Andrei Zhabinski

As we briefly discussed in FluxML/NNlib.jl#434, there's another way to think about the problem. If I understand the motivation correctly, there are functions that behave differently during training (i.e. being...

1. Wouldn't it be just `cond(somecondition && othercondition, () -> ..., () -> ...)`? Also, if you worry about compilation time due to multiple specializations, I believe `@nospecialize` should help...

> Well that works for my simple case above, but then you have `||`, `elseif`, etc. Hm, perhaps we put different meaning into `somecondition` and `othercondition`. I assume both are...

> This variable needs to be implicit and not passed into `cond`, because otherwise we run into the same issue `within_gradient` has with tracing. I think this is where disagree...

> To get a value for active, you ultimately need to call something which returns true when not differentiating and false when differentiating. That's the point - `active` and differentiation...

Yeah, it looks like we went quite off the topic :) The whole discussion also makes me think how hard it is to design common API for very different AD...

I ended up creating my own point type with built-in indexing: ``` type IndexedPoint2D

Great, thank you!

Any news on this issue? I'm planning to release library that uses Delaunay triangulation, so it will be nice to know status of the bug and either fix it or...

I already have kind of a fast fix for this, so I will go with it for now and just warn users until this issue is resolved normally. Thanks for...