https://github.com/Zymrael/torchSODE

PyTorch block-diagonal ODE CUDA solver, designed for gradient-based optimization
https://github.com/Zymrael/torchSODE

Last synced: about 2 months ago
JSON representation

PyTorch block-diagonal ODE CUDA solver, designed for gradient-based optimization

• Host: GitHub
• URL: https://github.com/Zymrael/torchSODE
• Owner: Zymrael
• Created: 2019-04-10T23:43:31.000Z (over 5 years ago)
• Default Branch: master
• Last Pushed: 2020-04-27T11:43:42.000Z (over 4 years ago)
• Last Synced: 2024-11-08T02:35:53.090Z (about 2 months ago)
• Language: Jupyter Notebook
• Homepage:
• Size: 1.42 MB
• Stars: 16
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

• awesome-neural-ode - repo

README

        

# torchSODE

CUDA solver callable from PyTorch. Optimized for independent ordinary differential equations (ODEs) that can be represented as a sparse block diagonal matrix. 

The solver itself is designed to be used during neural network training and thus accepts an additional argument `grad`

## Installation

In your terminal of choice:

`git clone https://github.com/Zymrael/torchODE.git`

`cd torchSODE/src`

`python setup.py install`

In your python files or notebooks:

`import torchSODE`

## API

`torchSODE.solve(F, x0, grad, dt, steps, method='Euler')` performs `steps` integration cycles with `dt` step size. 

For problems where the size of x0 is too large allocating a matrix of dimensions size * size is not always possible. In these cases we assume a compressed representation of `F` which exposes only its diagonal values.

The following convention is used (regardless of problem size):

1. For diagonal `F` allocate a torch.Tensor of shape (1).

2. For 4 block-diagonal `F`, allocate a torch.Tensor of shape (2,2) with values of upper-left diagonal in position [0,0], upper-right diagonal [0,1], lower-left diagonal [1,0], lower-right diagonal [1,1].

In any other scenario `torchSODE.solve` requires `x0.size(0)` to match `F.size(0)` and `F.size(1)`, with `F` assumed to be 4 block-diagonal.

### Methods

`'Euler'` = Euler

`'RK4'` = Runge-Kutta 4