I noticed that over the years I’ve adopted a more functional style of programming in Python. If anything because it’s easier to write tests.
What are good arguments in favour of using classes instead of functions?
why not classes
Eric Ma replies
Mostly to store data, I think. I pretty much only use classes for:
- Data classes,
- More conveniently parameterised functions (Equinox/PyTorch-style).
Eric uses objects for holding data, and functions for processing the data.
Point 2 refers to implementing configurable functions using a class with the __call__ function. This allows the pattern often used in neural network libraries:
layer = Dense(in_dims, out_dims)
out = layer(x)Here, the in_dims and out_dims parameters are attached to the object directly as part of the initialisation. This is useful when we configure reusable data associated with the function – here, the parameters of the layer.
Patrick Kidger adds
Classes are fine, just don’t mutate them :) This still counts as “functional”. Basically, use frozen dataclasses.
Classes:
- are containers for data
- provide abstract interfaces (=single dispatch)
- offer namespacing
Patrick’s Equinox library is designed around the base class eqx.Module, where every subclass is either abstract (it can be subclassed, but not instantiated); or final (it can be instantiated, but not subclassed). This means the __init__ method, and all dataclass fields, are defined once in one class.
Everything is abstract until the final class.
This promotes composing classes over class hierarchies. We never use super().__init__().
Using lots of classes could lead to spaghetti code, but using the abstract/final design pattern helps avoid this because __init__ always makes it clear when the class is final.
why classes
It is sometimes useful to put methods on a class to aid discoverability Think PyTorch tensor.pow(2).mean(dim=1).sqrt() vs NumPy np.sqrt(np.mean(array**2, axis=1)) (although a pipe operator could help here).
Classes are easily extensible.
For non-production data science work, it’s useful to have the data attached when doing stuff interactively and be able to tab complete.
summary
Use (data)classes for holding data, and functions for processing the data.
Coming from a maths background, functions have always clearer in my head as a way of transforming data over OOP, so I’ll try and implement this pattern wherever possible.