A signature pattern
There are many libraries around to specify what I call a ‘signature’ for an object, in other words a list of attributes that define its parameter set. I have heavily used Enthought’s Traits library for this purpose, but the concept is fairly general and can be found eg in ORMs (Object Relational Mappers) or web frameworks.
Specification of this interface of parameters may be used to answer a variety of needs:
- Typing: in the case of an ORM, to generate UIs, or for better error management, it may be desirable to have some control on the types of certain attributes of an object. In this case, specifying the signature corresponds to laying out a data model for the object.
- Reactive programming: using properties to react to changes to attributes, one can fully specify the API of an object in terms of these attributes. This gives a message-passing like programming style that can be very well suited to parallel-computing in particular because it can easily be made thread-safe.
Signatures for statistical learning objects
Recently, I considered the signature pattern in a new context. In the scikit-learn, we are interested in statistical learning. This entails fitting models to data and often tuning parameters to select a model that fits best (a problem called model selection). Each of our models is an object that implements a couple of key methods to fit to the data and to apply to new data (fit and predict).
The approach that we are currently taking for model selection is (more or less) to generate a list of models with different parameters and fit and test them on the data.
A very nice feature would be to find out the parameters to vary simply by inspecting the objects, and such a desire recently got us discussing of defining signatures for our objects. I must confess that I am a bit weary as this means either depending on a signature library, or building one. We don’t want to grow our dependencies, and most signature-definition code that I know involve meta-programming tricks to avoid code duplication.
Solving the simple problem: avoiding type checking
Today, I had to bite the bullet, because we were in a situation in which we had to instantiate new models from the existing one during model selection. For technical reasons, using a copy.copy to create these new models was not a great idea, and it was better to have the minimal list of parameters required. Here come signatures again.
After a bit of messing around with the code, I realized that typing information was useless, and most probably harmful, to our immediate goals and that I just needed the names of the relevant attributes. I finally settled down to the following solution (which might still change):
- All parameters need to be specified as keyword arguments of the __init__. The __init__ may not have positional arguments or ‘*’ arguments. Attributes on the objects have the same names as the __init__ parameters.
- A simple base class, with couple of methods relying on a simple use of the inspect module to find the signature of the __init__.
class BaseEstimator(object): @classmethod def _get_param_names(cls): args, varargs, kw, default = inspect.getargspec(cls.__init__) assert varargs is None, ( 'scikit learn estimators should always specify their ' 'parameters in the signature of their init (no varargs).' ) # Remove 'self' args.pop(0) return args def _get_params(self): out = dict() for key in self._get_param_names(): out[key] = getattr(self, key) return out def _set_params(self, **params): valid_params = self._get_param_names() for key, value in params.iteritems(): assert key in valid_params, ('Invalid parameter %s ' 'for estimator %s' % (key, self.__class__.__name__)) setattr(self, key, value)
The full code can be seen here and adds a bit more features, such as a clever __repr__.
What I like about this solution is that it (almost) does not use metaprograming, and avoids code duplication without forcing any specific pattern on the developer subclassing BaseEstimator.
The next step
This approach solves my immediate problem, but not the bigger one of finding what values can the different parameters take when varied for model selection. Of course this second problem is much more complicated, and maybe it is not worth solving it: the framework could very easily be bringing in more problems than it solves.
However, it seems that a fairly easy way of specifying possible values for parameters would be to decorate the __init__, giving the possible parameters to be tested during the model selection:
@cv_params(l1=np.logspace(1e-4, 1, 10)) def __init__(self, l1=.5, fit_intercept=True) # ...
All the decorator has to do is to store the information in an attribute attached to the __init__ (and probably to check that the parameters it was given are valid arguments, in order to raise errors early). Methods on the class can later inspect this information for model selection, or GUI building (data-model specification will probably require some typing language, rather than a simple list of possible parameters).
Once again, here we would be avoiding the difficulty of specifying type information in a non restrictive way, but avoiding a problem that we don’t have to solve is probably a good idea.