Python

LazyMatrix for Python applies column normalization during matrix operations. It accepts NumPy arrays and SciPy CSR/CSC matrices and sparse arrays. Install it with:

pip install lazymatrix

Python 3.11 or newer is required. Prebuilt wheels cover Linux x86_64, macOS on Intel and Apple silicon, and Windows x86_64. Building from source requires Rust 1.89 or newer.

Normalize and multiply

import numpy as np
from scipy import sparse
from lazymatrix import standardize

X = sparse.csr_matrix([[1.0, 0.0], [2.0, 3.0], [0.0, 4.0]])
Z = standardize(X)

beta = np.array([1.0, -1.0])
prediction = Z @ beta
gradient = Z.T @ prediction
column = Z[:, 0]

standardize fits column means and population standard deviations immediately. Pass center=False or scale=False to disable either component. Exact zero scales become one, so constant columns do not cause division by zero. Nonfinite values follow ordinary IEEE arithmetic; missing values are not ignored.

Z.shape, Z.ndim, and Z.dtype describe the logical matrix. Z.centers and Z.scales return copies of the fitted parameters, or None when the component is disabled. These parameters describe the original columns even for Z.T.

Both @ and np.matmul support dense vectors and matrices on either side. Multiple right-hand sides currently use repeated vector products. Z[:, j] returns one normalized column as an independent dense array, including support for negative column indices. Other slicing and elementwise operations are not part of the initial API.

Use SciPy solvers

The adapter exposes forward and transpose products without materializing the normalized input:

from scipy.sparse.linalg import lsqr

y = np.array([1.0, -2.0, 1.0])
coefficients = lsqr(Z.aslinearoperator(), y)[0]

Memory and conversion

LazyMatrix owns a copy of the raw input, so later changes to X do not change Z. Dense inputs require a full raw-data copy. Sparse copies remain sparse and are stored in canonical CSC form, with duplicates summed before normalization. All values and results use float64. Complex and nonnumeric inputs are rejected.

Products, transposes, and column reads do not construct a full normalized matrix. Call Z.toarray() when you explicitly want a dense copy:

dense = Z.toarray()

Implicit conversion with np.asarray(Z) raises an error. Existing estimators that require a concrete NumPy array or SciPy sparse matrix need an explicit conversion or their own operator integration.

The initial Python API fits each input independently. Reusing training normalization on new observations is planned; calling standardize on validation data computes new statistics.

See the Python API reference for public classes and methods.

Development and releases

From the repository root:

task python:check
task python:package

The first command builds an editable extension and runs formatting, linting, and correctness checks. The second builds a wheel and source distribution, installs the wheel in an isolated environment, rebuilds from the source distribution, and tests both installations. The environment lives in .devenv/python.

Python and Rust use independent versions and share one Versionary release PR. Rust releases also release Python bindings. Python-only releases leave the Rust version unchanged. Python tags use python-v<version> and Rust tags use v<version>.