Abstract
Training large deep neural networks on massive datasets is computationally very challenging. There has been recent surge in interest in using large batch stochastic optimization methods to tackle this issue. The most prominent algorithm in this line of research is L ARS, which by employing layerwise adaptive learning rates trains R ES N ET on ImageNet in a few minutes. However, L ARS performs poorly for attention models like B ERT, indicating that its performance gains are not consistent across tasks. In this paper, we first study a principled layerwise adaptation strategy to accelerate training of deep neural networks using large mini-batches. Using this strategy, we develop a new layerwise adaptive large batch optimization technique called L AMB; we then provide convergence analysis of L AMB as well as L ARS, showing convergence to a stationary point in general nonconvex settings. Our empirical results demonstrate the superior performance of L AMB across various tasks such as B ERT and R ES N ET-50 training with very little hyperparameter tuning. In particular, for B ERT training, our optimizer enables use of very large batch sizes of 32868 without any degradation of performance. By increasing the batch size to the memory limit of a TPUv3 Pod, B ERT training time can be reduced from 3 days to just 76 minutes (Table 1).