Fast Benchmarking of Accuracy vs. Training Time with Cyclic Learning Rates

Abstract

Benchmarking the tradeoff between neural network accuracy and training time is computationally expensive. Here we show how a multiplicative cyclic learning rate schedule can be used to construct a tradeoff curve in a single training run. We generate cyclic tradeoff curves for combinations of training methods such as Blurpool, Channels Last, Label Smoothing and MixUp, and highlight how these cyclic tradeoff curves can be used to efficiently evaluate the effects of algorithmic choices on network training.

PDF NeurIPSW OpenReview Semantic Scholar

Cite

Text

Portes et al. "Fast Benchmarking of Accuracy vs. Training Time with Cyclic Learning Rates." NeurIPS 2022 Workshops: HITY, 2022.

Markdown

[Portes et al. "Fast Benchmarking of Accuracy vs. Training Time with Cyclic Learning Rates." NeurIPS 2022 Workshops: HITY, 2022.](https://mlanthology.org/neuripsw/2022/portes2022neuripsw-fast/)

BibTeX

@inproceedings{portes2022neuripsw-fast,
  title     = {{Fast Benchmarking of Accuracy vs. Training Time with Cyclic Learning Rates}},
  author    = {Portes, Jacob and Blalock, Davis and Stephenson, Cory and Frankle, Jonathan},
  booktitle = {NeurIPS 2022 Workshops: HITY},
  year      = {2022},
  url       = {https://mlanthology.org/neuripsw/2022/portes2022neuripsw-fast/}
}