This repository contains an implementation of the Network Expressivity by Activation Rank (NEAR) score. NEAR is a zero-cost proxy for predicting the best performing neural network architecture in neural architecture search. It is based on the effective rank of the pre- and post-activation matrix of a neural network layer. NEAR can also be applied to identify suitable activation functions and weight initialization schemes. For a detailed description, we refer to our paper.
The module can be installed as follows:
git clone <near-score-repository>
cd <near-score-repository>
python3 -m pip install .A simple example on how to use the package is given in example.py. Please note that the example requires
the installation of torchvision.
The module near_score is distributed under the BSD 3-Clause "New" or "Revised" License. For more license and copyright information, see the file LICENSE.
When publishing results obtained with this package, please cite:
@Article{Husistein2024,
title = {{NEAR: A Training-Free Pre-Estimator of Machine Learning Model Performance}},
author = {Raphael T. Husistein and Markus Reiher and Marco Eckhoff},
journal = {arXiv:2408.08776 [cs.LG]}
year = {2024},
}
In case you encounter any problems or bugs, please write a message to [email protected].