it to an equivalent DCP form) before solving it.
* The keyword argument `gp` is False by default, ie, problems are assumed to
  be DCP by default.
* We add a new exception type, DGPError, which is analogous to DCP... (continued)

1480 of 1480 new or added lines in 94 files covered. (100.0%)

13526 of 16800 relevant lines covered (80.51%)

2.44 hits per line