xitorch

Differentiable scientific computing library


Keywords
project, library, linear-algebra, autograd, functionals
License
MIT
Install
pip install xitorch==0.2.0

Documentation

xitorch: differentiable scientific computing library

NOTE: the development is moved to https://github.com/xitorch/xitorch/

Build Docs Code coverage

xitorch is a PyTorch-based library of differentiable functions and functionals that can be widely used in scientific computing applications as well as deep learning.

The documentation can be found at: https://xitorch.readthedocs.io/

Example

Finding root of a function:

import torch
from xitorch.optimize import rootfinder

def func1(y, A):  # example function
    return torch.tanh(A @ y + 0.1) + y / 2.0

# set up the parameters and the initial guess
A = torch.tensor([[1.1, 0.4], [0.3, 0.8]]).requires_grad_()
y0 = torch.zeros((2,1))  # zeros as the initial guess

# finding a root
yroot = rootfinder(func1, y0, params=(A,))

# calculate the derivatives
dydA, = torch.autograd.grad(yroot.sum(), (A,), create_graph=True)
grad2A, = torch.autograd.grad(dydA.sum(), (A,), create_graph=True)

Modules

  • linalg: Linear algebra and sparse linear algebra module
  • optimize: Optimization and root finder module
  • integrate: Quadrature and integration module
  • interpolate: Interpolation

Requirements

  • python 3.6 or higher
  • pytorch 1.6 or higher (install here)

Getting started

After fulfilling all the requirements, type the commands below to install xitorch

python -m pip install xitorch

Or if you want to install from source:

git clone https://github.com/mfkasim1/xitorch/
cd xitorch
python -m pip install -e .

Gallery

Neural mirror design (example 01):

neural mirror design

Initial velocity optimization in molecular dynamics (example 02):

molecular dynamics