tₖᶠ

オレゴンで Theoretical Neuroscience やってるポスドク。

[tₖᶠ @tkf]

@apaszke @DougalMaclaurin @akatzzzzz @DavidDuvenaud @xnningxie @_ddjohnson I see, thanks, that makes sense. I'm likely ignoring a lot of subtleties as I'm thinking this in Julia which is dynamically typed.

[tₖᶠ @tkf]

@DougalMaclaurin @akatzzzzz @DavidDuvenaud @apaszke @xnningxie @_ddjohnson But since the Dex paper didn't discuss this, I wonder if I have some wrong intuition about it.

A bit of context: I've tried something similar in the "untyped" world https://t.co/A9aOidgObU

[tₖᶠ @tkf]

@DougalMaclaurin @akatzzzzz @DavidDuvenaud @apaszke @xnningxie @_ddjohnson Naively thinking, if a monoid can safely "mutate" (using the linear type) one of its arguments, wouldn't it be possible to express work-efficient programs? I'm no functional language researcher and I don't know if something like this has been done.

[tₖᶠ @tkf]

@DougalMaclaurin @akatzzzzz @DavidDuvenaud @apaszke @xnningxie @_ddjohnson By "implementing scan-reduce fusion" I mean a Dex user writing it manually in the pointful style. (Not the compiler doing it.)

[tₖᶠ @tkf]

@DougalMaclaurin @akatzzzzz @DavidDuvenaud @apaszke @xnningxie @_ddjohnson ... But implementing scan-reduce fusion correctly is rather tricky (you'd need to reinvent the semidirect product). Is there a plan for preventing "misuse" of parallelizable write+read effect?

[tₖᶠ @tkf]

@DougalMaclaurin @akatzzzzz @DavidDuvenaud @apaszke @xnningxie @_ddjohnson Thanks! As a big fan of monoid, learning why Dex puts "one more layer" (= effect) on top of was a nice brain teaser when reading the paper.

[tₖᶠ @tkf]

@DougalMaclaurin @akatzzzzz @DavidDuvenaud @apaszke @xnningxie @_ddjohnson So, I'm looking forward to learning what Dex will do with the read+write effect!

[tₖᶠ @tkf]

@DougalMaclaurin @akatzzzzz @DavidDuvenaud @apaszke @xnningxie @_ddjohnson But it's interesting that you are planning to add read to the effect for parallel reduction.

I'm interested in composing clever monoids for parallel reductions (e.g., here's my quick note on semidirect product https://t.co/R2ynisbUNR)

[tₖᶠ @tkf]

ちなみに IPython の master でやらないと一瞬で segfault するはずｗ

[tₖᶠ @tkf]

IPython の中で Julia の補完できるようになった! (楽しい) https://t.co/pbFWpuqfip

[tₖᶠ @tkf]

今日のJulia活ですhttps://t.co/V6EsU4x6W3 さっきマージされたPRがCIを壊してしまったので震えながら書いてた (誇張)

[tₖᶠ @tkf]

@ceptree はい...

[tₖᶠ @tkf]

1.0 でスコープルールが厳しくなったのでREPLでコード書くのが大変になったという話でめっちゃ揉めてる https://t.co/BSpxpHeLDc

[tₖᶠ @tkf]

マージされたぜっ https://t.co/n9WvxxtCEj 1.0.x にもバックポートされるらしい。

[tₖᶠ @tkf]

Julia 1.0 はスーパー最高な言語なんですが、メソッドの型指定があいまいでメソッドが呼べなくなる組み合わせがあったりするのが今の所ちょっと残念です...

A = Diagonal(ones(3))
B = ones(3, 3)
Y = similar(B)
mul!(Y, A, B')

とかエラーになってしまう... (ので送ったPRがさっきのやつ)

[tₖᶠ @tkf]

今日のJulia活です https://t.co/s9OGjfeTE1