Ability to specify different rules based on what combinations of inputs are actually being used
维护者通常 1 天内回复
还没有人认领这个 Issue。
评估
- 难度
- 5/5
- 预计耗时
- 一周以上
- 新手友好度
- 25/100
- Issue 类型
- 功能
- 描述清晰度
- 需要澄清
- 活跃度
- 停滞
- 技术栈
- julia
调研方向
首先查看 ChainRules API 和链接的 gist，然后将提议的扩展与现有的 ChainRulesCore 设计进行比较。检查围绕 PR #521 的讨论，并确定是否应由一个实验性扩展包承载这项工作；当输入组合特定的导数规则的 API 和实现计划确定后，即视为完成。
由索引模型根据 Issue 内容生成。
描述
It would be nice to have an extension to the ChainRules API that allowed for different rules to be written and hit depending on what combination of inputs the derivative is being taken with respect to.
(or maybe similar analogy for forwards mode)
Basically partial derivative rules.
Thunking is a simple approximation to this (with it's own set of struggles)
@wsmoses requested for Enzyme, though this is also relevant to every kind of operator overloading based AD (since only tracked types etc will have derivatives taken wrt to them).
In contrast it is useless for Zygote/Diffractor as they do no kind of activity analysis etc, and transform absolutely all code that is run.
A bit of a sketch for what that API might look-like is in https://gist.github.com/oxinabox/c6ad25c468b3108f8a799bda66c147f8/
This might also be useful for partial mutation support, since it is probably completely safe to have rules for things that mutate inputs that are not "active" on the derivative path? (cf https://github.com/JuliaDiff/ChainRules.jl/pull/521)
Though as the main reason we don't do mutation is tied to Diffractor/Zygote not supporting it, that might be kinda moot, unless they got some at least some basic activity analysis.
(NB: we may not initially implement this in ChainRulesCore. It might be better to make a little experimental extension package for it first.)
- 主要语言
- Julia
- 星标
- 267
- 派生
- 66
- 平均合并
- 1 天 1 小时
- 30 天内合并 PR
- 9
环境准备
这个项目没有提供开发容器、Dockerfile 或贡献指南，环境需要你自己搭建：先看它的 README，通用步骤见我们的新手贡献指南。
从这里开始
- 先读完整个 Issue，再读项目的贡献指南。
- 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
- Fork 仓库，在一个分支上完成修改。
- 提交 Pull Request，并在描述里引用这个 Issue 编号。
JuliaDiff/ChainRulesCore.jl 的其他 Issue
-
难度 4/5 3-5 天 新手友好度 28/100
JuliaDiff/ChainRulesCore.jl#712 ·
维护者通常 1 天内回复
-
variable scope of `@thunk`可能已有人在做 @chengchingwen 于 620 天前认领。 未关闭
难度 4/5 3-5 天 新手友好度 32/100
JuliaDiff/ChainRulesCore.jl#684 ·
维护者通常 1 天内回复
-
good first issue Structural Tangent
难度 2/5 1-3 小时 新手友好度 55/100
JuliaDiff/ChainRulesCore.jl#669 ·
维护者通常 1 天内回复
-
难度 5/5 一周以上 新手友好度 20/100
JuliaDiff/ChainRulesCore.jl#668 · 6 条评论 ·
维护者通常 1 天内回复
-
breaking
难度 3/5 1-2 天 新手友好度 35/100
JuliaDiff/ChainRulesCore.jl#665 ·
维护者通常 1 天内回复
查看 JuliaDiff/ChainRulesCore.jl 的全部 Issue
相似的 Issue
-
bug
难度 2/5 1-3 小时 新手友好度 76/100
QuantumSavory/QuantumClifford.jl#853 ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 66/100
JuliaPackaging/Yggdrasil#15441 · 1 条评论 ·
维护者通常 1 天内回复
-
bug
难度 1/5 1 小时以内 新手友好度 78/100
JuliaSmoothOptimizers/Penelopt.jl#260 ·
维护者通常 1 天内回复
-
authority-de-dommel ready-for-agent
难度 1/5 1 小时以内 新手友好度 78/100
Deltares/Ribasim-NL#997 ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 76/100
Sienna-Platform/InfrastructureSystems.jl#675 ·
维护者通常 2 天内回复