3. AstroData and Derivatives

The astrodata.core.AstroData class (or simply astrodata.AstroData) is the main interface to the package. When opening files or creating new objects, a derivative of this class is returned, as the AstroData class is not intended to be used directly. It provides the logic to calculate the tag set for an image, which is common to all data products. Aside from that, it lacks any kind of specialized knowledge about the different instruments that produce the FITS files. More importantly, it defines two methods (info and load) as abstract, meaning that the class cannot be instantiated directly: a derivative must implement those methods in order to be useful. Such derivatives can also implement descriptors, which provide processed metadata in a way that abstracts the user from the raw information (e.g., the keywords in FITS headers).

AstroData does define a common interface, though. Much of it consists on implementing semantic behavior (access to components through indices, like a list; arithmetic using standard operators; etc), mostly by implementing standard Python methods:

  • Defines a common __init__ function, that accepts a DataProvider as its single argument.
  • Implements __deepcopy__
  • Implements __iter__ to allow sequential iteration over the main set of components (e.g., FITS science HDUs, but this depends on the DataProvider implementation)
  • Implements __getitem__ to allow data slicing (e.g., ad[2:4] returns a new AstroData instance that contains only the third and fourth main components)
  • Implements __delitem__ to allow for data removal based on index. It does not define __setitem__, though. The basic AstroData series of classes only allows to append new data blocks, not to replace them in one sweeping move
  • Implements __iadd__, __isub__, __imul__, __itruediv__, and their not-in-place versions, based on them.

All of these provide default implementations that rely heavily on the DataProvider capabilities. There are a few other methods. For a detailed discussion, please refer to the API Reference Guide.

3.1. The tags Property

Additionally, and crucial to the package, AstroData offers a tags property, that under the hood calculates textual tags that describe the object represented by an instance, and returns a set of strings. Returning a set (as opposed to a list, or other similar structure) is intentional, because it is fast to compare sets, e.g., testing for membership; or calculating intersection, etc., to figure out if a certain dataset belongs to an arbitrary category.

The implementation for the tags property is just a call to AstroData.__process_tags(). This function implements the actual logic behind calculating the tag set (described below). A derivative class could redefine the algorithm, or build upon it.

3.2. Writing an AstroData Derivative

The first step when creating new AstroData derivative hierarchy would be to create a new class that knows how to deal with some kind of specific data in a broad sense. DRAGONS provide such a class for FITS files, astrodata.fits.AstroDataFits, that can be used as an example for future extensions (e.g., to support the ASDF format).

AstroDataFits implements both info and load in ways that are specific to FITS files. It also introduces a number of FITS-specific methods and properties, e.g.:

  • The properties phu and hdr, which return the primary header and a list of headers for the science HDUs, respectively.
  • A write method, which will write the data back to a FITS file
  • A _matches_data static method, which is very important, involved in guiding for the automatic class choice algorithm during data loading. We’ll talk more about this when dealing with registering our classes.

It also defines the first few descriptors, which are common to all Gemini data: instrument, object, and telescope, which are good examples of simple descriptors that just map a PHU keyword without applying any conversion.

A typical AstroData programmer will extend this class (AstroDataFits), unless introducing support for a different kind of data storage. Any of the classes under the gemini_instruments package can be used as examples, but we’ll describe the important bits here.

3.2.1. Create a package for it

This is not strictly necessary, but simplifies many things, as we’ll see when talking about registration. The package layout is up to the designer, so you can decide how to do it. For DRAGONS we’ve settled on the following recommendation for our internal process (just to keep things familiar):


Where instrument_name would be the package name (for Gemini we group all our derivative packages under gemini_instruments, and we would import gemini_instruments.gmos, for example). __init__.py and adclass.py would be the only required modules under our recommended layout, with lookup.py being there just to hold hard-coded values in a module separate from the main logic.

adclass.py would contain the declaration of the derivative class, and __init__.py will contain any code needed to register our class with the AstroData system upon import.

3.2.2. Create your derivative class

This is an excerpt of a typical derivative module:

from astrodata import astro_data_tag, astro_data_descriptor, TagSet
from astrodata import AstroDataFits

from . import lookup

class AstroDataInstrument(AstroDataFits):
    __keyword_dict = dict(
            array_name = 'AMPNAME',
            array_section = 'CCDSECT'

    def _matches_data(source):
        return source[0].header.get('INSTRUME', '').upper() == 'MYINSTRUMENT'

    def _tag_instrument(self):
       return TagSet(['MYINSTRUMENT'])

    def _tag_image(self):
        if self.phu.get('GRATING') == 'MIRROR':
            return TagSet(['IMAGE'])

    def _tag_dark(self):
        if self.phu.get('OBSTYPE') == 'DARK':
            return TagSet(['DARK'], blocks=['IMAGE', 'SPECT'])

    def array_name(self):
        return self.phu.get(self._keyword_for('array_name'))

    def amp_read_area(self):
        ampname = self.array_name()
        detector_section = self.detector_section()
        return "'{}':{}".format(ampname, detector_section)


An actual Gemini Facility Instrument class will derive from gemini_instruments.AstroDataGemini, but this is irrelevant for the example.

The class typically relies on functionality declared elsewhere, in some ancestor, e.g., the tag set computation is defined at AstroData, and the _keyword_for method is defined at AstroDataFits.

Some highlights:

  • __keyword_dict[1] defines one-to-one mappings, assigning a more readable moniker for an HDU header keyword. The idea here is to prevent hard-coding the names of the keywords, in the actual code. While these are typically quite stable and not prone to change, it’s better to be safe than sorry, and this can come in useful during instrument development, which is the more likely source of instability. The actual value can be extracted by calling self._keyword_for('moniker').

  • _matches_data is a static method. It does not have any knowledge about the class itself, and it does not work on an instance of the class: it’s a member of the class just to make it easier for the AstroData registry to find it. This method is passed some object containing cues of the internal structure and contents of the data. This could be, for example, an instance of HDUList, or DataProvider. Using these data, _matches_data must return a boolean, with True meaning “I know how to handle this data”.

    Note that True does not mean “I have full knowledge of the data”. It is acceptable for more than one class to claim compatibility. For a GMOS FITS file, the classes that will return True are: AstroDataFits (because it is a FITS file that comply with certain minimum requirements), AstroDataGemini (the data contains Gemini Facility common metadata), and AstroDataGmos (the actual handler!).

    But this does not mean that multiple classes can be valid “final” candidates. If AstroData’s automatic class discovery finds more than one class claiming matching with the data, it will start discarding them on the basis of inheritance: any class that appears in the inheritance tree of another one is dropped, because the more specialized one is preferred. If at some point the algorithm cannot find more classes to drop, and there is more than one left in the list, an exception will occur, as AstroData will have no way to choose one over the other.

  • A number of “tag methods” have been declared. Their naming is a convention, at the end of the day (the “_tag_” prefix, and the related “_status_” one, are just hints for the programmer): each team should establish a convention that works for them. What is important here is to decorate them using astro_data_tag, which earmarks the method so that it can be discovered later, and ensures that it returns an appropriate value.

    A tag method will return either a TagSet instance (which can be empty), or None, which is the same as returning an empty TagSet[2].

    All these methods will be executed when looking up for tags, and it’s up to the tag set construction algorithm (see Tags) to figure out the final result. In theory, one could provide just one big method, but this is feasible only when the logic behind deciding the tag set is simple. The moment that there are a few competing alternatives, with some conditions precluding other branches, one may end up with a rather complicated dozens of lines of logic. Let the algorithm do the heavy work for you: split the tags as needed to keep things simple, with an easy to understand logic.

    Also, keeping the individual (or related) tags in separate methods lets you exploit the inheritance, keeping common ones at a higher level, and redefining them as needed later on, at derived classes.

    Please, refer to AstroDataGemini, AstroDataGmos, and AstroDataGnirs for examples using most of the features.

  • The AstroDataFits.load method calls the FitsLoader.load method, which uses metadata in the FITS headers to determine how the data should be stored in the AstroData object. In particular, the EXTNAME and EXTVER keywords are used to assign individual FITS HDUs, using the same names (SCI, DQ, and VAR) as Gemini-IRAF for the data, mask, and variance planes. A SCI HDU must exist if there is another HDU with the same EXTVER, or else an error will occur.

    If the raw data do not conform to this format, the AstroDataFits.load method can be overridden by your class, by having it call the FitsLoader.load method with an additional parameter, extname_parser, that provides a function to modify the header. This function will be called on each HDU before further processing. As an example, the SOAR Adaptive Module Imager (SAMI) instrument writes raw data as a 4-extension MEF file, with the extensions having EXTNAME values im1, im2, etc. These need to be modified to SCI, and an appropriate EXTVER keyword added` [3]. This can be done by writing a suitable load method for the AstroDataSami class:

    def load(cls, source):
        def sami_parser(hdu):
            m = re.match('im(\d)', hdu.header.get('EXTNAME', ''))
            if m:
                hdu.header['EXTNAME'] = ('SCI', 'Added by AstroData')
                hdu.header['EXTVER'] = (int(m.group(1)), 'Added by AstroData')
        return cls(FitsLoader(FitsProvider).load(source, extname_parser=sami_parser))
  • Descriptors will make the bulk of the class: again, the name is arbitrary, and it should be descriptive. What may be important here is to use astro_data_descriptor to decorate them. This is not required, because unlike tag methods, descriptors are meant to be called explicitly by the programmer, but they can still be earmarked (using this decorator) to be listed when calling the descriptors property. The decorator does not alter the descriptor input or output in any way, so it is always safe to use it, and you probably should, unless there’s a good reason against it (e.g., if a descriptor is deprecated and you don’t want it to show up in lookups).

    More detailed information can be found in Descriptors.

3.2.3. Register your class

Finally, you need to include your class in the AstroData Registry. This is an internal structure with a list of all the AstroData-derived classes that we want to make available for our programs. Including the classes in this registry is an important step, because a file should be opened using astrodata.open or astrodata.create, which uses the registry to identify the appropriate class (via the _matches_data methods), instead of having the user specify it explicitly.

The version of AstroData prior to DRAGONS had an auto-discovery mechanism, that explored the source tree looking for the relevant classes and other related information. This forced a fixed directory structure (because the code needed to know where to look for files), and gave the names of files and classes semantic meaning (to know which files to look into, for example). Aside from the rigidness of the scheme, this introduced all sort of inefficiencies, including an unacceptably high overhead when importing the AstroData package for the first time during execution.

In this new version of AstroData we’ve introduced a more manageable scheme, that places the discovery responsibility on the programmer. A typical __init__.py file on an instrument package will look like this:

__all__ = ['AstroDataMyInstrument']

from astrodata import factory
from .adclass import AstroDataMyInstrument


The call to factory.addClass is the one registering the class. This step needs to be done before the class can be used effectively in the AstroData system. Placing the registration step in the __init__.py file is convenient, because importing the package will be enough!

Thus, a script making use of DRAGONS’ AstroData to manipulate GMOS data could start like this:

import astrodata
from gemini_instruments import gmos


ad = astrodata.open(some_file)

The first import line is not needed, technically, because the gmos package will import it too, anyway, but we’ll probably need the astrodata package in the namespace anyway, and it’s always better to be explicit. Our typical DRAGONS scripts and modules start like this, instead:

import astrodata
import gemini_instruments

gemini_instruments imports all the packages under it, making knowledge about all Gemini instruments available for the script, which is perfect for a multi-instrument pipeline, for example. Loading all the instrument classes is not typically a burden on memory, though, so it’s easier for everyone to take the more general approach. It also makes things easier on the end user, because they won’t need to know internal details of our packages (like their naming scheme). We suggest this “cascade import” scheme for all new source trees, letting the user decide which level of detail they need.

As an additional step, the __init__.py file in a package may do extra initialization. For example, for the Gemini modules, one piece of functionality that is shared across instruments is a descriptor that translates a filter’s name (say “u” or “FeII”) to its central wavelength (e.g., 0.35µm, 1.644µm). As it is a rather common function for us, it is implemented by AstroDataGemini. This class does not know about its daughter classes, though, meaning that it cannot know about the filters offered by their instruments. Instead, we offer a function that can be used to update the filter → wavelength mapping in gemini_instruments.gemini.lookup so that it is accessible by the AstroDataGemini-level descriptor. So our gmos.__init__.py looks like this:

__all__ = ['AstroDataGmos']

from astrodata import factory
from ..gemini import addInstrumentFilterWavelengths
from .adclass import AstroDataGmos
from .lookup import filter_wavelengths

# Use the generic GMOS name for both GMOS-N and GMOS-S
addInstrumentFilterWavelengths('GMOS', filter_wavelengths)

where addInstrumentfilterWavelengths is provided by the gemini package to perform the update in a controlled way.

We encourage package maintainers and creators to follow such explicit initialization methods, driven by the modules that add functionality themselves, as opposed to active discovery methods on the core code. This favors decoupling between modules, which is generally a good idea.


[1]Note that the keyword dictionary is a “private” property of the class (due to the double-underscore prefix). Each class can define its own set, which will not be replaced by derivative classes. _keyword_for is aware of this and will look up each class up the inheritance chain, in turn, when looking up for keywords.
[2]Notice that the example functions will return only a TagSet, if appropriate. This is OK, remember that every function in Python returns a value, which will be None, implicitly, if you don’t specify otherwise.
[3]An EXTVER keyword is not formally required as the FitsLoader.load method will assign the lowest available integer to a SCI header with no EXTVER keyword (or if its value is -1). But we wish to be able to identify the original im1 header by assigning it an EXTVER of 1, etc.