Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Normalize indices in promote_shape error messages (take 2) #41311

Merged
merged 5 commits into from
Feb 4, 2024

Conversation

tpapp
Copy link
Contributor

@tpapp tpapp commented Jun 22, 2021

Seeing implementation like Base.OneTo in error messages may be confusing to some users (cf discussion in #39242,
discourse).

This PR turns

julia> ones(2, 3) + ones(3, 2)
ERROR: DimensionMismatch("dimensions must match: a has dims (Base.OneTo(2), Base.OneTo(3)), b has dims (Base.OneTo(3), Base.OneTo(2)), mismatch at 1")

into

julia> ones(2, 3) + ones(3, 2)
ERROR: DimensionMismatch("dimensions must match: a has axes (1:2, 1:3), b has axes (1:3, 1:2), mismatch at 1")

Fixes #40118.

(This is basically #40124, but redone because I made a mess rebasing).

@mcabbott
Copy link
Contributor

mcabbott commented Jun 23, 2021

Might be cleaner just to say "a has size (2, 3)" when axes are OneTo?

Also, if changing things, could we kill the repetition in "DimensionMismatch("dimensions must match" ? It looks like a mistake although, when compared to ones(2, 3) .+ ones(3, 2)'s "DimensionMismatch("arrays could not be broadcast to a common size" perhaps it is trying to say that exact agreement is expected here. But surely it could say so without looking like a mistake, maybe "all nontrivial dimensions must agree" (since ones(2,3) + ones(2,3,1) is fine)

@tpapp
Copy link
Contributor Author

tpapp commented Jun 23, 2021

Might be cleaner just to say "a has size (2, 3)" when axes are OneTo?

That is indeed shorter, but at the same time I think the language should encourage people to think in terms of generalized indexing all the time since it is part of the <:AbstractArray interface, and error messages should be part of that. I think that the extra 1:s are a small price for that, and consistency.

kill the repetition

done, thanks.

@mcabbott
Copy link
Contributor

extra 1:s are a small price

My concern isn't brevity, it's more that zeros(1:2, 1:3) isa OffsetArray while zeros(2, 3) isa Array -- OneTo(3) has special printing precisely because it's not equivalent to 1:3.

@tpapp
Copy link
Contributor Author

tpapp commented Jun 23, 2021

I think the error message is pretty clear about these axes; I am not quite sure how zeros comes into it.

Also, what should happen for two arrays with axes eg (Base.OneTo(4),) and (3:5, )?

@mcabbott
Copy link
Contributor

mcabbott commented Jun 23, 2021

My point is that, for axes, 1:4 and OneTo(4) are meaningfully different. So I think we should be a little cautious about printing one as the other, which so far Base avoids doing. They are == of course, and the error is !=, so it's not crazy... but it strays from what Base would accept as input. (zeros is just an example, reshape etc. behave the same way.)

What should be printed in cases where some but not all of the axes in play are Base.OneTo, that's less obvious. If you're complaining that mixing size & axes in one message would be odd, then I agree, we shouldn't do that. This is certainly quite a mouthful at present:

julia> reshape(rand(3), 2:4) + ones(3)
ERROR: DimensionMismatch("dimensions must match: a has dims (OffsetArrays.IdOffsetRange(values=2:4, indices=2:4),), b has dims (Base.OneTo(3),), mismatch at 1")

Maybe printing something like axes(a) == (2:4,) but axes(b) == (1:3,) might make sense in the presence of offsets. To stress that this isn't printed in re-digestible form like repr, only up to ==.

@tpapp
Copy link
Contributor Author

tpapp commented Jun 23, 2021

are meaningfully different. [...] They are == of course

Sorry, I don't understand how values that are == can be considered meaningfully different.

I think that Base.OneTo is an irrelevant implementation detail that should be hidden from the user to the extent possible (cf the discussion in #39242). This is what this PR aims to do, nothing more.

@mcabbott
Copy link
Contributor

Sorry, I don't understand how values that are == can be considered meaningfully different.

I thought I gave lots of examples. One more, on a fresh session:

julia> reshape(rand(4), Base.OneTo(2), Base.OneTo(2)) |> summary
"2×2 Matrix{Float64}"

julia> reshape(rand(4), 1:2, 1:2)
ERROR: MethodError: no method matching reshape(::Vector{Float64}, ::Tuple{UnitRange{Int64}, UnitRange{Int64}})

@antoine-levitt
Copy link
Contributor

Ping, this is still grating and would be great to fix. Reading up the discussion, it seems that @mcabbott 's objection is that we need to carefully distinguish different types of axes, but it seems to me that it would be OK to have a simpler (possibly different) error message in the common cases of "standard" axes. @mcabbott do you think that's acceptable? If yes, any thoughts as to how to proceed?

@tpapp
Copy link
Contributor Author

tpapp commented Jul 18, 2022

@antoine-levitt: if triage gives a clear direction, I would be happy to update this PR.

FWIW, I would leave it as is: the only documented methods of reshape in Base are those that deal with dims::Tuple{Vararg{Union{Colon, Int64}}}. When (if) this part of the API is made more consistent, we can address that in another PR.

tpapp and others added 5 commits February 3, 2024 04:54
Seeing implementation like `Base.OneTo` in error messages may be
confusing to some users (cf discussion in JuliaLang#39242,
[discourse](https://discourse.julialang.org/t/promote-shape-dimension-mismatch/57529/)).

This PR turns
```julia
julia> ones(2, 3) + ones(3, 2)
ERROR: DimensionMismatch("dimensions must match: a has dims (Base.OneTo(2), Base.OneTo(3)), b has dims (Base.OneTo(3), Base.OneTo(2)), mismatch at 1")
```
into
```julia
julia> ones(2, 3) + ones(3, 2)
ERROR: DimensionMismatch("dimensions must match: a has axes (1:2, 1:3), b has axes (1:3, 1:2), mismatch at 1")
```

Fixes JuliaLang#40118.

Acked-by: Tamas K. Papp <[email protected]>
- avoid assuming the tuples have the same eltype
- coerce the inputs to simpler representations before printing, if
  unambiguous
- update a couple more error messages with the same text
- use string builder, instead of repeated concatenation
@vtjnash vtjnash force-pushed the tp/tp/promote_shape_error_msg2 branch from 4f66ee5 to ce7a16e Compare February 3, 2024 05:53
@vtjnash
Copy link
Member

vtjnash commented Feb 3, 2024

I have tried to take the feedback of the discussion into account, so now it reads, and I think should be ready to merge:

julia> ones(2, 3) + ones(2, 4)
ERROR: DimensionMismatch: a has size (2, 3), b has size (2, 4), mismatch at dim 2
Stacktrace:
 [1] throw_promote_shape_mismatch(a::Tuple{Base.OneTo{Int64}, Base.OneTo{Int64}}, b::Tuple{Base.OneTo{Int64}, Base.OneTo{Int64}}, i::Int64)
   @ Base ./indices.jl:135
 [2] promote_shape
   @ ./indices.jl:196 [inlined]
 [3] promote_shape
   @ ./indices.jl:188 [inlined]
 [4] +(A::Matrix{Float64}, Bs::Matrix{Float64})
   @ Base ./arraymath.jl:14
 [5] top-level scope
   @ REPL[1]:1

julia> include("test/testhelpers/OffsetArrays.jl")
Main.OffsetArrays

julia> reshape(rand(3), 2:4) + ones(3)
ERROR: DimensionMismatch: a has axes (2:4,), b has axes (1:3,), mismatch at dim 1
Stacktrace:
 [1] throw_promote_shape_mismatch(a::Tuple{Main.OffsetArrays.IdOffsetRange{Int64, Base.OneTo{Int64}}}, b::Tuple{Base.OneTo{Int64}}, i::Int64)
   @ Base ./indices.jl:135
 [2] promote_shape
   @ ./indices.jl:196 [inlined]
 [3] promote_shape
   @ ./indices.jl:188 [inlined]
 [4] +(A::Main.OffsetArrays.OffsetVector{Float64, Vector{Float64}}, B::Vector{Float64})
   @ Base ./arraymath.jl:7
 [5] top-level scope
   @ REPL[3]:1

@vtjnash vtjnash added linear algebra Linear algebra error messages Better, more actionable error messages labels Feb 3, 2024
@vtjnash vtjnash merged commit 47663bd into JuliaLang:master Feb 4, 2024
7 of 9 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
error messages Better, more actionable error messages linear algebra Linear algebra
Projects
None yet
Development

Successfully merging this pull request may close these issues.

Promote_shape dimension mismatch print.
4 participants