LinearInterpolatedResampler fails on uncertainty of UnknownUncertainty type

[dhomeier]

When (unsuccessfully) trying to reproduce #647 I raised a different exception by modifying the docs' example:

input_spectrum = Spectrum1D(flux=[1, 3, 7, 6, 20] * u.mJy,
                            spectral_axis=[2, 4, 12, 16, 20] * u.nm,
                            uncertainty=[0.4, 0.5, 0.6, 0.8, 1])
resample_grid = [1, 5, 9, 13, 14, 17, 21, 22, 23] * u.nm                                                                                                                        
fluxc_resample = LinearInterpolatedResampler()

input_spectrum.uncertainty                                                                                                                                                      
    UnknownUncertainty([0.4, 0.5, 0.6, 0.8, 1. ])
output_spectrum1D = fluxc_resample(input_spectra, resample_grid)

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
<ipython-input-153-d2ada17e1ff7> in <module>
----> 1 output_spectrum1D = fluxc_resample(input_spectrum, resample_grid)

~/lib/python3.8/site-packages/specutils/manipulation/resample.py in __call__(self, orig_spectrum, fin_spec_axis)
     37         Return the resulting `~specutils.Spectrum1D` of the resampling.
     38         """
---> 39         return self.resample1d(orig_spectrum, fin_spec_axis)
     40 
     41     @abstractmethod

~/lib/python3.8/site-packages/specutils/manipulation/resample.py in resample1d(self, orig_spectrum, fin_spec_axis)
    321                                                           unit=orig_spectrum.unit)
    322 
--> 323         return Spectrum1D(spectral_axis=fin_spec_axis,
    324                           flux=out_flux,
    325                           uncertainty=new_unc)

~/lib/python3.8/site-packages/specutils/spectra/spectrum1d.py in __init__(self, flux, spectral_axis, wcs, velocity_convention, rest_value, redshift, radial_velocity, **kwargs)
    157                         spectral_axis.shape[0], flux.shape[-1]))
    158 
--> 159         super(Spectrum1D, self).__init__(
    160             data=flux.value if isinstance(flux, u.Quantity) else flux,
    161             wcs=wcs, **kwargs)

~/lib/python3.8/site-packages/astropy/nddata/nddata.py in __init__(self, data, uncertainty, mask, wcs, meta, unit, copy)
    232         self._unit = unit
    233         # Call the setter for uncertainty to further check the uncertainty
--> 234         self.uncertainty = uncertainty
    235 
    236     def __str__(self):

~/lib/python3.8/site-packages/astropy/nddata/nddata.py in uncertainty(self, value)
    324                 # to be saved as weakref but that's done by NDUncertainty
    325                 # setter).
--> 326                 value.parent_nddata = self
    327         self._uncertainty = value

~/lib/python3.8/site-packages/astropy/nddata/nduncertainty.py in parent_nddata(self, value)
    231                 unit_from_data = self._data_unit_to_uncertainty_unit(parent_unit)
    232                 try:
--> 233                     unit_from_data.to(self.unit)
    234                 except UnitConversionError:
    235                     raise UnitConversionError("Unit {} of uncertainty "

AttributeError: 'NoneType' object has no attribute 'to'

Is there other functionality that would handle an astropy.nddata.nduncertainty.UnknownUncertainty correctly, or should this best be checked on initialising Spectrum1D?

[nmearl]

We're discussing ways to make the uncertainty creation easier, but currently, you should be initializing the Spectrum1D with an explicit uncertainty class in order for everything to Just Work™️

from astropy.nddata import StdDevUncertainty

input_spectrum = Spectrum1D(flux=[1, 3, 7, 6, 20] * u.mJy,
                            spectral_axis=[2, 4, 12, 16, 20] * u.nm,
                            uncertainty=StdDevUncertainty([0.4, 0.5, 0.6, 0.8, 1]))
resample_grid = [1, 5, 9, 13, 14, 17, 21, 22, 23] * u.nm                                                                                                                        
fluxc_resample = LinearInterpolatedResampler()

[dhomeier]

We're discussing ways to make the uncertainty creation easier, but currently, you should be initializing the Spectrum1D with an explicit uncertainty class in order for everything to Just Work™️

Yes, was just wondering if incorrect instantiation should be intercepted earlier – e.g. adding an additional check for uncertainty_type to if hasattr(self, 'uncertainty') and self.uncertainty is not None: ...

But an optional keyword like uncertainty_type='std' would be convenient, too.

[nmearl]

Indeed. See #529 for some early discussion. Most recent discussion has happened out-of-band.

[kelle]

I was also unable to write out a spectrum with an uncertainty of UnknownUncertainty type.