Panel partial index is transposed when minor_index is MultiIndex

[benmoran]

I can index a Panel to get a data frame, e.g. by major_axis:

>>> p = pd.Panel(np.arange(8).reshape(2,2,2),items=('A1','B1'),major_axis=('A2','B2'),minor_axis=('A3','B3'))
>>> p.loc[:,'A2',:]

    A1  B1
A3   0   4
B3   1   5

But if I do the same index when the minor_axis is a MultiIndex the resulting data frame is unexpectedly transposed:

mi = pd.MultiIndex((['A3','B3'],['a','b']), [(0,0),(1,1)])
p = pd.Panel(np.arange(8).reshape(2,2,2),items=('A1','B1'),major_axis=('A2','B2'),minor_axis=mi)
p.loc[:,'A2',:]

   A3   
    b  b
A1  0  1
B1  4  5

Actually I think the second behaviour is closer to what I expect - the items axis becomes dimension 0 of the DataFrame. But whichever is used, we should have stability, especially if 1-item MultiIndex objects automatically become Index objects.

[jreback]

so these hit 2 slightly different paths, and the one 2nd one has a 'fix' for this issue: #7516 and in the code here

I remember making this change and it seems reasonable.

If you'd like to read the other issue and see what is most logical would be helpful.

[benmoran]

@jreback sorry, I'm not sure what I'm looking at in the other issue you referenced?

This is the behaviour that seems "most logical" to me:

• most important is that whatever type my indices happen to be doesn't affect the transposition order of the output
• secondly, I'd expect that removing a dimension doesn't change the order of the other dimensions

I think the latter is what Numpy does, so "np.random.randn(3,4,5)[:,1,:].shape" gives (3,5) and not (5,3).

[jreback]

closing as Panels are deprecated