Installation¶

Requirements¶

Python 3.10–3.12
PyTorch 2.x — install a CPU or CUDA build appropriate for your machine

pinn-rk runs in float64 throughout. The residual divides differences of network outputs by the slab size k, which amplifies rounding error by 1/k; in float32 that noise can swamp the quantity being minimised.

From PyPI¶

pip install pinn-rk

That is the library and its only runtime dependencies, numpy and torch.

The optional groups are not pip extras

matplotlib, jupyter, plotly and the docs toolchain are Poetry groups, not pip extras, so pip install "pinn-rk[examples]" does not work. Install from a clone if you want the notebooks or the plotting flags.

From source¶

git clone https://github.com/DiogoRibeiro7/pinn-rk.git
cd pinn-rk
poetry install

That installs the runtime dependencies (numpy, torch) and the development group (pytest, ruff, mypy, bandit, pre-commit).

Optional dependency groups¶

Both are optional and off by default, so a plain install stays small.

poetry install --with examples   # matplotlib, jupyter, plotly -> notebooks and plots
poetry install --with docs       # mkdocs and friends -> this site

--with examples is what you need for examples/notebooks/ and for the --save-plots flag on the convergence study.

Verifying the install¶

poetry run pytest                      # full suite
poetry run python -c "import pinn_rk; print(pinn_rk.__version__)"

A quick end-to-end check that exercises the actual numerics rather than just the import:

poetry run python examples/04_convergence_study.py --methods radau3

This runs in seconds and prints measured convergence rates. If the fitted orders come out near the theoretical values, the installation is sound.

Development setup¶

poetry install
poetry run pre-commit install

Before opening a pull request:

poetry run ruff check .
poetry run ruff format --check .
poetry run mypy src
poetry run pytest

CI runs exactly these on Linux, macOS and Windows across Python 3.10–3.12.

A note on Windows¶

Installing PyTorch into a deeply nested virtualenv path can fail with WinError 206 (“filename or extension is too long”) while unpacking its bundled third-party licence tree. This is a MAX_PATH limitation rather than a packaging fault. Either enable long paths in Windows, or place the virtualenv nearer the drive root:

poetry config virtualenvs.in-project true