Application of mask_safe in MapDataset and SpectrumDataset does not handle nan values well

[lmohrmann]

Gammapy version
0.17

Bug description
The mask_safe is applied in the MapDataset and SpectrumDataset by multiplying the different maps (counts, background, etc.) with the mask_safe. This method fails when, for instance, there are nan values present in the background model map (which might occur when the background model is interpolated onto the used geometry). This is because nan * False = nan, which results in the corresponding pixels not being properly masked.

Expected behavior
All pixels outside the mask_safe should be properly masked. This could be achieved by using a scheme like dataset.background_model.map[~dataset.mask_safe.data] = 0 instead of the multiplication.

[adonath]

@lmohrmann I checked once again what the status here is and to apply the mask we consistently use Map.stack( weights=). The weights argument however is meant to be more generic and not strictly for boolean types. So I think we cannot apply the proposed solution and have to fix replace NaN values beforehand instead. A method for this has been added in #3219

[adonath]

However one could check the dtype here: https://github.com/gammapy/gammapy/blob/master/gammapy/maps/wcsnd.py#L782

[QRemy]

PR #3416 adds a method "bkg-clip" to the SafeMaskMaker in order create a mask from aberrant values in the background map (not finite and eventually very large values). It also offers the possibility to clip the aberrant values to zero in the background map. This should solve the issue.

[adonath]

Solved by #3419.