Multiprecision and Python 3.11 support for SDPA

[usamamuneeb]

Description

SDPA for Python now supports SDPA Multiprecision, a fork of SDPA-GMP. This change will test the multiprecision backend and Python 3.11 wheels in the CI pipeline.

SDPA with OpenBLAS backend does not pass LP5 test on the Windows runners, so I changed install_dependencies.sh to use the multiprecision backend on Windows and the regular (OpenBLAS) backend on Linux and macOS.

Type of change

• New feature (backwards compatible)
• New feature (breaking API changes)
• Bug fix
• Other (Documentation, CI, ...)

Contribution checklist

• Add our license to new files.
• Check that your code adheres to our coding style.
• Write unittests.
• Run the unittests and check that they’re passing.
• Run the benchmarks to make sure your change doesn’t introduce a regression.

[github-actions]

Benchmarks that have stayed the same:

   before           after         ratio
 [4f3799d6]       [5de91523]
      29.5±0s          30.0±0s     1.01  cvar_benchmark.CVaRBenchmark.time_compile_problem
      1.80±0s          1.81±0s     1.01  simple_LP_benchmarks.SimpleScalarParametrizedLPBenchmark.time_compile_problem
      24.0±0s          24.1±0s     1.00  simple_LP_benchmarks.SimpleLPBenchmark.time_compile_problem
      9.50±0s          9.53±0s     1.00  huber_regression.HuberRegression.time_compile_problem
      8.72±0s          8.74±0s     1.00  semidefinite_programming.SemidefiniteProgramming.time_compile_problem
      46.0±0s          46.1±0s     1.00  sdp_segfault_1132_benchmark.SDPSegfault1132Benchmark.time_compile_problem
      3.89±0s          3.89±0s     1.00  simple_LP_benchmarks.SimpleFullyParametrizedLPBenchmark.time_compile_problem
      7.35±0s          7.34±0s     1.00  optimal_advertising.OptimalAdvertising.time_compile_problem
      11.2±0s          11.2±0s     1.00  svm_l1_regularization.SVMWithL1Regularization.time_compile_problem

[rileyjmurray]

This is super cool!!

…

On Tue, Apr 4, 2023 at 8:14 PM Usama Muneeb ***@***.***> wrote: Description SDPA for Python now supports SDPA Multiprecision <https://github.com/sdpa-python/sdpa-multiprecision>, a fork of SDPA-GMP. This change will test the multiprecision backend and Python 3.11 wheels in the CI pipeline. SDPA with OpenBLAS backend does not pass LP5 test on the Windows runners, so I changed install_dependencies.sh to use the multiprecision backend on Windows and the regular (OpenBLAS) backend on Linux and macOS. Type of change - New feature (backwards compatible) - New feature (breaking API changes) - Bug fix - Other (Documentation, CI, ...) Contribution checklist <https://www.cvxpy.org/contributing/index.html#contribution-checklist> - Add our license to new files. - Check that your code adheres to our coding style. - Write unittests. - Run the unittests and check that they’re passing. - Run the benchmarks to make sure your change doesn’t introduce a regression. ------------------------------ You can view, comment on, or merge this pull request online at: #2097 Commit Summary - b160dd8 <b160dd8> Enable previously skipped tests - 9fa8ed1 <9fa8ed1> Use multiprecision backend for Windows tests - 21d17af <21d17af> Remove phasevalue flipping File Changes (4 files <https://github.com/cvxpy/cvxpy/pull/2097/files>) - *M* continuous_integration/install_dependencies.sh <https://github.com/cvxpy/cvxpy/pull/2097/files#diff-602fe9fc60aa156fa83df8056884a0707e6fd9edc8fe497cc61aedd077a24146> (13) - *M* cvxpy/reductions/solvers/conic_solvers/sdpa_conif.py <https://github.com/cvxpy/cvxpy/pull/2097/files#diff-edf428e85c14b14f102c1eed684b9b7ea198a0526f70fc3c29fa7c786d025d04> (31) - *M* cvxpy/tests/test_conic_solvers.py <https://github.com/cvxpy/cvxpy/pull/2097/files#diff-896acfa00647f17dac200b1cbe34230a583fd253a22286bff4481f600c6140c8> (3) - *M* cvxpy/tests/test_problem.py <https://github.com/cvxpy/cvxpy/pull/2097/files#diff-86d3f6ee4d1742fbbfb77be8a87aba79ea3462342fa658f792b28512a0140953> (2) Patch Links: - https://github.com/cvxpy/cvxpy/pull/2097.patch - https://github.com/cvxpy/cvxpy/pull/2097.diff — Reply to this email directly, view it on GitHub <#2097>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACRLIFDHVBCMSGR6BLQSNL3W7S2OZANCNFSM6AAAAAAWTMCNSQ> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[rileyjmurray]

@usamamuneeb tests are failing because of unrelated issues with COPT.

Can you please update our web documentation to mention these changes? You'll want to update the solver install guide and the material about choosing solver settings.

It would be nice to add a dedicated test that showcases the accuracy guarantees of SDPA-multiprecision. You can have a test under TestSDPA that only runs if multiprecision support is available. I don't know exactly what test problem you'd use here. It would be okay to define a new problem specifically for this purpose.

One possibility is to have a basic LP/SDP where some variable has optimal value sqrt(2). Using double-precision arithmetic, an IPM should only be able to compute a solution to accuracy 1e-8 or so. With multiprecision support, you can ask for a solution accurate to 1e-16. I don't think you can ask for a solution with accuracy better than 1e-16 since CVXPY will probably convert to double-precision somewhere in its rewriting process. (It would be good to check if that happens and to document it if so.)

[usamamuneeb]

Hi @rileyjmurray, I will go over the SDPA-GMP paper to select some problems which show the benefit of using the multiprecision variant. If I recall correctly, besides a more accurate solution, multiprecision version also has a secondary benefit in sense that it can solve some ill posed problems in SDPLIB that the regular SDPA package cannot.

As test cases, I will try to select some problems to demonstrate each of these benefits.

[rileyjmurray]

Hi @usamamuneeb. Thanks for adding the new test problem!

I think we should wrap test_lp_7 with a check that we're using SDPA-GMP, not just SDPA.

I suggest you add an enumeration like SDPA_Variants with values Standard and GMP. Then have SDPA.variant be a class property that returns Standard by default or GMP if appropriate to do so. If you make other versions of SDPA-Python in the future, then you can extend the SDPA_Variants enumeration and update the implementation of SDPA.variant accordingly.

[usamamuneeb]

Hi @rileyjmurray, just made the change. LP7 will now run only if GMP support is available.

In the next release of sdpa-python and sdpa-multiprecision, I can create an alias for sdpap.sdpacall.sdpacall.get_backend_info() to simplify it to sdpap.get_backend_info(). That might act as a suitable replacement for the enumeration.

[rileyjmurray]

Approved! @phschiele @SteveDiamond can one of you review? It's a little ambiguous as to whether this is changing the public API, but regardless it would be good to get more eyes on it.

[SteveDiamond]

LGTM

[phschiele]

@usamamuneeb Great feature, thanks a lot. LGTM