Mohamed Tarek

I think for Tracker ```julia sum(logpdf.(dist.f.(dist.θ), x)) ``` will be faster than ```julia sum(1:length(dist.θ)) do i logpdf(dist.f(dist.θ[i]), x[i]) end ``` So if either `θ` or `x` is a `TrackedArray`, all...

Thanks for the PR @cscherrer and sorry for the late review; I was busy the last few weeks. I will review your PR asap.

A few comments I have. 1. Doing constrained optimization by transforming the constrained variables is just one way of doing constrained optimization. There are optimization algorithms that can efficiently handle...

> I think MvNormal already does this, no? But what is the difference between this and the more general approach of allowing "lazy" transformations like what this issue is proposing?...

> You could still do this when P is, say, the actual function exp though, right? But maybe this has some issues I'm not fully aware of. Well in your...

>This puzzles me a bit Welcome to the club! It could be a Zygote bug.

Try making a MWE

Hi! Once you get it working, we can decide which code can go in here and which may need to go in Distributions.jl. We may also host some code here...

>@mohamed82008 any thoughts on how this implementation could be made less aggressive? This is a hard one. Without making Distributions.jl compatible with Zygote, we need a catchall adjoint here to...

So now when I define a new function, if I suspect ChainRules is hurting my performance/correctness, I can either define a correct and performant rule for my method or opt...