squeeze() does not infer number of dimensions in output array

[simonster]

e.g. code_typed(squeeze, (Matrix{Float64}, Int)) suggests that squeeze returns an Array{Float64,N}, but it seems like it should be possible to infer that it will be an Array{Float64,1}. This would be especially useful since squeeze is the natural way to deal with singleton dimensions returned by sum et al.

[JeffBezanson]

For this to be possible a bit of rewriting would be necessary. Currently some corner cases are possible, such as:

julia> squeeze(rand(2,2),(0,0,0))
2x2 Array{Float64,2}:
 0.154123  0.0170937
 0.296481  0.339155 

julia> squeeze(rand(2,2),(3,))
2x2 Array{Float64,2}:
 0.0312394  0.522369
 0.633781   0.698172

The first one is probably unintentional, but the second one arguably makes sense, unless we want to enforce that squeeze must be able to remove dimensions.

[simonster]

Enforcing that squeeze must be able to remove dimensions seems fine to me. Are there scenarios where one would want to squeeze a dimension that may or may not exist? It doesn't seem particularly unsafe, but it doesn't seem all that useful either, at least for the code I write.

[pao]

@StefanKarpinski explains the behavior here: https://groups.google.com/d/msg/julia-users/QQqVbyTiiPg/da5qkiMQVQIJ

[simonster]

I think that post is about a different aspect of squeeze. I agree with @StefanKarpinski that blindly dropping all singleton dimensions is a bad idea (and indeed, we definitely couldn't infer the number of dimensions in the output in that case). The question is whether squeeze should throw when the dimension to be squeezed exceeds the number of dimensions in the input (e.g. squeeze(Array{T,2}, 3)) so that squeeze(Array{T,N}, n) would always return an Array{T,N-1}.

[StefanKarpinski]

I think we should definitely fix squeeze so that the number of dimensions of the result can be statically inferred.

[pao]

Wow, I totally missed that--sorry for the noise. False positive on my FAQ alarm.