invlink for simplex distributions

[turinglang]

invlink for simplex distributions

[mohamed82008]

Done, #20. The main improvement came from replacing log(1/x) with -log(x) thus eliminating an unnecessary division. New benchmarks:

julia> @btime invlink_simplex($y);
39.384 μs (1 allocation: 7.94 KiB)

julia> @btime invlink($d, $y);
39.020 μs (1 allocation: 7.94 KiB)                                                                                                                                                                                                                                                                                          

[cpfiffer]

It looks to me like it's stable, but Mohamed the Master of Type Stability probably has a far better eye for that than me.

[mohamed82008]

Hi @trappmartin !

Your implementation looks type stable but I think we will run into numerical issues when inverting this. The main numerical stability problems were from link, invlink was just modified accordingly to be its inverse. I do think however there is a performance issue in Bijectors. I looked into it, and it seems the @debug sentences are causing some type instability. When removing them, and benchmarking I get:

julia> @btime invlink_simplex($y);
39.384 μs (1 allocation: 7.94 KiB)

julia> @btime invlink($d, $y);
43.395 μs (1 allocation: 7.94 KiB)

I assume the extra time is from the extra epsilon arithmetic that is done. The real difference is probably more when you use @inbounds. So I guess what it boils down to is if we are willing to throw away the inverse property between link and invlink, then we can gain some extra performance with your implementation of invlink.

[mohamed82008]

I will also make a PR to fix the @debug issues.

[trappmartin]

Sounds good. I wasn’t aiming for performance with my implementation but was surprised that the invlink is so much slower.

[trappmartin]

Maybe as a side note, it might be good to have internal functions that do not require to pass a distribution object. This would allow a knowledgeable user to use the transformations without instantiating a Distributions object. Similar to the StatsFuns package.

[trappmartin]

One last note. I think the invlink should be at least as fast as my code as I don’t optimise anything and invlink looks rather tuned. Your test shows that even after removing the @debug the invlink is still slower. I guess we should think about improving the code as this gets called frequently during sampling.

[mohamed82008]

@trappmartin The slowdown is from the epsilons to make invlink a proper inverse of the numerically stable link. Are you proposing making invlink as fast as possible even if it is not a good inverse of the numerically stable link?

[trappmartin]

No, I propose we try to get to code faster while keeping it numerical stable.