Example notebook: deep kernel learning #322
Conversation
| using Flux | ||
| using Distributions, LinearAlgebra | ||
| using Plots | ||
| using ProgressMeter | ||
| using AbstractGPs |
There was a problem hiding this comment.
This also introduces a circular dependency and hence leads to the same problems as mentioned in #316 (comment). It might be better to move it to AbstractGPs or the JuliaGaussianProcesses webpage. We can always link to these examples from the documentation, so it would still be possible to make it discoverable from the KernelFunctions docs.
There was a problem hiding this comment.
Now that each example has its own project environment & pinning, shouldn't this no longer be a concern? The example depends on KernelFunctions and AbstractGPs, and AbstractGPs depends on KernelFunctions, but neither KernelFunctions nor AbstractGPs depend on the example.
There was a problem hiding this comment.
AbstractGPs depends on KernelFunctions
This is exactly the problem and the circular dependency I was referring to. If we make breaking changes in KernelFunctions, the example will break - there is no version of AbstractGPs that is compatible with this version yet. Hence either we have to switch to an old version of KernelFunctions and there will be a disconnect between the version of KernelFunctions discussed in the documentation and used in the example until AbstractGPs is updated and the example is fixed or we have to remove the example for a while. Both alternatives seem a bit annoying.
This problem is the main reason why I think it is unfortunate to use downstream packages in the documentation of KernelFunctions and these examples should rather go into AbstractGPs where we would not have to deal with these problems.
There was a problem hiding this comment.
A breaking change doesn't imply the example must break, just that it's not guaranteed to not break... In any case, I'm fine with moving this notebook over to AbstractGPs. Just wanted to save the current state out of #234.
There was a problem hiding this comment.
Often the example itself won't break but this is not relevant: AbstractGPs is just not compatible with KernelFunctions before and when (and possibly for some time after) a breaking release of KernelFunctions is made. A new release of AbstractGPs is needed with updated compatibility bounds and possibly some fixes, even if the example itself does not require any changes.
Codecov Report
@@ Coverage Diff @@
## master #322 +/- ##
===========================================
- Coverage 83.21% 43.14% -40.07%
===========================================
Files 52 52
Lines 1251 1240 -11
===========================================
- Hits 1041 535 -506
- Misses 210 705 +495
Continue to review full report at Codecov.
|
No description provided.