Welcome to augpy

augpy is a lightweight library with minimal dependencies that provides a comprehensive tensor implementation for Cuda-enabled GPUs, with most of the functionality you are used to from numpy and Pytorch with a similar syntax.

What sets augpy apart is its focus on saturating math (no under or overflows possible), as well as comprehensive support for all data types in all functions. For example, augpy allows you to fill a uint8 tensor with Gaussian distributed random numbers.

augpy’s tensors are based on the DLPack specification and can be exchanged copy-free with other frameworks such as Jax or Pytorch.

While augpy’s support for tensors is quite generic, it also includes additional functionality to work on 2D images, such as high quality affine warps with supersampling and Gaussian blurs.

Indices and tables

Examples