Implement FD step size adaptation #671
Conversation
Codecov Report
@@ Coverage Diff @@
## develop #671 +/- ##
===========================================
+ Coverage 88.16% 90.33% +2.16%
===========================================
Files 79 96 +17
Lines 5257 6363 +1106
===========================================
+ Hits 4635 5748 +1113
+ Misses 622 615 -7
Continue to review full report at Codecov.
|
| CONSTANT = "constant" | ||
| SPACE = "space" | ||
| STEPS = "steps" | ||
| UPDATE_CONDITIONS = [CONSTANT, SPACE, STEPS] |
There was a problem hiding this comment.
I'd rather leave these here, s.t. one has to call them pypesto.FDDelta.CONSTANT, as they are only used in this specific context.
| assert np.allclose(val, v, atol=atol, rtol=rtol), attr | ||
| # cannot completely coincide | ||
| assert (val != v).any(), attr | ||
| assert np.allclose(val, val_fd, atol=atol, rtol=rtol), attr |
There was a problem hiding this comment.
Can we maybe double check, whether we achieve better accuracies when using adaptive FDs than when using static FDs? Maybe even down to 1e-5?
There was a problem hiding this comment.
The default is 1e-6, which already gives quite a good approximation (automatically chosen are 1e-7 to 1e-8), so on this problem not much gain, rather a "automatic detection does not screw things up" check.
Step sizes adaptation for finite differences, realized via a
pypesto.objective.finite_difference.FDDeltaclass.Also, some simplifications of grad+hess+sres derivative calculations (i.e. rather interpret those as 1st or 2nd order derivatives with code reuse).
Reference: https://github.com/ICB-DCM/PESTO/blob/master/private/getStepSizeFD.m