Shatter

Data driven programming; input data and output nice functional code ;)

Introduction

This is a python 3.6+ project that uses algorithms to transform a set of conditions into functional python code. See some examples.

Package Setup

Install:

    $ pip install shatter

Dependencies

pyeda==0.28.0
pandas==0.19.1
sympy==1.1
Keras==2.0.6
numpy==1.13.1
pip==9.0.1
scikit_learn==0.19.0

Examples

Get Started

Copy paste this snippet:

from shatter.solver import Rules


def my_func(a, b):
    pass

r = Rules(a=True, b=True, output=True)
r.solve(my_func)

Run it and see how my_func code changes from pass to return a and b. We just specified that when a and b are true then the output should be True, that is equivalent to the logical and operator.

We can add further conditions and shatter will compute the optimal function to get there.

Adding more conditions

Now we add 2 additional conditions with r.add():

from shatter.solver import Rules


def my_func(a, b):
    pass


r = Rules(a=True, b=True, output=True)
r.add(a=False, b=True, output=True)
r.add(a=True, b=False, output=True)
r.solve(my_func)

In this case the solution is a or b.

If conditionals

What if the output for a given logical condition is not a boolean? In that case a programmer would use an if. In the next example this package solves this case:

Change output to 1:

from shatter.solver import Rules


def my_func(a, b):
    pass

r = Rules(a=True, b=True, output=1)
r.solve(my_func)

The solution will be:

def my_func(a, b):

    if a and b:
        return 1

    return False

Returns 1 or False otherwise.

Adding pieces of code

Say you want to add a arbitrary piece of code that evaluates to boolean, then:

from shatter.solver import Rules, Code


def any_code(a):
    pass

r = Rules(condition=Code(code_str='isinstance(a, str)'), output=2)
r.solve(any_code)

The result should be:

def internal_code(a):

    if isinstance(a, str):
        return 2

    return False

Here the piece of code isinstance(a, str) was added as the if condition to output 2

Iteration

Run this code:

from shatter.solver import Rules, Code, Output


def recursive(a):
    pass

a = Code()
args = {'a': a + 1}
out = Output(function=recursive, arguments=args)

r = Rules(stopping_condition=  a > 2, output=a, default=out)
solution = r.solve(recursive)

The result this time will be a recursive counting function :)

def recursive(a):

    if a > 2:
        return a

    return recursive(a + 1)

With a = Code() variable a is initialized as a code piece. Then with

args = {'a': a + 1}

A dictionary for the inputs of the recursive function is declared. Those inputs are fed into a Output object:

out = Output(function=recursive, arguments=args)

After out is passed via default keyword when initializing the Rules object. This default keyword is used to override the last return statement of the recursive function.

Solve Small ML problem

Copy paste this snippet:

import pandas as pd
from sklearn import datasets
from shatter.solver import Rules, solve


@solve()
def solve_iris(x1, x2, x3, x4):
    pass


iris = datasets.load_iris()

x = iris.data
y = iris.target

data_frame = pd.DataFrame(x, columns=['x1', 'x2', 'x3', 'x4'])

# Make binary and add to df
data_frame['output'] = [int(bool(e)) for e in y]


print(data_frame)

r = Rules(data_frame)

solution = r.solve(solve_iris)

Outputs:

def solve_iris():
    return x3 >= 2.45

Going deeper

Setup

Clone repository:

`git clone git@github.com:jisazaTappsi/shatter.git`

More examples

See examples.

How does shatter work?

Takes a function and a truth table which is processed using the Quine-McCluskey Algorithm. Then finds an optimal boolean expression. This expression is inserted in the method definition.

Rules Class

Is initialized with one rule. Other rules can be added with Rules.add() method. To generate the solution call Rules.solve() method.

Each rule

The arguments of each rule are specified as optional arguments inside a Rules constructor or inside a Rules.add() call. There are reserved keywords:

output: Determines the value to be returned when the given condition is True.

output_args: Dictionary with the values for the arguments when output is a function.

default: Value returned when non of the rules are True.

Arguments of Rules.solve()

• function: passed as a callable. This function is going to be filled with the solution to the present task.

• unittest=None: Test Case to be able to run and test the code generated each time the test runs. See example for a deeper understanding.

Output Class

solver.Output: Class that helps define a function with arguments as an output. Has fields:

• function: A callable object.
• arguments Dictionary with the function inputs.

Code class

solver.Code: Class that helps represent pieces of code. The code is fed as a string (with optional argument str_code) or it can be declared as variables. eg:

from shatter.solver import Code

a = Code()
b = Code()
print(a > b)

This will literally print the code a > b rather than the objects or any result.

Solution class

solver.Solution: Class that contains the solution of the problem it includes:

• rules: The information given by the user.
• implementation: Plain code.
• ast: Abstract syntax tree