I’ve recently taken the monumental undertaking of refactoring my cosmology code, COFFE, and building a semi-decent, user-friendly Python wrapper for it in Cython. In the first iteration of it, I’ve decided to just shove all of the parameters (which the C version reads from a parameter file) into a single class (cue Pylint complaining about too many attributes), and make them Python properties, so we could do something like this:

cdef class Coffe:
    cdef double _omega_cdm
    def __init__(self):
        # stuff here
        self._omega_cdm = 0.25
        # more stuff here

    @property
    def omega_cdm(self):
        return self._omega_cdm

    @omega_cdm.setter
    def omega_cdm(self, value):
        # type and value checking for `value`
        self._omega_cdm = value

# cythonize the module

from coffe import Coffe

cosmo = Coffe()
print(cosmo.omega_cdm) # prints 0.25
cosmo.omega_cdm = 0.27 # setting omega_cdm, the safe way

We also want a method which is similar to classy’s (the Python wrapper to CLASS) set method, say, set_parameters, which takes a dictionary, and performs the bulk setting of the parameters (while doing the type and value checking of course). This also comes in handy if we want output that’s convertible to a Pandas DataFrame object. Now, assuming we have a bunch of properties like above, how would we get a dictionary, with keys as the properties of the class, and values as, well, their values? The first idea that came to mind was to just try getattr and manually specify the parameters:

def get_properties(self):
    return {
        key : getattr(self, key) \
        for key in ('omega_cdm', 'omega_baryon', ...)
    }

Unfortunately, this doesn’t really work well for at least two reasons:

  1. we need to make sure we don’t leave out any of the parameters in our list
  2. what happens if we decide to change the names of the properties? Then we would have to rewrite the list of names every time, and we all know that naming things is one of the hardest things in computer science.

Additionally, we’re dealing with Cython, and not Python, for which this solution sadly doesn’t work.

As a start, we will need the dir builtin, which more or less provides us with a list of attributes and methods our class has:

dir(cosmo)

which outputs something like this:

['__class__',
 '__delattr__',
 '__dict__',
 '__dir__',
 '__doc__',
# more stuff here
 '_omega_cdm',
 'omega_cdm']

Now, if we look at the docs of the property builtin, we’ll see that it defines, among others, the __set__ method, so we may have some luck with hasattr:

hasattr(cosmo.omega_cdm, '__set__')

which, to my surprise, outputs False! Upon closer inspection though, this makes sense, since cosmo.omega_cdm actually returns a float, which clearly doesn’t have the __set__ method. On the other hand, type(Coffe.omega_cdm) returns getset_descriptor, so we may have better luck with it:

hasattr(Coffe.omega_cdm, '__set__')

which returns True, so we’re on the right track. The question now is though, how do we access the attribute, without knowing the actual name of the attribute (in this case, omega_cdm)? Why, with getattr of course!

hasattr(getattr(Coffe, 'omega_cdm'), '__set__')

Now we just need to go through the contents of the output of dir and we’re good to go. Going back to the implementation of get_properties, our refined version looks like this:

def get_properties(self):
    return {
        key : getattr(self, key) \
        for key in dir(self) \
        if hasattr(getattr(Coffe, key), '__set__')
    }

That Coffe is sticking out a bit like a sore thumb though, since it’s not passed to the function, and we want to make our function as generic as possible. Fortunately, Python has a solution for that as well: each object has a __class__ attribute which returns the class itself, so instead of Coffe, we may as well write self.__class__.

There’s a small problem in the above implementation: it will also return read-only properties, since the property decorator doesn’t discriminate and just defines the __set__ method for everything it’s applied to, and we want to be able to do something like the below, without raising any errors1:

p1 = Coffe()
p2 = Coffe()
# change parameters in p1
...
# we want something like this to work
p2.set_parameters(p1.get_properties())

Since setting a read-only property always raises an AttributeError, we may as well attempt to set it to its current value, and if we’re successful, the property is writable, otherwise it’s read-only, and we don’t do anything. This brings us to the final implementation:

def get_properties(self):
    properties = {
        key : getattr(self, key) \
        for key in dir(self.__class__) \
        if hasattr(getattr(self.__class__, key), '__set__')
    }
    writable = {}
    for key in properties:
        try:
            # try to set the attribute;
            # if it doesn't raise an error,
            # we add it to the new dictionary
            setattr(self, key, getattr(self, key))
            writable[key] = getattr(self, key)
        except AttributeError:
            pass
    return writable

And there we have it, a method that returns all of the writable attributes of a given class, in both Cython and pure Python (well, as long as the class in question isn’t terribly complex). Of course, to get the read-only attributes, we can just invert the try-except block in the above.

One thing to note (as pointed out in this SO answer) is that we are looping through the class (self.__class__) in the above, but checking the object (self); this is intentional, since it’s possible for an object to have additional attributes, as can be observed with this simple code:

cosmo = Coffe()
set(dir(cosmo)) - set(dir(Coffe))

which outputs {'_omega_cdm'}, since this attribute doesn’t technically exist until we create an object.

Note that if we want get_property to be a property itself, we need to change our conditional to if key != 'get_properties' and hasattr(...) in the final implementation (sadly, getting the name of a function is a bit convoluted, so we opt for specifying it as a string literal), otherwise the code would get stuck in a loop since get_properties would have a __set__ method, and getattr would call get_properties, which would have a __set__ method, and getattr would call get_properties…well, you probably get the picture :)

Footnotes

  1. at this point, we could just copy the whole object using the builtin copy module (possibly defining our custom __copy__ and __deepcopy__ methods along the way), but then we wouldn’t have had so much fun with getattr and hasattr ↩