=========================
Installing NeuralOperator
=========================

The ``neuraloperator`` Python package provides all necessary tools for implementing operator learning.
Once installed, you can import it as ``neuralop``::

    import neuralop

~~~~~~~~~~~~~~~~~~~~~~~~~~

Basic Install
-------------

Pre-requisites
++++++++++++++

You will need to have Python 3 installed, as well as NumPy, Scipy, PyTorch, TensorLy and TensorLy-Torch.
If you are starting with Python or generally want a pain-free experience, we recommend that you 
install the `Anaconda distribution <https://www.anaconda.com/download/>`_. It comes ready to use with all prerequisite packages.

~~~~~~~~~~~~~~~~~~~~~~~~~~

Installing with pip
+++++++++++++++++++

We periodically package `neuraloperator` for release on PyPI. This version is not guaranteed to be up-to-date with
the latest changes to our code. 

To install via pip, simply run, in your terminal::

   pip install -U neuraloperator

(the `-U` is optional, use it if you want to update the package).

~~~~~~~~~~~~~~~~~~~~~~~~~~

Building ``neuraloperator`` from source
+++++++++++++++++++++++++++++++++++++++

First ensure that you are in an environment with Python 3, NumPy, SciPy, PyTorch, TensorLy and TensorLy-Torch. 
Then clone the repository and cd there::

   git clone https://github.com/neuraloperator/neuraloperator
   cd neuraloperator


Then, install the requirements ::

   pip install -r requirements.txt


Then install the package (here in editable mode with `-e`, or equivalently `--editable`)::

   pip install -e .


.. raw:: html

   <br/>

~~~~~~~~~~~~~~~~~~~~

Advanced dependencies
---------------------

.. _open3d_dependency:

Fast 3D spatial computing with Open3D
+++++++++++++++++++++++++++++++++++++

To accelerate spatial computing for 3D applications, we include 
`Open3D <https://github.com/isl-org/Open3D>`_ as an optional dependency. Open3D includes
utilities for reading 3D mesh files and fast 3D neighbor search. To install::

   pip install open3d

Note that Open3D is only
compatible with specific builds of PyTorch and CUDA. Check the sub-package 
`Open3D-ML <https://github.com/isl-org/Open3D-ML>`_ documentation for more details. 

~~~~~~~~~~~~~~~~~~~~~

.. _torch_scatter_dependency:

Sparse computations with PyTorch-Scatter
++++++++++++++++++++++++++++++++++++++++

We use the package ``torch_scatter`` (available on PyPI) to speed up neighborhood reductions during 
the final stage of the ``IntegralTransform`` layer, linked here: :ref:`gno_api`.

``torch_scatter`` uses compiled CUDA kernels 
to perform this reduction, providing both memory efficiency and speed advantages over a native PyTorch implementation.

However, since these kernels are precompiled, it is crucial to install the version of ``torch_scatter`` that is built for
your specific combination of OS, PyTorch and CUDA. All available wheels can be found `here <https://data.pyg.org/whl/>`_.  

.. note :: 
    The latest PyTorch version for which ``torch_scatter`` is built is 2.5.1, so you should downgrade if you are using 
    the latest (``torch>=2.6.0``) release:


To force-downgrade PyTorch, run 

.. code ::

   pip install -U torch==2.5.1