Advect is a focused automatic differentiation library for scientific Python, with broad NumPy API coverage and first-class support for the Python Array API standard. For repeated workloads, it can stage a function into a reusable, optimized program that can itself be differentiated, saved, and loaded.

Try it

python -m pip install advect
# or
conda install -c conda-forge advect

import numpy as np

import advect as ad


def energy(x):
    return np.sum(np.sin(x) ** 2)


x = np.array([0.0, 0.5, 1.0])
value, gradient = ad.value_and_grad(energy)(x)

print(f"loss: {value:.6f}")
print("gradient:", np.round(gradient, 6))
# loss: 0.937922
# gradient: [0.       0.841471 0.909297]

value_and_grad evaluates energy and differentiates it with respect to x. When the same kinds of inputs will run repeatedly, stage the function once:

program = ad.stage(energy, x)
restored = ad.StagedProgram.from_dict(program.to_dict())
print(f"restored loss: {restored(x):.6f}")
# restored loss: 0.937922

From the same code, Advect can compute JVPs, VJPs, Jacobians, Hessians, and higher-order or implicit derivatives. The core API also covers checkpointing and custom primitives. Optional integrations add SciPy and xarray workflows and bridges to other autodiff frameworks.

Related projects

Other NumPy-focused autodiff projects include HIPS Autograd and MyGrad. Advect also offers optional bridges that let a NumPy-backed function participate in JAX or PyTorch.

Learn and contribute

Start with the runnable tutorials, then use the API reference and compatibility tables for exact contracts. Contributors should begin with CONTRIBUTING.md and the developer guide.

Report security issues through GitHub's private advisory form. Advect is distributed under the MIT License.