AdaScale SGD: A User-Friendly Algorithm for Distributed Training
This addresses the challenge of resource-intensive learning rate tuning for distributed training, offering a user-friendly solution for practitioners, though it is incremental as it builds on existing scaling rules and warm-up heuristics.
The paper tackles the problem of adapting learning rates for large-batch training in stochastic gradient descent to maximize speed-ups without degrading model quality, proposing AdaScale SGD which automatically adjusts learning rates based on gradient variance and achieves training beyond batch size limits with no model degradation across tasks like machine translation and image classification.
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.