Closed denizokt closed 10 years ago
Some style things and a significant issue with how mvndst_parameters is stored in the EI class.
Thanks for getting this together quickly!
One more thing... now that I read through where abseps and releps are used: http://www.math.wsu.edu/faculty/genz/software/fort77/mvndstpack.f and the default values in scipy:
Other Parameters
----------------
maxpts : input int, optional
Default: 2000
abseps : input float, optional
Default: 1e-06
releps : input float, optional
Default: 1e-06
I'm actually a bit concerned/confused. mvndst says it uses eps = MAX(abseps, finest*releps)
. 1.0e-6 is pretty large absolute error (i think). That might mean these super low releps values we're setting don't actually do anything and only maxpts matters.
It'd be worth understanding what's happening here--maybe we should be using abseps or setting abseps to 0 so that only releps matters. Otherwise we're probably being overridden by abseps.
okay i made abseps accessible as well.
now test_qd_with_self is running too slowly... the full test took 60 seconds to run with abseps=1.0e-11 and 14 for 1.0e-7 welp
1.0e-6 runs in < 3 seconds... good enough?
@suntzu86 currently i don't actually access the private variable anywhere outside the expected improvement class. i was thinking that making getters/setters would be very redundant in this case. i could make them using the property(getter, setter) format but that would just make it have two copies of the same thing.
Looks good. A few minor nits.
Also, thanks again so much for getting this together so quickly :)
@bulutcocuk fair enough, you can leave the getters/setters off then. I didn't realize all the private accesses outside the class were gone. still want the comment about what the different eps mean and how they relate to each other
@bulutcocuk i don't understand what you're asking in that post about timing results. I think your numbers don't all refer to releps...? or something?
:ship:
shipit pending travis
***** PEOPLE ***** Primary reviewer: @sc932
Reviewers: @suntzu86
***** DESCRIPTION ** Branch Name: denizokt_376_fixflakes Ticket(s)/Issue(s): Closes #376 Closes #374
edit (@suntzu86): this also closes the test performance issue in #374
***** TESTING DONE ***** make test make style-test made sure demos and pretty endpoints still work