Julia 0.7 update #205

lostella · 2018-07-07T07:54:00Z

Addressing #191. This is an updated re-issuing of #198, taking into account all recent contributions to IterativeSolvers and changes to Julia 0.7. It relies on this PR for LinearMaps, so CI will fail before that is merged.

Open issues:

3 occurrences of spdiagm (1 in benchmark, 2 in test) still need be updated and raise warnings as of now (I'm looking into this now).
I temporarily commented "Dampened tests" for LSMR, apparently the utility code for DampenedVector and DampenedMatrix needs to be extended to have this test working; I'm happy to fix that with some help from the authors.
Some tests (lobpcg.jl and stationary.jl) worked after changing the rng seed, otherwise some assertions complained about tolerances not being met (by a small margin) or matrices not being positive semi-definite.
- stationary.jl
- lobpcg.jl

I'd like to get inputs especially on the latter two items, as well as on the rest of the code of course.
For the rest, this PR should be merged as soon as this PR for LinearMaps is.

Edit: since this PR is 0.6-incompatible, I would also immediately

Update iterables interface.

lostella · 2018-07-07T08:24:45Z

This is also the place where to get rid of #201, but after a quick check IterativeSolvers doesn't seem to be needing opnorm. Am I missing something?

dkarrasch · 2018-07-07T14:53:29Z

Regarding the norm issue, a quick check indeed seems to say that only norms of vectors are computed, which is to be expected in the context of iterative solvers. At times, however, this is done via vecnorm (I checked the current version, not the PR), which is deprecated in favor of norm.

mohamed82008 · 2018-07-09T00:27:03Z

Tolerances for lobpcg were a bit funny before, often working on my machine but not on Travis, so I imagine the residual norm is sensitive to the machine code and possibly also the seed. I will take a closer look. Thanks for this btw.

lostella · 2018-07-09T09:07:06Z

I added somewhat minimal changes to make the package compliant with the new iteration protocol. These can easily be extended by completely getting rid of start and done.

Further restructuring & uniformization of the iterators I would defer to a later stage.

haampie · 2018-07-09T09:53:01Z

Thanks, this looks great, I'll look into that issue with stationary.jl. The DampenedVector stuff is imho a bit too much for a test, but let's see if it can be upgraded easily as well.

Edit, w.r.t stationary.jl: there is no stopping criterion based on the residual or something, so the test was a bit shaky. Current fix LGTM.

haampie · 2018-07-09T11:09:00Z

Hm, I think I'll just simplify the DampenedVector bit, cause as far as I see the only special thing is the DampenedMatrix.

… that is not in LinearAlgebra for dense matrices.

lostella · 2018-07-09T15:59:28Z

Yeah it definitely looks like there's something fishy with test tolerances in lobpcg.jl.

lostella · 2018-07-09T16:21:17Z

src/lobpcg.jl

@@ -869,9 +872,9 @@ function lobpcg!(iterator::LOBPCGIterator; log=false, tol=nothing, maxiter=200,
    end
    n = size(X, 1)
    sizeX = size(X, 2)
-    residualTolerance = (tol isa Void) ? (eps(real(T)))^(real(T)(4)/10) : real(tol)
+    residualTolerance = (tol isa Nothing) ? (eps(real(T)))^(real(T)(4)/10) : real(tol)


Could the problem be here? @mohamed82008
Tests may fail since

(eps(real(T)))^(real(T)(4)/10) > sqrt(eps(real(T)))

but the test tolerance is set to sqrt(eps(real(T))).

Nevermind, tol is actually not a Nothing in tests: it is passed in and matches the test tolerance.

LOBPCG tests pass on my machine.

Oh I see you changed the seed.

It also runs fine on my machine (yes, with a modified seed), but apparently not on Travis.

[WIP] Tolerance increase and systematic tolerance setting

mohamed82008 · 2018-07-10T11:02:00Z

All good now. It seems like LOBPCG works, but some other tests fail on my machine and there are other warnings. I will try to fix the warnings.

lostella · 2018-07-10T12:09:25Z

Yep, thank you @mohamed82008! However, I don't see any warning on macos, and Travis doesn't show any on Linux.

Coverage decrease may be due to the @test_skipped tests: these will be re-enabled as soon as the PR for LinearMaps is merged (& version-bumped?)

mohamed82008 · 2018-07-10T13:43:37Z

Yes the warnings are gone when I went to nightly, seems like it was some bug in the early v0.7 beta. Looking good!

Test Summary:    | Pass  Total
Basic operations |    4      4
Test Summary:     | Pass  Total
Orthogonalization |   24     24
Test Summary: | Pass  Total
Hessenberg    |    4      4
Test Summary:      | Pass  Total
Stationary solvers |   89     89
Test Summary:       | Pass  Total
Conjugate Gradients |   39     39
Test Summary: | Pass  Total
BiCGStab(l)   |   48     48
Test Summary: | Pass  Total
MINRES        |   36     36
Test Summary: | Pass  Total
GMRES         |   35     35
Test Summary: | Pass  Total
IDR(s)        |   26     26
Test Summary: | Pass  Total
Chebyshev     |   36     36
Test Summary:       | Pass  Total
Simple Eigensolvers |   24     24
Test Summary:                                           | Pass  Total
Locally Optimal Block Preconditioned Conjugate Gradient |  322    322
Test Summary: | Pass  Total
SVD Lanczos   |   32     32
Test Summary:         | Pass  Total
BrokenArrowBidiagonal |    8      8
Test Summary: | Pass  Total
LSQR          |   11     11
Test Summary: | Pass  Total
LSMR          |    9      9
Test Summary:      | Pass  Total
ConvergenceHistory |   17     17
   Testing IterativeSolvers tests passed

haampie · 2018-07-10T22:28:04Z

LGTM. I'll merge tomorrow and tag a new version :). Thanks very much @lostella!

clason · 2018-07-11T17:38:45Z

Out of interest: I notice that https://github.com/JuliaRegistries/Uncurated / https://github.com/JuliaLang/METADATA.jl aren't actually picking up the new version (nor 0.6.0) -- is it by design that these two versions aren't tagged and published there (yet)?

haampie · 2018-07-11T18:10:57Z

@clason thanks for reminding me; there was an issue with the tests when tagging IterativeSolvers.jl on METADATA.jl. @mohamed82008 http://juliarun-ci.s3.amazonaws.com/1789ca0b09724ec0b42d7f890223157a2e4900eb/pull_request_of_IterativeSolvers_on_julia_0_6.log this is probably resolved with the fixes of this PR? If so I'll check if I can redo the 0.6.0-tag with these changes and then publish again.

mohamed82008 · 2018-07-11T18:26:24Z

I believe this should be fixed, but Travis can be surprising in bad ways sometimes.

lostella added 2 commits July 7, 2018 09:01

updated code to support julia 0.7

09b377c

test fix

b4926b7

lostella mentioned this pull request Jul 7, 2018

Update to Julia 0.7 #198

Closed

updated spdiagm syntax

5e3f14b

lostella added 2 commits July 9, 2018 11:02

added usings to further make tests independent

d1dbf09

updated iteration protocol

74eed98

Mark tests that rely on LinearMaps temporarily as skipped

00b9467

Simplify the tests of lsmr and remove the 5-argument mul! stuff since…

e9bf96d

… that is not in LinearAlgebra for dense matrices.

haampie mentioned this pull request Jul 9, 2018

Bring back 5-arg mul! to lsmr #206

Open

lostella commented Jul 9, 2018

View reviewed changes

mohamed82008 and others added 2 commits July 10, 2018 19:55

Tolerance increase and systematic tolerance setting

5458ca8

Merge pull request #1 from mohamed82008/v0.7fixes

15ca958

[WIP] Tolerance increase and systematic tolerance setting

lostella added 2 commits July 10, 2018 14:41

updated REQUIREs

8497abb

re-enabled skipped tests

8d9ac8d

haampie merged commit 781bcef into JuliaLinearAlgebra:master Jul 11, 2018

pkofod mentioned this pull request Jul 25, 2019

norm(matrix) -> opnorm(matrix) in Julia 0.7 #201

Closed

lostella deleted the julia-0.7-update branch May 17, 2020 13:35

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Julia 0.7 update #205

Julia 0.7 update #205

lostella commented Jul 7, 2018 •

edited

Loading

lostella commented Jul 7, 2018 •

edited

Loading

dkarrasch commented Jul 7, 2018

mohamed82008 commented Jul 9, 2018

lostella commented Jul 9, 2018

haampie commented Jul 9, 2018 •

edited

Loading

haampie commented Jul 9, 2018

lostella commented Jul 9, 2018

lostella Jul 9, 2018

lostella Jul 9, 2018

mohamed82008 Jul 9, 2018

mohamed82008 Jul 9, 2018

lostella Jul 9, 2018

mohamed82008 commented Jul 10, 2018

lostella commented Jul 10, 2018

mohamed82008 commented Jul 10, 2018 •

edited

Loading

haampie commented Jul 10, 2018 •

edited

Loading

clason commented Jul 11, 2018 •

edited

Loading

haampie commented Jul 11, 2018

mohamed82008 commented Jul 11, 2018

Julia 0.7 update #205

Julia 0.7 update #205

Conversation

lostella commented Jul 7, 2018 • edited Loading

lostella commented Jul 7, 2018 • edited Loading

dkarrasch commented Jul 7, 2018

mohamed82008 commented Jul 9, 2018

lostella commented Jul 9, 2018

haampie commented Jul 9, 2018 • edited Loading

haampie commented Jul 9, 2018

lostella commented Jul 9, 2018

lostella Jul 9, 2018

Choose a reason for hiding this comment

lostella Jul 9, 2018

Choose a reason for hiding this comment

mohamed82008 Jul 9, 2018

Choose a reason for hiding this comment

mohamed82008 Jul 9, 2018

Choose a reason for hiding this comment

lostella Jul 9, 2018

Choose a reason for hiding this comment

mohamed82008 commented Jul 10, 2018

lostella commented Jul 10, 2018

mohamed82008 commented Jul 10, 2018 • edited Loading

haampie commented Jul 10, 2018 • edited Loading

clason commented Jul 11, 2018 • edited Loading

haampie commented Jul 11, 2018

mohamed82008 commented Jul 11, 2018

lostella commented Jul 7, 2018 •

edited

Loading

lostella commented Jul 7, 2018 •

edited

Loading

haampie commented Jul 9, 2018 •

edited

Loading

mohamed82008 commented Jul 10, 2018 •

edited

Loading

haampie commented Jul 10, 2018 •

edited

Loading

clason commented Jul 11, 2018 •

edited

Loading