Skip to content

Latest commit

 

History

History
274 lines (230 loc) · 8.19 KB

CHANGELOG.md

File metadata and controls

274 lines (230 loc) · 8.19 KB

Version 1.0 (targets)

  • TODO change *args to lists.
  • TODO change size to size[0]*size[1] and shape to size.
  • TODO return 2D arrays instead of matrices.
  • TODO add sets and domains for Parameters/Variables that affect DCP properties.
  • TODO PSD parameters?
  • TODO warmstart based on variable values.
  • TODO change diag for matrices to diagm so can choose which diagonal to access.
  • TODO make sure Bool.value = ? etc satisfy constraints.
  • TODO make upper_tri_to_full atom?
  • TODO parameterize upper_tri (so takes elems above kth diagonal).
  • TODO replace norm with vecnorm and norm (for matrices) (see how numpy does it).
  • TODO separate constant and linear components of expressions so potentially can cache matrix factorizations.
  • TODO add informative errors for single constraint in problem instead of list, etc.
  • TODO sort out norms. i.e. norm(X, 1) is wrong.
  • TODO support more axis operations (norm, log_sum_exp, sum_squares).
  • TODO add Lambert W function W(z) = max x, exp(-entr(x)) <= z.
  • TODO change reshape to take size tuple argument.
  • TODO make identical to numpy in every behavior (ndarrays instead of matrices etc.)
  • TODO make @ work when imported from future.

Version 0.4.9 (next release)

  • TODO make division work elementwise (and with division of scalar by variable)
  • TODO test and fix updating constraints.
  • TODO improve error message when solving with missing parameter values.
  • TODO add domains for special variables like Semidef and NonNegative (var with domain). Canonicalize then add in domains separately.
  • TODO force variable value assignments to satisfy variable properties.
  • TODO make quad_form(pos cvx, pos def pos) work.
  • TODO add cummax/cummin.
  • TODO canonicalize to QP hack (split objective into quad + PWL + constant)

Version 0.4.8

  • Fixed test with nonzero called in Python 3.

Version 0.4.7

  • Fixed bug with power of negative values outside domain.
  • Fixed bug with nonzero being removed on Python 3 by conda build.

Version 0.4.6

  • Made cumsum definition implicit and O(n).
  • Error for parameter P to quad_form.
  • Fix for Gurobi interface and Python 3.
  • Fixed bug with grad of sparse*var.
  • Added .is_pwl() # Piece-wise linear.

Version 0.4.5

  • Add residual to constraints (vector/matrix), use for violation.
  • Added cumsum.

Version 0.4.4

  • Dropped LS as default solver.

Version 0.4.3

  • Dropped dependency on CVXOPT.

Version 0.4.2

  • Fixed bug with gradient of expressions where a variable only appears in a constant term.
  • Added special solver for linearly constrained least-squares problems.

Version 0.4.1

  • Made error message for chaining constraints clearer.
  • Switched from toolz.memoize to fastcache.cru_cache to fix memory leak in Python 3.
  • Added support for matmul and rmatmul.

Version 0.4.0

  • Added domains and gradients.
  • Made curvature, sign recursive and affine == convex + concave || constant. Eliminated old DCPAttr system. Memoized important info.

Version 0.3.9

  • Fixed bug in diag with row vectors.
  • Fixed kl_div to be elementwise.

Version 0.3.8

  • Fixed bug with cvxopt solver on windows (conversion from scipy to cvxopt sparse matrix failed).

Version 0.3.7

  • Fixed bug where partial optimize didn't work with maximize objective.
  • log_sum_exp axis now works for rectangular matrices.

Version 0.3.6

  • Fixed bug with DCP attributes of partial optimize.

Version 0.3.5

  • Made to work on Windows.

Version 0.3.4

  • Added indexing with boolean ndarrays and lists of indices.

Version 0.3.3

  • Converted indices and slices to ints always.
  • Adding axis argument to *_entries atoms.

Version 0.3.2

  • Requires CVXcanon 0.0.21

Version 0.3.1

  • Fixed Gurobi interface.
  • Removed failing test that expected CVXOPT solver failure.
  • Added logistic function.

Version 0.3.0

  • Added CVXcanon integration.

Version 0.2.28

  • Fixed bug with partial_optimize not working for nonlinear constraints.

Version 0.2.27

  • Put partial_optimize in cvxpy namespace.

Version 0.2.26

  • Fixed bug with printing slices.
  • Added MOSEK SOCP interface thanks to Enzo.
  • Fixed bug with mul_elemwise by scalar.
  • Fixed compatibility issue with power atom and new NumPy.
  • Fixed Windows compatibility issue.

Version 0.2.25

  • Added NonNegative variables.
  • Added partial_optimize as first transform.
  • Require ECOS 2.0.

Version 0.2.24

  • Added Elemental interface.
  • Can add problems and objectives.
  • Basic remove redundant rows for CVXOPT chol. Made 'chol' default kktsolver.
  • Added symmetric variables and positive definite inequalities (<<, >>).
  • Added bmat atom for making block matrices.
  • Added kron.

Version 0.2.23

  • Made to work with SCS 1.1.3.

Version 0.2.22

  • Fixed issue where using "is" instead of "==".
  • Required SCS version 1.0.7 so 1.1.0 can be updated on pip.

Version 0.2.21

  • Made operator overloading work with scipy sparse matrices (with scipy 0.15).
  • Removed Expression shape function.
  • Removed Expression array function.
  • Caught c.T*x where c is a NumPy 1D array.
  • Added power.

Version 0.2.20

  • sum no longer crashes on scalar expressions, though you shouldn't use it.
  • Added mixed integer and SOCP support to Gurobi.
  • Added geo_mean

Version 0.2.19

  • Fixed Python 3 runtime error where modifying the cvxopt.solvers.options dictionary while reading it.

Version 0.2.18

  • Requires ECOS 1.1.1.
  • Removed norm_largest.
  • Added GLPK interface.
  • Updated ECOS BB exit codes.
  • Factored out ability to get parameterized entries of matrix.
  • Added GLPK_MIP interface.

Version 0.2.17

  • Optimizes rmul to mul.
  • Added warmstart.
  • Disabled MIP tests until nondeterminism is resolved.
  • Added sum_largest, sum_smallest, norm_largest, lambda_sum_largest, lambda_sum_largest, and lambda_sum_smallest.

Version 0.2.16

  • Added log1p.
  • Added scalene penalty.
  • Made LinOp for multiplication on the right.
  • TODO break up DCPAttr so only use atom DCP function.
  • Added boolean and integer variables.
  • Made geo_mean elementwise.
  • Python 3 support!!!
  • Error checking for x**p.
  • Added upper_tri.
  • Changed BoolVar, IntVar to Bool, Int.
  • Simplified Huber loss graph_implementation.

Version 0.2.15

  • Made it so you can assign a value to a variable for an initial value.
  • Overloaded power operator.

Version 0.2.14

  • Added unpack_results function to update problem state given results from a solver.
  • Fixed bug where parameters were cached.

Version 0.2.13

  • Got rid of NumPy 1D array warning.
  • Made vstack and hstack not create new variables.
  • Replaced memoize with lazyprop in constraints.
  • Refactored code for caching problem parsing.

Version 0.2.12

  • Changed cvxpy to use new SCS interface.

Version 0.2.11

  • Renamed semidefinite to Semidef.
  • Switched solver_specific_opts to **kwargs.
  • Added vec and documentation for reshape and vec.
  • Changed repr to print names in line with class names.
  • Changed str for Problem.
  • Made < and > map to <= and >=.
  • Added warning for NumPy 1D arrays.

Version 0.2.10

  • CVXPY throws an error when a solver error is encountered.
  • Presolver removes constraints with no variables.
  • Conversion from non-linear constraints to linear constraints is cached. This fixes issue #120.2.

Version 0.2.9

  • Fixed bug with sign multiplication.
  • Added check that objective is Minimize/Maximize.
  • Fixed bug with key error when solving exponential cone problem with CVXOPT.

Version 0.2.8

  • Removed a stray println.

Version 0.2.7

  • CVXPY import can succeed even if SCS import fails.
  • The sign of vector and matrix constant is positive (negative) if all the entries are positive (negative), instead of always being unknown.
  • Can now use negative indices.
  • Added *_INACCURATE return codes.

Version 0.2.6

  • Made all lin_to_matrix functions return SciPy sparse matrix or NumPy matrix (instead of ndarray).