hdlib

Hyperdimensional Computing Library for building Vector Symbolic Architectures in Python 3.

Install

It is available through pip and conda. Please, use one of the following commands to start playing with hdlib:

# Install with pip
pip install hdlib

# Install with conda
conda install -c conda-forge hdlib

Usage

The hdlib library provides two main modules, space and arithmetic. The first one contains constructors of Space and Vector objects that can be used to build vectors and the space that hosts them. The second module, called arithmetic, contains a bunch of functions to operate on vectors.

from hdlib.space import Space, Vector
from hdlib.arithmetic import bind, bundle, permute

Hyperdimensional Vectors

Vector objects in hdlib can be created through the Vector class whose constructor requires the following parameters:

There are three different ways to initialize Vector objects:

# With no spacific parameters
# This creates a random bipolar vector with size 10,000 by default
vector = Vector()

# By creating a numpy.ndarray object first
# A binary vector in this case
import numpy as np
ndarray = np.random.randint(2, size=size)
vector = Vector(vector=ndarray)

# By loading a precomputed Vector object
vector = Vector(from_file="~/vector.pkl")

Note In this last example, a Vector object is built by loading the content of a pickle file. Vector objects can be saved to pickle files with the dump() method as in the following example: vector.dump(to_file="~/vector.pkl")

Here is the list of Vector class methods:

Hyperdimensional Space

Vectors are stored into a so called hyperdimensional space that can be defined through the Space constructor that requires the following parameters:

There are two ways to initialize Space objects:

# With no specific parameters
# This creates a space that can host random bipolar vectors with size 10,000 by default
space = Space()

# By loading a precomputed Space object
space = Space(from_file="~/space.pkl")

Note In this last example, similarly to Vector objects, a Space object is built by loading the content of a pickle file. Space objects can be saved to pickle files with the dump() method as in the following example: space.dump(to_file="~/space.pkl")

Here is the list of Space class methods:

Arithmetic Operations

A Vector Symbolic Architecture (a.k.a. Hyperdimensional Computing) is composed of vectors in the hyperdimensional space and a series of arithmetic operations to manipulate vectors.

The hdlib library provides three operators under the arithmetic module: bundle, bind, and permute.

Here are the characteristics of these operators:

Contributing

Long-term discussion and bug reports are maintained via GitHub Issues, while code review is managed via GitHub Pull Requests.

Please, (i) be sure that there are no existing issues/PR concerning the same bug or improvement before opening a new issue/PR; (ii) write a clear and concise description of what the bug/PR is about; (iii) specifying the list of steps to reproduce the behavior in addition to versions and other technical details is highly recommended.