torch_interpolations

This package implements interpolation routines in PyTorch, making them GPU-capable and differentiable. The only interpolation routine supported so far is RegularGridInterpolator, from scipy.

Installation

Install the preview of PyTorch here: https://pytorch.org/. Ensure that you have torch version >= 1.7.0. Then navigate to the main directory of this repository and run:

$ python setup.py install

API

First, you construct a torch_interpolations.RegularGridInterpolator object by supplying it points (a list of torch Tensors) and values (a torch Tensor):

rgi = torch_interpolations.RegularGridInterpolator(points, values)

Then, to interpolate a set of points, you run:

rgi(points_to_interp)

where points_to_interp is a list of torch Tensors, each with the same shape.

Tests

First, install pytest:

pip install pytest

Then, from the main directory, run:

pytest .

Examples

To run the examples, first install the dependencies:

pip install scipy matplotlib

Then navigate to the examples folder. The examples are the basic example:

python basic_example.py

and the two-dimensional example:

python two_dimensional.py

Performance

We include a performance script in the perf/ folder. To run it, navigate to the perf/ folder, and run:

python perf.py

It compares the torch CPU and GPU implementations, and the scipy CPU implementation. The test case involves interpolating 4.5 million points on a 300 x 300 grid. The timing results on my machine (Intel i9-9980XE and Nvidia GeForce 1080 TI) are:

Interpolating 4500000 points on 300 by 300 grid
PyTorch took 146.812 +\- 7.167 ms
PyTorch Cuda took 14.799 +\- 6.816 ms
Scipy took 545.935 +\- 1.089 ms

So the torch GPU implementation is 20 times faster than the torch CPU implementation, which itself is twice as fast as the scipy implementation.

License

torch_interpolations carries an Apache 2.0 license.