Feat: xarray extra dims coordinates

[tmichela]

Adds coordinates for extra dimensions, particularly useful when roi is defined.

[tmichela]

A potential bug I noticed while testing this feature: if a roi contains an integer (e.g. np.s_[0, :10]) the assigned dimensions will skip that axis, i.e.:
The resulting xarray has shape=(<n-trains>, 10) and dims=['trainId', 'dim_0'].
But I think it should rather be dims=['trainId', 'dim_1'] (dim_0 is dropped by the slicing).

?

[takluyver]

if a roi contains an integer (e.g. np.s_[0, :10]) the assigned dimensions will skip that axis, i.e.:
The resulting xarray has shape=(<n-trains>, 10) and dims=['trainId', 'dim_0'].

I think you're right. I didn't really give any attention to the logical consistency of dimension names, they just have to be called something. 🤷

[philsmt]

I don't think I stumbled on this problem yet, but this seems to be the better solution.

If arbitrary coordinates are a concern here, we should rather make using those more convenient than have the returned dimensions be more useless.

The question around dim_0 or dim_1 if a dimension is omitted is tricky. Given that extra_dims must match the final shape, this is more in line with starting with dim_0. In other words, if we skip dimensions sliced out of the result, extra_dims should also look at the full shape rather than the final one.

[takluyver]

I was surprised this works, but it does:

In [11]: 'dim_0' in da.coords
Out[11]: False

In [12]: da.coords['dim_0'].values
Out[12]: array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])

TIL

[takluyver]

LGTM

[tmichela]

thanks