AdaScale SGD: A User-Friendly Algorithm for Distributed Training

TL;DR

AdaScale SGD is a new algorithm that automatically adjusts learning rates during large-batch training, ensuring model quality and speed without extensive re-tuning or additional hyperparameters.

Contribution

It introduces AdaScale SGD, a principled method for adaptive learning rate scaling that maintains model quality across various batch sizes without extra hyperparameters.

Findings

Achieves large-batch training without model degradation.

Outperforms traditional linear scaling rules.

Maintains convergence bounds similar to standard SGD.

Abstract

When using large-batch training to speed up stochastic gradient descent, learning rates must adapt to new batch sizes in order to maximize speed-ups and preserve model quality. Re-tuning learning rates is resource intensive, while fixed scaling rules often degrade model quality. We propose AdaScale SGD, an algorithm that reliably adapts learning rates to large-batch training. By continually adapting to the gradient's variance, AdaScale automatically achieves speed-ups for a wide range of batch sizes. We formally describe this quality with AdaScale's convergence bound, which maintains final objective values, even as batch sizes grow large and the number of iterations decreases. In empirical comparisons, AdaScale trains well beyond the batch size limits of popular "linear learning rate scaling" rules. This includes large-batch training with no model degradation for machine translation, image classification, object detection, and speech recognition tasks. AdaScale's qualitative behavior is similar to that of "warm-up" heuristics, but unlike warm-up, this behavior emerges naturally from a principled mechanism. The algorithm introduces negligible computational overhead and no new hyperparameters, making AdaScale an attractive choice for large-scale training in practice.