pyfop 0.2.0

A forward-oriented programming paradigm for Python.

pyfop

A novel forward-oriented programming paradigm for Python.

Dependencies: None
Developer: Emmanouil (Manios) Krasanakis
Contant: [e-mail]

About

pyfop is a package that introduces the concept of forward-oriented programming in Python. This aims to simplify development by sharing parameters across multiple components and defining those after main business logic definition.

Features

• Simplified code that focuses on business logic.
• Value sharing between arguments.
• Non-intrusive API (minimal changes to source code).
• Priority-based conflict resolution.
• Scoped method modification.
• Cached optimization.

Quickstart

Overall, there are three steps to using the library:

1. wrapping some components with a lazy execution decorator
2. assigning some arguments of these components as aspects
3. calling components

To see these in action, let us create a system where we transform (e.g. normalize) numpy arrays and then compare them with known data mining measures. We will make this system modular by allowing combination of various transformation and comparison components.

First, we define a couple of single-input array transformation methods tautology and normalize, as well as two pairwise array comparison methods dot and KLdivergence. In addition to array inputs, some of these methods also make use of optional parameter values, such as norm to indicate the type of normalization and epsilon to offset division with or logarithms of zero.

We make arguments share-able by name between objects by wrapping their default values with the @pyfop.Aspect class. For example, if normalize and KLdivergence are used together in the same call, their norm argument would obtain the same value. This value is either determined through the priority defaults (the package would throw an error if the same priorities tried to set different values with the same priorities) and can be customized during calls.

To parse aspects as values, we also need to set up our methods for lazy execution required by the package. This is achieved by adding a @pyfop.lazy decorator.

import pyfop as pfp
import numpy as np

@pfp.lazy
def tautology(x):
    return x

@pfp.lazy
def normalize(x, norm=pfp.Aspect(2)):
    return x / (np.sum(x**norm))**(1./norm)

@pfp.lazy
def dot(x, y):
    return np.sum(x*y)

@pfp.lazy
def KLdivergence(x, y, 
                 norm=pfp.Aspect(1, priority=pfp.Priority.INCREASED), 
                 epsilon=pfp.Aspect(np.finfo(float).eps)):
    if norm != 1:
        raise Exception("KLDivergence should not work on non-L1 normalizations")
    return np.sum(-x*np.log(x/(y+epsilon)+epsilon))

We finally bring together various normalization and comparison strategies in the following class. This stores lazy execution methods as well as any additional keyword arguments kwargs to be used for aspect values. Then, when comparing arrays, it runs the lazy execution with these arguments.

class Comparator :
    def __init__(self, transform, measure, **kwargs):
        self.transform = transform
        self.measure = measure
        self.kwargs = kwargs

    def __call__(self, x, y):
        transformed_x = self.transform(x)
        transformed_y = self.transform(y)
        return self.measure(transformed_x, transformed_y).call(**self.kwargs)

For example, we can write the following expression to compute the cosine similarity between two arrays.

x = np.array([1., 1., 1.])
y = np.array([1., 1., 1.])
print(Comparator(normalize, dot, norm=2)(x, y))

If we did not provide a norm argument to the constructor to be eventually passed to lazy execution, the first default value would be inferred (in this case, norm=2 based on the default of normalization).

This default can change depending on what is being executed. For example, the following code automatically infers norm=1 based on priority conflict resolution.

print(Comparator(normalize, KLdivergence, epsilon=0)(x, y))

pyfop makes error checking trivial; we just needed to add the normalization aspect to KLdivergence and check for the shared value. For example, adding a norm=2 argument to the previous command will throw an error. There is no need for conditional checks at other parts of the code.

Functionalities

Making a method lazily execute can be achieved with the @pyfop.lazy decorator. Aspect variables are assigned as pfp.Aspect variables. These can have a default value. Aspect values can change after lazy methods are first called.

import pyfop as pfp

@pfp.lazy
def increase(x, inc=pfp.Aspect(1)):
    return x + inc

y = increase(2)
assert y.call() == 3
assert y.call(inc=2) == 4

For minimal intrusiveness, a @pyfop.aytoaspects is provided can turn all default arguments into aspects. In the above snippet, the method definition could change to the one bellow. Note that lazy decorators should remain the topmost ones.

@pfp.lazy
@pfp.autoaspects
def increase(x, inc=1):
    return x + inc

Caching (also known as memoization) is automatically supported to prevent lazy calls from re-running for the exact same inputs. It is based on object identifiers, but it prevents the garbage collector from running on past method inputs. To clear the garbage collector after operations, these can run withing a cached scope, which is available in the form of the context. This can be achieved per the following code:

import pyfop as pfp

@pfp.eager
@pfp.autoaspects
def zeros(length=10):
    return [0] * length

with pfp.CacheScope():
    id1 = id(zeros(10))
    assert id1 == id(zeros(10))

assert id1 != id(zeros(10))

In the above example, the @pyfop.eager decorator defines immediately runnable methods that support lazy execution arguments. Calling methods decorated this way is equivalent to calling the call() method immediately. All aspects should somehow obtain values at least once, either via defaults or through normal pythonic argument parsing.

import pyfop as pfp

@pfp.lazy
def add(x, permutation=pfp.Aspect()):
    return x + permutation

@pfp.eager
def mult(x, permutation=pfp.Aspect()):
    return x * permutation

assert mult(add(1, 3)) == 12
assert mult(add(1), 3) == 12