Section sampler #279
Section sampler #279
Conversation
|
Hard to tell if that's a related CI failure from a quick glance. |
|
Not sure how to solve this: the point of failure of the tests at Significantly, the issue seems to reproduce on the Having a deeper look to verify whether Zygote v0.6.12 tests pass on my machine. |
|
This looks like it was caused by FluxML/Zygote.jl#962. That makes the gradient of getindex a special immutable type, which probably should not be saved in So my guess is that the fix is something like this: |
|
Yeah the user probably shouldn't be exposed to that type. That seems leaky. |
|
Yes, I mean as a quick check! If this fixes it, then it's a Zygote bug, and this should become a PR, with tests... |
|
Added the workaround suggested by Michael Abbott
to The |
Codecov Report
@@ Coverage Diff @@
## master #279 +/- ##
==========================================
+ Coverage 80.61% 80.79% +0.17%
==========================================
Files 19 19
Lines 2229 2301 +72
==========================================
+ Hits 1797 1859 +62
- Misses 432 442 +10
Continue to review full report at Codecov.
|
|
👍 upstream it to Zygote. |
|
Flux.jl was pushed 6 days ago to v0.12.6: I hope we can make it with the merge :-) |
|
Just eliminated the stale "workaround" that was still hanging in the code... |
|
It looks like the coverage went down 0.09%, which is preventing one last check from being passed. Looking into it.... |
|
Simplified a return statement. Also tried compressing comments - which should not count in coverage measurements anyway... |
Simplified return statement
|
All checks passed! :-) |
| @@ -34,6 +34,12 @@ LowDiscrepancySample{T} | |||
| sample(n,lb,ub,S::LowDiscrepancySample) | |||
| ``` | |||
|
|
|||
| * Sample on section | |||
| ```@docs | |||
| SectionSample | |||
There was a problem hiding this comment.
this doesn't have a docstring?
| @@ -1711,3 +1711,82 @@ end | |||
| end | |||
| return pareto_set,pareto_front | |||
| end | |||
|
|
|||
|
|
|||
| function surrogate_optimize(obj::Function,::EI,lb,ub,krig,sample_type::SectionSample;maxiters=100,num_new_samples=100) | |||
There was a problem hiding this comment.
This specialization may need docs.
|
Nice! Follow up with a bit more docs? |
In surrogate-based identification and control,
optimization can alternate between unconstrained sampling
in the full-dimensional parameter space,
and sampling constrained on specific sections (e.g. a planes in a 3D volume),
A
SectionSamplerallows sampling and optimizingon a subset of 'free' dimensions while keeping 'fixed' ones constrained.
The sampler is defined as in e.g.
where the first argument is a
Vector{T}in which numbers are fixed coordinates
and
NaNs correspond to free dimensions,and the second argument is a
SamplingAlgorithmwhich is used to sample in the free dimensions.