Effectful

Effectful is an algebraic effect system for Python, intended for use in the implementation of probabilistic programming languages. It is a core component of the ChiRho causal modeling language.

Installation

Install From Source

git clone git@github.com:BasisResearch/effectful.git
cd effectful
git checkout master
pip install -e .[pyro]

Install With Optional PyTorch/Pyro Support

effectful has optional support for:

• PyTorch (tensors with named dimensions)

• Pyro (wrappers for Pyro effects)

• Jax (tensors with named dimensions)

• Numpyro (operations for Numpyro distributions)

To enable PyTorch support:

pip install effectful[torch]

Pyro support (which includes PyTorch support):

pip install effectful[pyro]

Jax support:

pip install effectful[jax]

Numpyro support (which includes Jax support):

pip install effectful[numpyro]

Getting Started

Here’s an example demonstrating how effectful can be used to implement a simple DSL that performs arithmetic on terms with free variables.

import functools

from effectful.ops.types import Term
from effectful.ops.syntax import defdata, defop
from effectful.ops.semantics import handler, evaluate, coproduct, fwd

add = defdata.dispatch(int).__add__

def beta_add(x: int, y: int) -> int:
    match x, y:
        case int(), int():
            return x + y
        case _:
            return fwd()

def commute_add(x: int, y: int) -> int:
    match x, y:
        case Term(), int():
            return y + x
        case _:
            return fwd()

def assoc_add(x: int, y: int) -> int:
    match x, y:
        case _, Term(op, (a, b)) if op == add:
            return (x + a) + b
        case _:
            return fwd()

beta_rules = {add: beta_add}
commute_rules = {add: commute_add}
assoc_rules = {add: assoc_add}

eager_mixed = functools.reduce(coproduct, (beta_rules, commute_rules, assoc_rules))

We can represent free variables as operations with no arguments, generated using defop:

>>> x = defop(int, name="x")
>>> y = defop(int, name="y")

If we evaluate an expression containing free variables, we get a term:

>>> e = 1 + 1 + (x() + 1) + (5 + y())
>>> print(e)
add(2, add(add(x(), 1), add(5, y())))

We can make the evaluation strategy smarter by taking advantage of the commutativity and associativity of addition, as expressed by the commute_add and assoc_add handlers.

>>> with handler(eager_mixed):
>>>     print(evaluate(e))
add(8, add(x(), y()))

Learn More

More examples and API documentation can be found in the docs.