The Limit of the Batch Size

Large-batch training is an efficient approach for current distributed deep learning systems. It has enabled researchers to reduce the ImageNet/ResNet-50 training from 29 hours to around 1 minute. In this paper, we focus on studying the limit of the batch size. We think it may provide a guidance to AI supercomputer and algorithm designers. We provide detailed numerical optimization instructions for step-by-step comparison. Moreover, it is important to understand the generalization and optimization performance of huge batch training. Hoffer et al. introduced "ultra-slow diffusion" theory to large-batch training. However, our experiments show contradictory results with the conclusion of Hoffer et al. We provide comprehensive experimental results and detailed analysis to study the limitations of batch size scaling and "ultra-slow diffusion" theory. For the first time we scale the batch size on ImageNet to at least a magnitude larger than all previous work, and provide detailed studies on the performance of many state-of-the-art optimization schemes under this setting. We propose an optimization recipe that is able to improve the top-1 test accuracy by 18% compared to the baseline.

[1]  Olatunji Ruwase,et al.  ZeRO: Memory Optimization Towards Training A Trillion Parameter Models , 2019, SC.

[2]  Yuanzhou Yang,et al.  Highly Scalable Deep Learning Training System with Mixed-Precision: Training ImageNet in Four Minutes , 2018, ArXiv.

[3]  Kurt Keutzer,et al.  Inefficiency of K-FAC for Large Batch Size Training , 2019, AAAI.

[4]  Yoram Singer,et al.  Adaptive Subgradient Methods for Online Learning and Stochastic Optimization , 2011, J. Mach. Learn. Res..

[5]  Frank Hutter,et al.  SGDR: Stochastic Gradient Descent with Warm Restarts , 2016, ICLR.

[6]  Leslie N. Smith,et al.  Cyclical Learning Rates for Training Neural Networks , 2015, 2017 IEEE Winter Conference on Applications of Computer Vision (WACV).

[7]  Jorge Nocedal,et al.  On Large-Batch Training for Deep Learning: Generalization Gap and Sharp Minima , 2016, ICLR.

[8]  Alex Krizhevsky,et al.  One weird trick for parallelizing convolutional neural networks , 2014, ArXiv.

[9]  Mohammad Shoeybi,et al.  Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism , 2019, ArXiv.

[10]  Hiroaki Mikami,et al.  Massively Distributed SGD: ImageNet/ResNet-50 Training in a Flash , 2018 .

[11]  Joseph Gonzalez,et al.  On the Computational Inefficiency of Large Batch Sizes for Stochastic Gradient Descent , 2018, ArXiv.

[12]  A. Bray,et al.  Statistics of critical points of Gaussian fields on large-dimensional spaces. , 2006, Physical review letters.

[13]  James Demmel,et al.  ImageNet Training in Minutes , 2017, ICPP.

[14]  Jascha Sohl-Dickstein,et al.  Measuring the Effects of Data Parallelism on Neural Network Training , 2018, J. Mach. Learn. Res..

[15]  Dennis DeCoste,et al.  Stochastic Weight Averaging in Parallel: Large-Batch Training that Generalizes Well , 2020, ICLR.

[16]  Takuya Akiba,et al.  Extremely Large Minibatch SGD: Training ResNet-50 on ImageNet in 15 Minutes , 2017, ArXiv.

[17]  Surya Ganguli,et al.  Identifying and attacking the saddle point problem in high-dimensional non-convex optimization , 2014, NIPS.

[18]  J. Bouchaud,et al.  Anomalous diffusion in disordered media: Statistical mechanisms, models and physical applications , 1990 .

[19]  Satoshi Matsuoka,et al.  Large-Scale Distributed Second-Order Optimization Using Kronecker-Factored Approximate Curvature for Deep Convolutional Neural Networks , 2018, 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR).

[20]  Elad Hoffer,et al.  Exponentially vanishing sub-optimal local minima in multilayer neural networks , 2017, ICLR.

[21]  Pongsakorn U.-Chupala,et al.  ImageNet/ResNet-50 Training in 224 Seconds , 2018, ArXiv.

[22]  Mu Li.  Proposal Scaling Distributed Machine Learning with System and Algorithm Co-design , 2016 .

[23]  James Demmel,et al.  Large Batch Optimization for Deep Learning: Training BERT in 76 minutes , 2019, ICLR.

[24]  Geoffrey E. Hinton,et al.  Learning representations by back-propagating errors , 1986, Nature.

[25]  Boris Ginsburg,et al.  Stochastic Gradient Methods with Layer-wise Adaptive Moments for Training of Deep Networks , 2019, ArXiv.

[26]  Jimmy Ba,et al.  Adam: A Method for Stochastic Optimization , 2014, ICLR.

[27]  Xu Sun,et al.  Adaptive Gradient Methods with Dynamic Bound of Learning Rate , 2019, ICLR.

[28]  Liyuan Liu,et al.  On the Variance of the Adaptive Learning Rate and Beyond , 2019, ICLR.

[29]  Tao Lin,et al.  Don't Use Large Mini-Batches, Use Local SGD , 2018, ICLR.

[30]  Quoc V. Le,et al.  Don't Decay the Learning Rate, Increase the Batch Size , 2017, ICLR.

[31]  Tao Wang,et al.  Scale MLPerf-0.6 models on Google TPU-v3 Pods , 2019, ArXiv.

[32]  Yang You,et al.  Scaling SGD Batch Size to 32K for ImageNet Training , 2017, ArXiv.

[33]  Masafumi Yamazaki,et al.  Yet Another Accelerated SGD: ResNet-50 Training on ImageNet in 74.7 seconds , 2019, ArXiv.

[34]  Roger B. Grosse,et al.  Optimizing Neural Networks with Kronecker-factored Approximate Curvature , 2015, ICML.

[35]  Kaiming He,et al.  Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour , 2017, ArXiv.

[36]  Jian Sun,et al.  Deep Residual Learning for Image Recognition , 2015, 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR).

[37]  Yoram Singer,et al.  Memory Efficient Adaptive Optimization , 2019, NeurIPS.

[38]  Tao Wang,et al.  Image Classification at Supercomputer Scale , 2018, ArXiv.

[39]  Razvan Pascanu,et al.  On the difficulty of training recurrent neural networks , 2012, ICML.

[40]  Geoffrey E. Hinton,et al.  Regularizing Neural Networks by Penalizing Confident Output Distributions , 2017, ICLR.

[41]  Elad Hoffer,et al.  Train longer, generalize better: closing the generalization gap in large batch training of neural networks , 2017, NIPS.

[42]  James Demmel,et al.  A view of the parallel computing landscape , 2009, CACM.