Transpose doesn't change the number of dimensions, just reverses their order. I think another function would be needed, and the multidimensional case should also be included.

Yeah, I think what's really needed here is a "transpose last two axes" function (#7495), in which automatic promotion from 1d to 2d seems reasonable

@jolespin: Notice that np.transpose([x]) is not the same as np.transpose(x). In the first case, you're effectively doing np.array([x]) as a (somewhat confusing and non-idiomatic) way to promote x to a 2-dimensional row vector, and then transposing that.

@eric-wieser: So would a 1d array be promoted to a row vector or a column vector before being transposed? If we had an operation to swap the last two axes, then it would make more sense to me if it fails when there aren't two axes to swap. In the face of ambiguity, refuse the temptation to guess, and all that.

@njsmith You'd promote an (N,) array to a (1, N) (row) array (and then transpose to (N,1) (column)), as broadcasting is always done by adding singular axes to the left. Perhaps this is still too magic, but I would say the choice between row and column is clear

Also, np.transpose([x]) will bite you if x is already 2d, as that produces a 3d array.

I should draw attention to the edit to my original comment linking to a mostly-duplicate issue, #7495

I don't think broadcasting is relevant here at all. Broadcasting is a way of lining up parallel loops, not a method for arbitrary munging of arrays to meet some shape requirement. Note that np.dot also coerces 1d arrays to 2d, but it does not use the broadcasting rule.

Also, if that's the promotion behavior you want you can already get it by explicitly writing np.atleast_2d(x).T, or if we had some sort of stacked_transpose it could be stacked_transpose(np.atleast_2d(x)).

I don't think broadcasting is relevant here at all. Broadcasting is a way of lining up parallel loops, not a method for arbitrary munging of arrays to meet some shape requirement.

+1

Using the model of generalized ufuncs, the "transpose last two axes" function would have the core dimensions signature (m,n)->(n,m). Following the rules for generalized ufuncs, that means it should error, not broadcast when given 1D inputs. (See #8811 and #5077 for discussion about broadcasting inside core dimensions of gufuncs.)

In #8827 (comment), @charris suggested .M[atrix]T[ranspose]=.MT as the name for the possible new property. With such a property, transposing a shape (N,) to (N, 1) would seem reasonable -- the implementation would then be np.swapaxes(atleast_2d(array), -2, -1).

Changing the operation's spelling from a function to a property does not change my opinion on the semantics :-)

I just merged .mT, so closing.