Comments (13)
Hmm is this a weird kind of invalidation not because we are defining a more specific method,
but because we define the only method that doesn't error?
I think it would probably be fine for us to error in this case too.
It's a weird corner case: https://github.com/JuliaDiff/ChainRulesCore.jl/blob/main/src/tangent_types/tangent.jl#L110
from chainrulescore.jl.
This one has a lot of children and will hit a lot of packages, so it would be nice to fix it if possible. @oxinabox are you suggesting deleting the method?
from chainrulescore.jl.
Current version:
julia> import Pkg; Pkg.activate(temp=true); Pkg.add("ChainRulesCore")
julia> using SnoopCompileCore; invalidations = @snoopr(using ChainRulesCore); using SnoopCompile
julia> trees = invalidation_trees(invalidations);
julia> length(uinvalidated(invalidations))
267
julia> trees
3-element Vector{SnoopCompile.MethodInvalidations}:
inserting *(::Any, ::ZeroTangent) in ChainRulesCore at /home/hendrik/.julia/packages/ChainRulesCore/6Sl9y/src/tangent_arithmetic.jl:105 invalidated:
mt_backedges: 1: signature Tuple{typeof(*), String, Any} triggered MethodInstance for (::Test.var"#7#9")(::Any) (0 children)
2: signature Tuple{typeof(*), String, Any} triggered MethodInstance for (::Test.var"#8#10")(::Any) (0 children)
inserting convert(::Type{<:Number}, x::ChainRulesCore.NotImplemented) in ChainRulesCore at /home/hendrik/.julia/packages/ChainRulesCore/6Sl9y/src/tangent_types/notimplemented.jl:63 invalidated:
backedges: 1: superseding convert(::Type{Union{}}, x) in Base at essentials.jl:213 with MethodInstance for convert(::Core.TypeofBottom, ::Any) (13 children)
inserting tail(t::Tangent{<:NamedTuple{<:Any, <:Tuple{}}}) in ChainRulesCore at /home/hendrik/.julia/packages/ChainRulesCore/6Sl9y/src/tangent_types/tangent.jl:110 invalidated:
mt_backedges: 1: signature Tuple{typeof(Base.tail), Any} triggered MethodInstance for Base._cshp(::Int64, ::Tuple{Bool}, ::Tuple{Int64}, ::Any) (0 children)
2: signature Tuple{typeof(Base.tail), Any} triggered MethodInstance for Base._cshp(::Int64, ::Tuple{Bool}, ::Tuple{Any, Vararg{Any}}, ::Any) (0 children)
3: signature Tuple{typeof(Base.tail), Any} triggered MethodInstance for Base.Iterators._zip_isdone(::Tuple, ::Any) (0 children)
4: signature Tuple{typeof(Base.tail), Any} triggered MethodInstance for Base.Iterators._zip_iterate_some(::Tuple, ::Any, ::Tuple{Missing, Vararg{Any}}, ::Missing) (0 children)
5: signature Tuple{typeof(Base.tail), Any} triggered MethodInstance for Base.Iterators._zip_iterate_some(::Tuple, ::Any, ::Tuple{Any, Vararg{Any}}, ::Missing) (0 children)
6: signature Tuple{typeof(Base.tail), Any} triggered MethodInstance for Base.Iterators._zip_iterate_some(::Tuple, ::Any, ::Tuple{Bool, Vararg{Any}}, ::Bool) (0 children)
7: signature Tuple{typeof(Base.tail), Any} triggered MethodInstance for Base.Iterators._zip_iterate_some(::Tuple, ::Any, ::Tuple{Any, Vararg{Any}}, ::Bool) (0 children)
8: signature Tuple{typeof(Base.tail), Any} triggered MethodInstance for iterate(::Base.Iterators.Enumerate{Vector{VersionNumber}}, ::Any) (0 children)
9: signature Tuple{typeof(Base.tail), Any} triggered MethodInstance for Base.tail(::NamedTuple{names}) where names (335 children)
24 mt_cache
Removing https://github.com/JuliaDiff/ChainRulesCore.jl/blob/main/src/tangent_types/tangent.jl#L110:
julia> import Pkg; Pkg.activate(temp=true); Pkg.develop("ChainRulesCore")
julia> using SnoopCompileCore; invalidations = @snoopr(using ChainRulesCore); using SnoopCompile
julia> trees = invalidation_trees(invalidations);
julia> length(uinvalidated(invalidations))
271
julia> trees
3-element Vector{SnoopCompile.MethodInvalidations}:
inserting *(::Any, ::ZeroTangent) in ChainRulesCore at /home/hendrik/.julia/dev/ChainRulesCore/src/tangent_arithmetic.jl:105 invalidated:
mt_backedges: 1: signature Tuple{typeof(*), String, Any} triggered MethodInstance for (::Test.var"#7#9")(::Any) (0 children)
2: signature Tuple{typeof(*), String, Any} triggered MethodInstance for (::Test.var"#8#10")(::Any) (0 children)
18 mt_cache
inserting convert(::Type{<:Number}, x::ChainRulesCore.NotImplemented) in ChainRulesCore at /home/hendrik/.julia/dev/ChainRulesCore/src/tangent_types/notimplemented.jl:63 invalidated:
backedges: 1: superseding convert(::Type{Union{}}, x) in Base at essentials.jl:213 with MethodInstance for convert(::Core.TypeofBottom, ::Any) (168 children)
9 mt_cache
inserting tail(t::Tangent{<:NamedTuple{<:Any, <:Tuple{Any}}}) in ChainRulesCore at /home/hendrik/.julia/dev/ChainRulesCore/src/tangent_types/tangent.jl:109 invalidated:
mt_backedges: 1: signature Tuple{typeof(Base.tail), Any} triggered MethodInstance for Base._cshp(::Int64, ::Tuple{Bool}, ::Tuple{Int64}, ::Any) (0 children)
2: signature Tuple{typeof(Base.tail), Any} triggered MethodInstance for Base._cshp(::Int64, ::Tuple{Bool}, ::Tuple{Any, Vararg{Any}}, ::Any) (0 children)
3: signature Tuple{typeof(Base.tail), Any} triggered MethodInstance for Base.Iterators._zip_isdone(::Tuple, ::Any) (0 children)
4: signature Tuple{typeof(Base.tail), Any} triggered MethodInstance for Base.Iterators._zip_iterate_some(::Tuple, ::Any, ::Tuple{Missing, Vararg{Any}}, ::Missing) (0 children)
5: signature Tuple{typeof(Base.tail), Any} triggered MethodInstance for Base.Iterators._zip_iterate_some(::Tuple, ::Any, ::Tuple{Any, Vararg{Any}}, ::Missing) (0 children)
6: signature Tuple{typeof(Base.tail), Any} triggered MethodInstance for Base.Iterators._zip_iterate_some(::Tuple, ::Any, ::Tuple{Bool, Vararg{Any}}, ::Bool) (0 children)
7: signature Tuple{typeof(Base.tail), Any} triggered MethodInstance for Base.Iterators._zip_iterate_some(::Tuple, ::Any, ::Tuple{Any, Vararg{Any}}, ::Bool) (0 children)
8: signature Tuple{typeof(Base.tail), Any} triggered MethodInstance for iterate(::Base.Iterators.Enumerate{Vector{VersionNumber}}, ::Any) (0 children)
9: signature Tuple{typeof(Base.tail), Any} triggered MethodInstance for Base.tail(::NamedTuple{names}) where names (184 children)
So there are still many invalidations left...
The code removing https://github.com/JuliaDiff/ChainRulesCore.jl/blob/main/src/tangent_types/tangent.jl#L110 lives in https://github.com/ranocha/ChainRulesCore.jl/tree/hr/fix_invalidations. Shall I make a PR? Or does anybody have time to hunt these invalidations?
from chainrulescore.jl.
The ones on *
are not a concern.
They aren't affecting anything real that will ever matter. Though it might be nice to find out what actually causes them.
The one on convert
is a bit of a worry, I would like to know where it came from. I.e. what changed in Base (I guess?), as that wasn't there before.
Or did SnoopCompile get smarter.
Re:tail. We should also be able to delete
So the remaining ones are on line 109.
I guess this is because now there exist multiple possible return types for Base.tail(::Any)
still.
Rather than just Tuple/NamedTuple? Which I guess it was relying on via world splitting.
Since Base.tail
isn't part of the exported API, maybe we shouldn't have overloaded it in the first place.
@mcabbott you added that in #567
is it actually being used anywhere?
from chainrulescore.jl.
Base has these methods for tail
:
julia> methods(Base.tail)
# 3 methods for generic function "tail" from Base:
[1] tail(t::NamedTuple{names}) where names
@ namedtuple.jl:321
[2] tail(::Tuple{})
@ essentials.jl:330
[3] tail(x::Tuple)
@ essentials.jl:329
There's no overlap with Tangent
. Why exactly is anything invalidated?
from chainrulescore.jl.
IIRC the rule which uses it is https://github.com/JuliaDiff/ChainRules.jl/blob/39c2d17df672836659493d6adb7d4ad8593250a5/src/rulesets/Base/indexing.jl#L175-L178 , obviously that could be re-written if someone wants to.
from chainrulescore.jl.
There's no overlap with
Tangent
. Why exactly is anything invalidated?
I believe it is a so called "world splitting" optimization.
Where even if we can't infer which method of Base.tail
is going to be hit,
if we know that all methods of Base.tail
return one of a small number of types,
then we can generate the code for all of them and just put in a if
(It's like the small union's optimization)
from chainrulescore.jl.
Ah ok.
So every week the mental model you need grows more complex. I guess this is a strong argument against defining your own types, when you can avoid it. "Your struct is as good as Base's" isn't quite true, it seems.
from chainrulescore.jl.
So every week the mental model you need grows more complex.
Yeah, I miss the days before julia had an optimizer. It used to be easier to reason about.
Julia is IMO a very hard language to reason about these days, luckily we have good tools now; and this stuff only matters right at the pointy-end of micro-optimizations which is done by experts.
Apparently world-splitting was introduced at the same time as small-union splitting.
But possibly SnoopCompile couldn't recognize it til recently.
idk
from chainrulescore.jl.
I think it can also be totally independent of any of these optimizations. If anything is inferred only as Any
in some code in base or the stdlibs before calling tail
on it, the methods in question need to be re-inferred and re-compiled when loading this package introducing a new method for tail
. So the best fix would be to hunt the type instabilities in Julia base/stdlibs and fix them there. But someone needs to spend some time on this.
from chainrulescore.jl.
I think it can also be totally independent of any of these optimizations. If anything is inferred only as Any in some code in base or the stdlibs before calling tail on it, the methods in question need to be re-inferred and re-compiled when loading this package introducing a new method for tail.
No, I don't think so.
Without this optimization that should result in a dynamic dispatch, whcih would avoid need to recompilation.
from chainrulescore.jl.
I wonder how much we'd have to change in Base to delete Tangent entirely. Being able to add them is the big virtue, what chance we could get these methods accepted?
Base.:+(x::NamedTuple{s}, y::NamedTuple{s}) where {s} = map(+, x, y)
Base.:+(x::Tuple{Vararg{Any,N}}, y::Tuple{Vararg{Any,N}}) where {N} = map(+, x, y)
from chainrulescore.jl.
JuliaLang/julia#46762 helps a bit but does not fix all invalidations.
from chainrulescore.jl.
Related Issues (20)
- ProjectTo access to undefined reference HOT 5
- writing rules for <:AbstractArray HOT 1
- ChainRulesCore v1.15.4 seems to have broken ChainRules HOT 2
- Loading ChainRulesCore.jl breaks complex number arithmetic on Julia 1.8.1 HOT 7
- ProjectTo(::Vector{Vector{Float64}}) type unstable HOT 2
- `Reusable` and `NonReusable` reverse mode capability
- `test_rrule` fails on the pedagogical example HOT 2
- Precompiling ChainRules stack not defined HOT 3
- Warnings about unused type variable HOT 3
- Document how to define own ProjectTo's
- Document how to construct Wirtinger derivative
- Making `getproperty`, `getindex` and `iterate` more type stable for `Tangent`s HOT 1
- Make ZeroTangent() == 0 HOT 3
- Using the gradient function from Flux / Zygote with a custom rrule HOT 5
- Cannot generate `frule` seed via `one(x)` HOT 1
- overload iszero for Tangent
- `ProjectTo(::AbstractSparseMatrix{Bool})` should be trivial
- No method matching `(::ChainRulesCore.ProjectTo)(::Tuple{Float64})` HOT 9
- ProjectTo causes scalar indexing when taking adjoints of complex CuArray HOT 1
- FAQ Broken Links HOT 3
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from chainrulescore.jl.