Checking for non-descent direction in qN

[mateuszbaran]

A slightly generalized check originally proposed in #339 . Ideally we would also have an option that purges the memory but that's fairly difficult to do in Manopt. This change may be a little difficult to test.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 99.73%. Comparing base (5253f86) to head (92d47b1).

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #361      +/-   ##
==========================================
+ Coverage   99.58%   99.73%   +0.14%     
==========================================
  Files          73       73              
  Lines        6801     6805       +4     
==========================================
+ Hits         6773     6787      +14     
+ Misses         28       18      -10     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[kellertuer]

That is a very neat idea. I have to check how to best test that, maybe starting it with -grad for one test?

Also for the warning, I usually did more like message passing and / or have specific debug to catch warnings. I can think a bit what best to do here with that, maybe the warning is also ok.

[mateuszbaran]

I'm trying to add a test by defining a custom struct QuasiNewtonGradientDirectionUpdate <: AbstractQuasiNewtonDirectionUpdate end that returns gradient as the direction.

[kellertuer]

I think my only remark would be that this changes the default. (:ignore was the default before), but this could also be seen as an enhancement.

We could merge this the next few days together with the typo PR as a new version then :)

[mateuszbaran]

I think my only remark would be that this changes the default. (:ignore was the default before), but this could also be seen as an enhancement.

Yes, I think this is an enhancement. It adds a bit of additional cost for checking but I think it's worth it.

We could merge this the next few days together with the typo PR as a new version then :)

Sure, that's fine.