Overview | Example | Installation | Documentation | Support
Jaxdf is a package based on JAX that provides a coding framework for creating differentiable numerical simulators with arbitrary discretizations.
The primary objective of Jaxdf is to aid in the construction of numerical models for physical systems, like wave propagation, or the numerical resolution of partial differential equations, in a manner that is easily tailored to the user's research requirements. These models are pure functions that can be seamlessly integrated into arbitrary differentiable programs written in JAX. For instance, they can be employed as layers within neural networks, or utilized in constructing a physics loss function.
The script below constructs the non-linear operator (∇2 + sin), applying a Fourier spectral discretization on a square 2D domain. It then utilizes this operator to define a loss function. The gradient of this loss function is calculated using JAX's Automatic Differentiation.
from jaxdf import operators as jops
from jaxdf import FourierSeries, operator
from jaxdf.geometry import Domain
from jax import numpy as jnp
from jax import jit, grad
# Defining operator
@operator
def custom_op(u, *, params=None):
grad_u = jops.gradient(u)
diag_jacobian = jops.diag_jacobian(grad_u)
laplacian = jops.sum_over_dims(diag_jacobian)
sin_u = jops.compose(u)(jnp.sin)
return laplacian + sin_u
# Defining discretizations
domain = Domain((128, 128), (1., 1.))
parameters = jnp.ones((128,128,1))
u = FourierSeries(parameters, domain)
# Define a differentiable loss function
@jit
def loss(u):
v = custom_op(u)
return jnp.mean(jnp.abs(v.on_grid)**2)
gradient = grad(loss)(u) # gradient is a FourierSeries
Before proceeding with the installation of jaxdf
, ensure that JAX is already installed on your system. If you intend to utilize jaxdf
with NVidia GPU support, follow the instructions to install JAX accordingly.
To install jaxdf
from PyPI, use the pip
command:
pip install jaxdf
For development purposes, install jaxdf
by either cloning the repository or downloading and extracting the compressed archive. Afterward, navigate to the root folder in a terminal, and execute the following command:
pip install --upgrade poetry
poetry install
This will install the dependencies and the package itself (in editable mode).
If you encounter any issues with the code or wish to suggest new features, please feel free to open an issue. If you seek guidance, wish to discuss something, or simply want to say hi, don't hesitate to write a message in our Discord channel.
Contributions are absolutely welcome! Most contributions start with an issue. Please don't hesitate to create issues in which you ask for features, give feedback on performances, or simply want to reach out.
To make a pull request, please look at the detailed Contributing guide for how to do it, but fundamentally keep in mind the following main guidelines:
kacl-cli
mkdocs serve
An initial version of this package was presented at the Differentiable Programming workshop at NeurIPS 2021.
@article{stanziola2021jaxdf,
author={Stanziola, Antonio and Arridge, Simon and Cox, Ben T. and Treeby, Bradley E.},
title={A research framework for writing differentiable PDE discretizations in JAX},
year={2021},
journal={Differentiable Programming workshop at Neural Information Processing Systems 2021}
}
plum
, check out this amazing project: https://github.com/wesselb/plumodl
Operator Discretization Library (ODL) is a python library for fast prototyping focusing on (but not restricted to) inverse problems.deepXDE
: a TensorFlow and PyTorch library for scientific machine learning.SciML
: SciML is a NumFOCUS sponsored open source software organization created to unify the packages for scientific machine learning.