Memory efficient stack of multiple 2D sparse arrays.

Installation

Requirements

Python 3.8 or higher

Pip Install

Simply install using pip: pip install sparsestack

First code example

import numpy as np
from sparsestack import StackedSparseArray

# Create some fake data
scores1 = np.random.random((12, 10))
scores1[scores1 < 0.9] = 0  # make "sparse"
scores2 = np.random.random((12, 10))
scores2[scores2 < 0.75] = 0  # make "sparse"
sparsestack = StackedSparseArray(12, 10)
sparsestack.add_dense_matrix(scores1, "scores_1")

# Add second scores and filter
sparsestack.add_dense_matrix(scores2, "scores_2", join_type="left")

# Scores can be accessed using (limited) slicing capabilities
sparsestack[3, 4]  # => scores_1 and scores_2 at position row=3, col=4
sparsestack[3, :]  # => tuple with row, col, scores for all entries in row=3
sparsestack[:, 2]  # => tuple with row, col, scores for all entries in col=2
sparsestack[3, :, 0]  # => tuple with row, col, scores_1 for all entries in row=3
sparsestack[3, :, "scores_1"]  # => same as the one before

# Scores can also be converted to a dense numpy array:
scores2_after_merge = sparsestack.to_array("scores_2")

Adding data to a sparsestack-array

Sparsestack provides three options to add data to a new layer.

1. .add_dense_matrix(input_array) Can be used to add all none-zero elements of input_array to the sparsestack. Depending on the chosen join_type either all such values will be added (join_type="outer" or join_type="right"), or only those which are already present in underlying layers ("left" or "inner" join).
2. .add_sparse_matrix(input_coo_matrix) This method will expect a COO-style matrix (e.g. scipy) which has attributes .row, .col and .data. The join type can again be specified using join_type.
3. .add_sparse_data(row, col, data) This essentially does the same as .add_sparse_matrix(input_coo_matrix) but might in some cases be a bit more flexible because row, col and data are separate input arguments.

Accessing data from sparsestack-array

The collected sparse data can be accessed in multiple ways.

1. Slicing. sparsestack allows multiple types of slicing (see also code example above).

sparsestack[3, 4]  # => tuple with all scores at position row=3, col=4
sparsestack[3, :]  # => tuple with row, col, scores for all entries in row=3
sparsestack[:, 2]  # => tuple with row, col, scores for all entries in col=2
sparsestack[3, :, 0]  # => tuple with row, col, scores_1 for all entries in row=3
sparsestack[3, :, "scores_1"]  # => same as the one before

2. .to_array() Creates and returns a dense numpy array of size .shape. Can also be used to create a dense numpy array of only a single layer when used like .to_array(name="layerX").
   Carefull: Obviously by converting to a dense array, the sparse nature will be lost and all empty positions in the stack will be filled with zeros.
3. .to_coo(name="layerX") Returns a scipy sparse COO-matrix of the specified layer.