Patrick Aschermayr

Thanks for your answer! I used the log_likelihood function in another function and the compiler wasn't able to correctly guess the output - so I thought you might get some...

Thank you, that sounds good!

Worked for me - I just added a few comments but none of them should influence the merge. Thank you for your work!

> My 2 cents: > 3. Would it be possible to introduce a `strict=false` keyword that certains types of performance code could use to enforce symmetry/stability in both `flatten` and...

@willtebbutt : I think I managed to implement a method that allows us to keep initial types and lets us work with AD. I uploaded a version here: https://github.com/paschermayr/Shared-Code/blob/master/parameterhandling.jl I...

@willtebbutt Not urgent at all, happy to work with what I have. Thank you in any case!

Since this PR started, I have created another package, because my needs were slightly different than the ParameterHandling.jl case, https://github.com/paschermayr/ModelWrappers.jl . I managed to incorporate all possible cases (Flatten/unflatten performant...

> ModelWrappers.jl looks great, but does it provide the same functionality? E.g. `ParameterHandling.positive_definite` is something that I use a lot. IMHO this PR should still be merged. I wonder what...

@oschulz : There are 2 issues that come to my mind: 1) How do you handle the unflatten part? If AD is in mind, then the output type is determined...

> An approach like x_flat, reconstruct_function = flatten(x) didn't work out performance/type-stability wise I guess you could make methods in case `reconstruct_function` is not needed.