Alpine: A Flexible, User-Friendly, and Distributed PyTorch Library for Implicit Neural Representation Development

Kushal Vyas, Vishwanath Saragadam, Ashok Veeraraghavan, Guha Balakrishnan

Abstract:

Implicit neural representations (INRs) are the workhorse
data structure in neural field algorithms, offering a flexible, continuous, and compact encoding of complex signals. While simple in concept, INR designs now vary along
various axes, such as nonlinearities, parameter initialization schemes, training procedures, and interpretability techniques. As such, there is a growing need for a systematic
library to ensure rapid, scalable, and reproducible INR development. To fill this need, we present Alpine, an opensource PyTorch library for flexible development, fitting, and
function visualization for INRs, with a focus on rapid prototyping and ease of extensibility across a variety of scientific applications, from applied physics to medical imaging.
Alpine provides a clean API to set up custom INR workflows, train them using gradient-based or sophisticated metalearners, and visualize learned features, learned INR geometry, and metrics. This paper presents the components of
Alpine, and its capabilities