This repository provides the code and data used in the On Network Design Spaces for Visual Recognition work, including full training statistics for over 100,000 models spanning multiple model families.
Comparing networks. (a) Early work on neural networks for visual recognition tasks used point estimates to compare architectures, often irrespective of model complexity. (b) More recent work compares curve estimates of error vs. complexity traced by a handful of selected models. (c) We propose to sample models from a parameterized model design space, and measure distribution estimates to compare design spaces. This methodology allows for a more complete and unbiased view of the design landscape.
Data is available for download here. We provide notebooks to reproduce all figures from the paper, that serve as examples of how to use the data and apply our methodology. All models were trained using pycls.
If you use the code or data in your research, please use the following BibTex entry:
@InProceedings{Radosavovic2019,
title = {On Network Design Spaces for Visual Recognition},
author = {Radosavovic, Ilija and Johnson, Justin and Xie, Saining and Lo, Wan-Yen and Doll{\'a}r, Piotr},
booktitle = {ICCV},
year = {2019}
}
The code is released under the MIT license. Please see the LICENSE file for more information.