Ein is a pointful array language embedded in Python. It is compiled into efficient point-free whole-array operations as present in libraries like NumPy.
The ein library forms Ein's frontend in Python, with a lean API focused on array comprehensions constructed via the array function. It has NumPy, PyTorch, and JAX backends, and can thus be integrated with existing code.
To simply install the ein package into pip (which should work without poetry), use pip install . (where . is assumed to be the repository's root directory).
Alternatively, use pip install git+https://github.com/jbachurski/ein.git (once the repository is public).
Ein might also be published on PyPI in the future.
To check the basics work, in your Python REPL of choice try:
In Ein, you can use array comprehensions, defined via the array function. Ein defers computations -- when you're done, use the .eval() method to evaluate. Here's how you can compute a triangular mask of given size:
from ein import array, where
print(array(
lambda i, j: where(i <= j, 1.0, 0.0),
size=(3, 4)
).eval())
# -> [[1. 1. 1. 1.]
# [0. 1. 1. 1.]
# [0. 0. 1. 1.]]If you would like to involve existing numpy.ndarrays in computations, use wrap and then evaluate as usual. Here we compute the outer product of a pair of vectors:
import numpy
from ein import array, wrap
u, v = wrap(numpy.array([1., 2.])), wrap(numpy.array([-0.5, 0.0, 0.5]))
print(array(lambda i, j: u[i] * v[j]).eval())
# [[-0.5 0. 0.5]
# [-1. 0. 1. ]]Ein is strongly typed and in contrast in NumPy it does not perform type promotion, which can cause type errors unwillingly.
from ein import array
array(lambda i: i, size=5).eval()
# -> array([0, 1, 2, 3, 4])Environment. Ein uses Poetry for project configuration and managing environments. To set up your development environment, use poetry install, at which point you can use poetry shell to activate the virtual environment. For reproducibility, a poetry.lock file was used to keep track of dependencies - but this can be adapted accordingly.
Tests. The repository uses the pytest test framework. In a Poetry environment (spawned by poetry shell), you may use pytest tests to run the test suite.
Pre-commit hooks. If you wish to use the same suite of linters, configure the pre-commit hooks, as configured by .pre-commit-config.yaml. This can be done by executing pre-commit install, which sets up a dedicated environment (this might take a minute). Then use pre-commit run --all-files to run them manually.
There is currently no full documentation, so this section briefly explains the main functions available in ein.
Types T allowed and returned by Ein functions are given by:
T ::= Scalar (scalars)
| Vec[T] (vectors)
| tuple[T, ...] | dict[str, T] | DataclassT (records)
Scalars have operator overloading and various numeric methods. For instance, a + b is addition, and a.float() converts and integer to a float. Python ints and floats are automatically promoted to an Ein Scalar.
Vectors Vec[T] can be indexed, returning an instance of (exactly) T. For instance, indexing into a vector of tuples (and records generally) Vec[tuple[Scalar, Scalar]] actually returns a Python tuple.
For record types, DataclassT is constrained to just frozen dataclasses of other Ein values of T. However, these may have custom methods implemented, and will be reconstructed accordingly.
Array comprehensions, introduced with array, are used to construct vectors (type Vec).
def array(f: (Int) -> T, size: Int | None = None) -> Vec[T]: ...
array takes a function f, such that the element at index i is given by f(i). For convenience, f can take multiple arguments, in which case the array defines multiple dimensions.
from ein import array
array(lambda i: i, size=5).eval()
# -> array([0, 1, 2, 3, 4])
array(lambda i, j: 3*i + j, size=(2, 3)).eval()
# -> array([[0, 1, 2],
# [3, 4, 5]])If size is None, Ein attempts to infer it from sites where the introduced index is used to directly index into an array. If the defined array has multiple dimensions, a tuple of sizes may be provided. Ein supports dynamically-sized arrays (as long as sizes do not depend on an index, which would lead to jagged arrays), so besides usual ints Ein integers can also be passed to size.
(Indexed) folds are available via a simple for loop-like primitive. Folds don't iterate over an existing vector, and instead over some range [0, 1, ..., count - 1].
def fold(init: T, step: (Int, T) -> T, count: Int | None = None) -> T: ...
Similarly to array, folds can also perform size inference (in this case to assign a value to count).
from ein import array, fold
print(fold(1, lambda i, acc: acc * (i + 1), count=5).eval())
# -> 120 == 5!
print(fold(1., lambda i, acc: (i + 1).float() ** acc, count=3).eval())
# -> 9.0 == 3 ** (2 ** 1)
a = array(lambda i: i, size=5)
print(fold(0, lambda i, acc: acc + a[i]).eval())
# -> 10 == 0 + 1 + 2 + 3 + 4Wrapping introduces existing arrays into the Ein program. This can also be used with torch.Tensors and jax.Arrays when using the appropriate backends.
def wrap(x: numpy.ndarray | ...) -> T: ...
def eval() -> T: ...
By default, Ein evaluates using its NumPy backend. There are also specialised methods - .numpy(), .torch(), and .jax() - for calling specific backends.
A function can be decorated with @function, in which case it automatically wraps its arguments and evals the returned value.
Vectors can be reduced with custom reductions, which can be much faster than a fold.
def reduce(vec: Vec[T], ident: T, cat: (T, T) -> T) -> T: ...
The arguments (ident, cat) must form a monoid (cat must be associative, with an identity ident).
For instance:
import numpy
from ein import array, wrap
a = wrap(numpy.array([3, 6, -1, 5]))
print(a.reduce(0, lambda x, y: x + y).eval())
# -> 13 == 3 + 6 - 1 + 5This also works with records:
import numpy
from ein import array, wrap, where
a = wrap(numpy.array([3.0, 6.0, -1.0, 5.0]))
value, index = array(
lambda i: {"value": a[i], "index": i}
).reduce({"value": float('inf'), "index": 0}, lambda x, y: where(
x["value"] < y["value"],
x, y
)).values()
print(value.eval(), index.eval())
# -> -1.0 2 as min(a) == -1.0 / argmin(a) == 2Ein was developed as part of my Bachelor's Thesis research at the University of Cambridge. It resulted in a publication in the ARRAY 2024 proceedings: "Points for Free: Embedding Pointful Array Programming in Python". Ein was also recognised in the POPL 2024 SRC and the 2024 ACM SRC Grand Finals.
If you use Ein or the techniques developed in it, use this citation:
@inproceedings{bachurski2024points,
title={Points for Free: Embedding Pointful Array Programming in Python},
author={Bachurski, Jakub and Mycroft, Alan},
booktitle={Proceedings of the 10th ACM SIGPLAN International Workshop on Libraries, Languages and Compilers for Array Programming},
pages={1--12},
year={2024}
}