Add general Background classes

[larrybradley]

This large PR contains the following changes:

1. Turns the background.py module into a background subpackage.
2. Adds many new classes for computing a (scalar) background value or background rms value from a (possibly masked) array of any dimension:

• MeanBackground
• MedianBackground
• MMMBackground
• SExtractorBackground
• BiweightLocationBackground
• StdBackgroundRMS
• MADStdBackgroundRMS
• BiweightMidvarianceBackgroundRMS

Sigma-clipping is performed by default, but can be turned off by setting sigclip = False.
3. The existing BackgroundBase, Background and BackgroundIDW classes are now called BackgroundBase2D, Background2D, and BackgroundIDW2D, respectively, since they return 2D background images.

My plan is to submit another PR to integrate the general background classes into the 2D background classes (with a similar interface).

Example

from photutils import MMMBackground
bkg = MMMBackground(sigma=3., iters=10)
bkg_value = bkg.calc_background(data)

The classes are designed to be used as an object-oriented interface for other functions, e.g.:

bkg = MMMBackground(sigma_lower=5.5, sigma_upper=3., iters=10)
result = psf_photometry(data, ... , background_sub=bkg)

in function form (requires calc_background -> __call__):

result = psf_photometry(data, ... , background_sub=np.ma.mean)

where the user can define custom background classes if desired.

cc: @eteq, @hcferguson

[larrybradley]

The background classes are now callable. The __call__ method simply passes the data to the calc_background() method.

[eteq]

I guess this comment really belongs just a bit further up... But I wonder if you should add a short example in the previous section that demos all of the various Background (not 2D) classes? Right now it all uses the sigma clipping functions, but now you can insert all of the various estimators implemented here. Maybe this could just be a (logarithmic) histogram of the pixel values and then mark the output of the estimators as vertical lines?

[larrybradley]

Yes, I didn't add the new Background classes to narrative docs yet. I wanted to wait until the API was firmly decided.

[eteq]

@larrybradley - left some inline comments above, and I've got one more general suggestion:
I found the base hierarchy at first conceptually confusing - that is, why is there a BkgBase at all, when it didn't seem to do anything related to background. After some thought, I realized it's there because it provides sigma-clipping... But I think that doesn't actually need to be the base of the background heirarchy... And in fact it's a bit confusing because it implies background subtraction must involve sigma clipping.

So I have an alternate suggestion that I think is just a small perturbation on what you've done already: change BkgBase to something like SigmaClipper, and use it as a mixin, rather than the base. That is, have BackgroundBase and BackgroundRMSBase be object subclasses, but have all the other classes have inheritance signatures along the lines of MMMBackground(BackgroundBase, SigmaClipper). You might have to be a bit more careful with the super calls, but I think it's a non-issue here because there's no BackgroundBase.__init__ anyway.

[larrybradley]

Thanks, @eteq! Please review the updates.

[eteq]

looks great to me now, @larrybradley ... except for the test failures. Any idea what might be specific about numpy 1.9 that would cause this?

[larrybradley]

@eteq Thanks.

I guess you missed the ongoing saga with conda and ci-helpers [e.g. 1, 2, 3]. The very short summary is that conda 4.1.0 broke a lot of things last week (among other things leaving out the default channel). Our conda channels are in a state of flux. I think the plan was to wipe astropy-ci-extras and start over in conda-channel-ci-extras, but nothing is in there yet. travis-ci tests are failing across all astropy packages, including the core package. conda 4.1.1 and 4.1.2 are out now, but I think we need to re-build astropy with older numpy versions.

[1] astropy/ci-helpers#97
[2] astropy/ci-helpers#98
[3] astropy/ci-helpers#100

[eteq]

Ohhh, gotcha - I was confused by the fact that all of the other versions still work, but I see now from those ci issues that 1.9 is indeed problematic.

But given that the others passed, I would think this is good to merge (assuming you're done making changes, @larrybradley ?)