David Widmann
David Widmann
> So with 1.7 (maybe 1.6 too?), there is an issue with Zygote gradients and the expec_loglik function I'm not familiar with this package but this sounds like it could...
(Completely unrelated: You could use `invsqrtπ` instead of `/ √π` or `/ sqrtπ`)
SciML noticed the same Zygote issue it seems: https://github.com/SciML/DiffEqSensitivity.jl/pull/511#issuecomment-984434332
> Yet the linear algebra is always the fastest It seems `broadcasted` is the fastest for `N = 500` and best allocation-wise in all examples?
That's only a workaround, not a proper fix. The manifest files and package information obtained with the linked PR (which BTW introduced formatting issues and should not have been merged...
Using a relative path to the package makes the resulting Manifest a bit less dependent on the directory structure in the Github action ("../.." instead of something like "~/work/ApproximateGPs.jl/ApproximateGPs.jl/"). However,...
Hmm not sure if I understand your question correctly. This PR and the suggestions above are only concerned with the examples, the docs are built in the same standard way....
FYI the PR contains some bugs but the approach was quite nicely it seems. I used and developed this Literate setup in a bunch of personal and other packages, so...
It seems to work also for the stable docs (tagged a release yesterday for the first time since the update): https://devmotion.github.io/CalibrationErrors.jl/stable/examples/classification/
No, in CalibrationErrors I used a different implementation. As said above, this PR is broken, ie contains some bugs.