pytorch_interpolation

Fast bilinear interpolation in PyTorch on a regular grid on a unstructured set of 2D query points, analogous to Scipy's RegularGridInterpolator.

pytorch_interpolation runs entirely in C++/CUDA backends, thus significantly outperforming Scipy.

This repository implements a C++/CUDA extensions for PyTorch (in the style of https://github.com/pytorch/extension-cpp)

At present, only bilinear 2D interpolation is implemented and padding/linear extrapolation is applied to points outside the domain of the cartesian grid. The method assumes that the function is known is a on a regular grid.

Installation intructions

First, you need to have PyTorch installed. Now you are set up and ready to install the repository

pip install .

To run the examples provided, install numpy, scipy and matplotlib:

pip install numpy scipy matplotlib

The repository was tested with Python3.8 and PyTorch with CUDA version 12.1

Usage example

The package works similarly to Scipy's RegularGrid. In example.py we provided a simple example and comparison between our interpolation and Scipy's RegularGridInterpolator.

The syntax is analogous to that of Scipy's RegularGridInterpolator

In a nutshell, suppose that:

• x and y are the x and y grid of data points you want to interpolate over (given in ascending order)
• F is the function values at these points
• xpt and ypt be the N dimensional tensor of interpolation query points

then you can call

from pytorch_interp import RegularGridInterpolator

interp = RegularGridInterpolator(F,x,y,xpt,ypt)
G = interp(xpt,ypt)

Optionally, the constant fill value padding can modified as fill_value=value. Also, bilinear extrapolation is now possible by setting fill_value=None (just like in Scipy).

Performance

The script performance.py tests the performance test for pytorch_interp in cpu/cuda vs scipy and torch_interpolations, another pytorch package for bilinear interpolation written entirely in Python.

Our implementation outperformes all others libraries. This is because we use precompiled C++/CUDA code does not require type checking. In addition our implementation allow for larger batch sizes than torch_interpolations in CUDA (we could test up to <= 2^29 query points, whereas torch_interpolations supported <= 2^27 query points).

All tests are done on a 11th Gen Intel(R) Core(TM) i7-11700 @ 2.50GHz CPU and a NVIDIA GeForce RTX 3060 GPU. For the CPU performance tests, 8 threads are run in parallel in Torch.