Vecta

A computational maths library for ML, built from scratch in C++ with Python bindings. The long-term goal is to work chapter-by-chapter through Mathematics for Machine Learning (Deisenroth, Faisal, Ong), implementing every concept myself instead of reaching for a library.

Current state

The linear-algebra core is up and running as a header-only C++ template library, exposed to Python through a single pybind11 module:

• Matrix<T> — row-major dense matrix with (i, j) access
• matmul, add, scalar_mul — arithmetic (with dimension checking)
• transpose, identity, is_symmetric

C++ and Python test suites are planned but not written yet.

Stack

• C++17 header-only core (no external maths dependencies — everything implemented from scratch)
• pybind11 for Python bindings (fetched via CMake FetchContent)
• scikit-build-core + CMake for packaging
• uv for the Python venv and all package installs (no pip commands)

Layout

include/vecta/linalg/        C++ headers (header-only core)
    matrix.hpp               Matrix<T> class
    ops.hpp                  matmul, add, scalar_mul, transpose, identity, is_symmetric
src/bindings.cpp             pybind11 binding code (module `_vecta`)
python/vecta/                Python package wrapper (imports `_vecta`)
examples/                    scratch space for worked examples from the book
tests/cpp/                   Catch2 tests (planned)
tests/python/                pytest tests (planned)

Dev workflow (uv-based, no pip)

Create the venv and install the package in editable mode:

uv venv
source .venv/bin/activate
uv pip install -e . --no-build-isolation

Rebuild + reinstall the Python module after changing C++ code:

uv pip install -e . --no-build-isolation

Try it from Python:

from vecta import _vecta as v

A = v.Matrix(2, 2)
A.set(0, 0, 1); A.set(1, 1, 1)
print(v.matmul(A, v.identity(2)))
print(v.is_symmetric(A))

Roadmap

Follow the book chapter by chapter, folding each topic into include/vecta/<chapter>/, src/bindings.cpp, and python/vecta/: