neuraloperator is a comprehensive library for learning neural operators in PyTorch. It is the official implementation for Fourier Neural Operators and Tensorized Neural Operators.

Unlike regular neural networks, neural operators enable learning mapping between function spaces, and this library provides all of the tools to do so on your own data.

Neural operators are also resolution invariant, so your trained operator can be applied on data of any resolution.

Just clone the repository and install locally (in editable mode so changes in the code are immediately reflected without having to reinstall):

git clone https://github.com/NeuralOperator/neuraloperator
cd neuraloperator
pip install -e .
pip install -r requirements.txt

You can also just pip install the most recent stable release of the library on PyPI:

pip install neuraloperator

After you've installed the library, you can start training operators seamlessly:

from neuralop.models import FNO

operator = FNO(n_modes=(16, 16), hidden_channels=64,
                in_channels=3, out_channels=1)

Tensorization is also provided out of the box: you can improve the previous models by simply using a Tucker Tensorized FNO with just a few parameters:

from neuralop.models import TFNO

operator = TFNO(n_modes=(16, 16), hidden_channels=64,
                in_channels=3,
                out_channels=1,
                factorization='tucker',
                implementation='factorized',
                rank=0.05)

This will use a Tucker factorization of the weights. The forward pass will be efficient by contracting directly the inputs with the factors of the decomposition. The Fourier layers will have 5% of the parameters of an equivalent, dense Fourier Neural Operator!

Checkout the documentation for more!

Create a file in neuraloperator/config called wandb_api_key.txt and paste your Weights and Biases API key there. You can configure the project you want to use and your username in the main yaml configuration files.

NeuralOperator is 100% open-source, and we welcome all contributions from the community! If you spot a bug or a typo in the documentation, or have an idea for a feature you'd like to see, please report it on our issue tracker, or even better, open a Pull-Request on GitHub.

NeuralOperator has additional dependencies for development, which can be found in requirements_dev.txt:

pip install -r requirements_dev.txt

Before you submit your changes, you should make sure your code adheres to our style-guide. The easiest way to do this is with black:

black .

Testing and documentation are an essential part of this package and all functions come with unit-tests and documentation. The tests are run using the pytest package.

To run the tests, simply run, in the terminal:

pytest -v neuralop

The HTML for our documentation website is built using sphinx. The documentation is built from inside the doc folder.

cd doc
make html

This will build the docs in ./doc/build/html.

Note that the documentation requires other dependencies installable from ./doc/requirements_doc.txt.

To view the documentation locally, run:

cd doc/build/html
python -m http.server [PORT_NUM]

The docs will then be viewable at localhost:PORT_NUM.

If you use NeuralOperator in an academic paper, please cite [1], [2]:

@misc{li2020fourier,
   title={Fourier Neural Operator for Parametric Partial Differential Equations},
   author={Zongyi Li and Nikola Kovachki and Kamyar Azizzadenesheli and Burigede Liu and Kaushik Bhattacharya and Andrew Stuart and Anima Anandkumar},
   year={2020},
   eprint={2010.08895},
   archivePrefix={arXiv},
   primaryClass={cs.LG}
}

@article{kovachki2021neural,
   author    = {Nikola B. Kovachki and
                  Zongyi Li and
                  Burigede Liu and
                  Kamyar Azizzadenesheli and
                  Kaushik Bhattacharya and
                  Andrew M. Stuart and
                  Anima Anandkumar},
   title     = {Neural Operator: Learning Maps Between Function Spaces},
   journal   = {CoRR},
   volume    = {abs/2108.08481},
   year      = {2021},
}