MNT: Suppress LineSearchWarning's in scipy.optimize tests #4570
Conversation
| warnings.simplefilter('ignore', RuntimeWarning) | ||
| s, fc, gc, fv, ofv, gv = ls.line_search_wolfe2(f, fprime, x, p, | ||
| g0, f0, old_f, | ||
| amax=smax) |
There was a problem hiding this comment.
I guess you could only filter LineSearchWarning (from scipy.optimize.linesearch) because plain RuntimeWarnings are not expected here IIRC.
There was a problem hiding this comment.
I thought about it after making the commit. I changed it now.
nmayorov
force-pushed
the
filter_convergence_warnings
branch
from
67d6485
to
a4a08c3
Compare
| x = optimize.fmin_bfgs(func, x0, fprime, disp=False) | ||
| with warnings.catch_warnings(): | ||
| warnings.simplefilter('ignore', LineSearchWarning) | ||
| x = optimize.fmin_bfgs(func, x0, fprime, disp=False) |
There was a problem hiding this comment.
I think the warning should not be filtered at this level, because BFGS actually uses a wrapper to line_search_wolfe2, the function raising the warning. That is _line_search_wolfe12 in optimize.py, which handles the fact that the line search algorithm may not converge. So, IMO, the warning should be filtered within the wrapper, and it would hence not appear in tests.
There was a problem hiding this comment.
So the only valid filtering in this file is for direct calls to line_search_wolfe2.
There was a problem hiding this comment.
But if we filter warnings in _line_search_wolfe12 a user won't see them when this wrapper is called. It looks like undesired behaviour, doesn't it?
I might be confused with LineSearchWarning vs _LineSearchError. The former is raised in line_search_wolfe2 when derphi_star is None, the latter is raised in _line_search_wolfe12 when alpha_star is None.
As far as I understand we still want to give a warning (but not an error), when derphi_star is None. Although currently I have no idea about algorithm details.
There was a problem hiding this comment.
Nikolay Mayorov wrote:
But if we filter warnings in |_line_search_wolfe12| a user won't see
them when this wrapper is called. It looks like undesired behaviour,
doesn't it?
No, because this wrapper handles the case of line_search_wolfe2
returning None and then raise an exception (which is then catched in the
caller). So there's no point warning the user about something they have
no control on.
I might be confused with |LineSearchWarning| vs |_LineSearchError|. The
former is raised in |line_search_wolfe2| when |derphi_star| is |None|,
the latter is raised in |_line_search_wolfe12| when |alpha_star| is |None|.As far as I understand we still want to give a warning (but not an
error), when |derphi_star| is |None|. Although currently I have no idea
about algorithm details.
I think we want a warning only when line_search_wolfe2 is called
directly by the user, to let them know about converge issue, although
the function returns something (this behavior must be kept for backwards
compatibility).
nmayorov
force-pushed
the
filter_convergence_warnings
branch
from
a4a08c3
to
267e1d8
Compare
|
@dlax, I think I understood what you were saying. I updated the commit (I think more than one is not appropriate here). |
MNT: Suppress LineSearchWarning's in scipy.optimize tests
|
Merged, thanks @nmayorov ! |
I tried to fix #4554.
Tell me if I did it correctly.