Returning scalars from NDData

[embray]

Currently if one wants to return a scalar value from an NDData object it's a bit of a mess:

In [1]: from astropy.nddata import NDData

In [2]: d = NDData([1, 2, 3])

In [3]: d[0]
Out[3]: <astropy.nddata.nddata.NDData at 0x2b33e90>

In [4]: d[0].data
Out[4]: array(1)

In [5]: d[0].data.item()
Out[5]: 1

Right now even a scalar is wrapped in the NDData class with metadata and everything. I realize there are issues with this involving masks, error propagation, units (for arrays with units a scalar can just be returned as a Quantity object). But I think that right now for simple cases like, "I just want an array with metadata attached" the current behavior is bad. It would be better to implement reasonable behavior for now than to leave the class in this somewhat unusable state in anticipation of "eventually" solving all conceivable problems with how the class should ideally work.

[Cadair]

I would think that a quantity is the way to go, the only obvious metadata loss in the current implementation would be coordinate data. (?) Which would be avoided if slicing by coordinate is implemented. I think returning a quantity is the most sensible option.

[astrofrog]

Potentially controversial (this has come up before) - now that Quantity inherits from np.ndarray, should NDData inherit and extend Quantity? It would behave more intuitively for people used to Quantity then.

[Cadair]

I, from a SunPy point of view, would be adverse to NDData inheriting ndarray in any form, having just removed this situation from our code base!!

[astrofrog]

Ok - I just wanted to make sure we considered that option, but it would be great to have your insight into why you decided to move away from this?

[embray]

I think the thing that makes NDData complicated is that it's really a container for multiple arrays that need to stay associated with each other. Still, I don't like that we now have these two specialized array types that have very different purposes. Can a Quantity array be used as the science data in an NDData? Has this even been tested?

I think the NDData class really needs some loving now, as I want to be able to extend it to add more specialized science data types like Image and Spectrum classes.

[astrofrog]

@iguananaut - that's a good idea (using Quantity for NDData.data).

[Cadair]

I agree that could be useful but I would suggest that it not be required. What happens if you want to use a memmap of a ndarray or a file i.e. HDF5 or FITS? The wonderful thing about NDData is that it can have any data attribute, be it a ndarray, a pandads dataframe etc.

@embray You should have a look at the SunPy map module, it is close to being a NDData specilisation for a Image. I am working on designing one for a multi-wavelength image and a timeseries' there of.

[astrofrog]

@Cadair - those are very good points - we could keep NDData as simply requiring an ndarray-compatible object inside, rather than strictly an ndarray.

[Cadair]

I think that would be the best solution. Obviously subclasses could be more specific.

[embray]

@Cadair That was my thinking too. I think it would be often very useful to use Quantity for the NDData's data object, but there's no reason it should be strictly required (though I would say if NDData.data is a Quantity then NDData.unit should be tied to NDData.data.unit somehow.

Note also an mmap'd array can still be the base for a Quantity array too.

[embray]

This issue will probably become mostly irrelevant as the purpose and scope of NDData is being reconsidered. However, I'll leave it open until it actually becomes irrelevant and/or is fixed in the new implementation.

[wkerzendorf]

@embray still want to leave it open?

[embray]

Hmm, well, I don't know if the NDData APE mentions this (probably not), but it might be nice if we required that when a slice returns a scalar it returns an actual scalar instead of a zero-dimensional array.

[mwcraig]

Closing, since it appears to be fixed in the latest (v1.3) nddata:

In [2]: from astropy.nddata import NDData

In [3]: d = NDData([1, 2, 3])

In [4]: d[0]
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-4-17371c6688f6> in <module>()
----> 1 d[0]

TypeError: 'NDData' object does not support indexing

In [5]: d.data[0]
Out[5]: 1