Sparse vectors optimized for memory and NumPy integrations.

numpy handles densely populated n-dimemsional arrays. scipy.sparse handles sparsely populated 2-dimensional arrays, i.e., matrices. What's missing from the ecosystem is sparsely populated 1-dimensional arrays, i.e., vectors.

NumPy	Python	Spector
1-dim bool numpy.array	set[int]	spector.indices
1-dim float numpy.array	dict[int, float]	spector.vector
scipy.sparse.dok_matrix	dict[int, dict[int, float]]	spector.matrix

Indices and vectors are implemented in Cython as hash sets and maps. All native operations are optimized and release the GIL.

• conversion between sparse numpy arrays
• conversion between dense numpy arrays
• binary set operations
• binary math operations
• map, filter, and reduce operations with numpy universal functions

Usage

indices

A sparse boolean array with a set interface.

>>> from spector import indices
>>> ind = indices([0, 2])
>>> ind
indices([2 0])
>>> 1 in ind
False
>>> ind.add(1)
True
>>> ind.todense()
array([ True,  True,  True])
>>> ind.fromdense(_)
indices([2 1 0])

vector

A sparse float array with a mapping interface.

>>> from spector import vector
>>> vec = vector({0: 1.0, 2: 2.0, 4: 1.0})
>>> vec
vector([4 2 0], [1. 2. 1.])
>>> vec[2] += 1.0
>>> vec[2]
3.0
>>> vec.todense()
array([1., 0., 3., 0., 1.])
>>> vector.fromdense(_)
vector([4 2 0], [1. 3. 1.])
>>> vec.sum()
5.0
>>> vec + vec
vector([0 2 4], [2. 6. 2.])

Vectors support math operations with scalars, and with vectors if the set method is unambiguous.

vector operation	set method	ufunc
+	union	add
*	intersection	multiply
-		subtract
/		true_divide
**		power
|	union	max
&	intersection	min
^	symmetric_difference
difference	difference

matrix

A mapping of keys to vectors.

>>> from spector import matrix
>>> mat = matrix({0: {1: 2.0}})
>>> mat
matrix(<class 'spector.vector.vector'>, {0: vector([1], [2.])})
>>> mat.row, mat.col, mat.data
(array([0]), array([1]), array([2.]))

Installation

% pip install spector

Tests

100% branch coverage.

% pytest [--cov]