Add matrix, an operation to convert an array to a matrix (like vec)

[oxinabox]

We have vec which is a short-hand for reshape(x, Val(1)).
I propose that we should have mat as well, as a shorthand for reshape(x,Val(2)).

Obviously vec is the most common case of reshaping,
but I suggest that mat comes in second, and is much more common than the general reshape.
In particular for converting a Vector into a single column matrix.
Very common for interfacing with various libraries.

My old trick for that was to double transpose it (x''), but since as of 0.6 we take vector transpose seriously, that won't work. (Plus this way is nonallocating).

[andreasnoack]

This would be useful to have. I think I'd prefer matrix instead of mat even though mat is similar to vec. However, vec is standard whereas mat is not and I prefer to spell things out.

[andreasnoack]

Should be AbstractMatrix instead of Matrix.

[oxinabox]

Makes sense to spell it out. mat is pretty vague.
I liked the 3 letter match to vec, but explict is better.
I'll fix the doc on vec too w.r.t Abstract

[GunnarFarneback]

Would it be reasonable to expect this to convert a RowVector to a 1xN matrix? The matrix(x')' workaround isn't that attractive.

[oxinabox]

Would it be reasonable to expect this to convert a RowVector to a 1xN matrix? The matrix(x')' workaround isn't that attractive.

What do you mean?
A RowVector is a 1xN AbstractMatrix already.
If you want to convert it to a 1xN concrete Matrix, use Matrix, or convert(Matrix, x)

julia> supertype(RowVector) == AbstractMatrix
true

julia> x = RowVector([1,2,3])
1×3 RowVector{Int64,Array{Int64,1}}:
 1  2  3

julia> size(x)==(1,3)
true

julia> matrix(x)
1×3 RowVector{Int64,Array{Int64,1}}:
 1  2  3

julia> matrix(x) == x
true


julia> Matrix(x)
1×3 Array{Int64,2}:
 1  2  3

[GunnarFarneback]

Yes, I wanted a concrete Matrix without copying the data. Maybe I should just look the other way to make the receivers of the data more open to abstract matrices.

[andyferris]

I like the direction. There are a bunch of cases where you want to cast your quasi-1d collection into a specific shape, so that's cool.

I totally haven't got around to this, but I was hoping to rather have this little family of functions:

• vec (as now)
• rowvec (basically rowvec(x) = vec(x).' with non-recursive transpose)
• rowmat (never makes a RowVector)
• colmat

since sometimes you want a column matrix and sometimes a row matrix.

[GunnarFarneback]

There's still something bothering me with the pseudomatrixness of RowVectors. If I say matrix(x) I would expect to get something that is matrixy enough that I can transpose it and always have a matrix, not something that occasionally reverts to a vector.

[oxinabox]

@andreasnoack is this good to go now?
As far as satisfying your review

Should I add @andyferris 's stuff?
@andyferris so colmat === matrix as currently defined, right?

[andyferris]

The same, yes.

@GunnarFarneback I agree. matrix or rowmat or whatever should never let the RowVector wrapper survive.

[andreasnoack]

Now that we have a RowVector I think Vectors are effectively columns so I think it would be fine to make matrix(Vector) return a column matrix and matrix(RowVector) return a row matrix.

[Sacha0]

With the present direction of array constructors, Matrix constructors could eventually do the trick? And perhaps likewise with vec, Vector constructors? Best!

[mbauman]

I don't think constructors can fill this role since I'd expect them to copy the data; vec and reshapes are views. Maybe #23821?

Of course, convert(AbstractMatrix, x) is even more verbose than reshape, so the primary motivation here still remains.

[Sacha0]

I don't think constructors can fill this role since I'd expect them to copy the data; vec and reshapes are views.

Ah, right! Thank you :).

Maybe #23821?

👍

[oxinabox]

@andyferris's set of functions is progressively growing on me.

[jebej]

rowvec (basically rowvec(x) = vec(x).' with non-recursive transpose)

@andyferris Shouldn't that operation go row major, i.e. rowvec(x) = vec(x.')

[andyferris]

@jebej I wouldn't have thought so... the idea here is to turn any iterable into a RowVector such that iteration is preserved. If users want to change the iteration ordering to row major they should do that explicitly with .'.

[oxinabox]

I think we can close this. We have matrix literals even for matrixes of length 1, and we have reshape.