{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/large-batch-training-for-lstm-and-beyond","title":"Large-Batch Training for LSTM and Beyond","arxiv_id":"1901.08256","date":"2019-01-24","proceeding":null,"authors":["Yang You","Jonathan Hseu","Chris Ying","James Demmel","Kurt Keutzer","Cho-Jui Hsieh"],"abstract":"Large-batch training approaches have enabled researchers to utilize\nlarge-scale distributed processing and greatly accelerate deep-neural net (DNN)\ntraining. For example, by scaling the batch size from 256 to 32K, researchers\nhave been able to reduce the training time of ResNet50 on ImageNet from 29\nhours to 2.2 minutes (Ying et al., 2018). In this paper, we propose a new\napproach called linear-epoch gradual-warmup (LEGW) for better large-batch\ntraining. With LEGW, we are able to conduct large-batch training for both CNNs\nand RNNs with the Sqrt Scaling scheme. LEGW enables Sqrt Scaling scheme to be\nuseful in practice and as a result we achieve much better results than the\nLinear Scaling learning rate scheme. For LSTM applications, we are able to\nscale the batch size by a factor of 64 without losing accuracy and without\ntuning the hyper-parameters. For CNN applications, LEGW is able to achieve the\nsame accuracy even as we scale the batch size to 32K. LEGW works better than\nprevious large-batch auto-tuning techniques. LEGW achieves a 5.3X average\nspeedup over the baselines for four LSTM-based applications on the same\nhardware. We also provide some theoretical explanations for LEGW.","url_abs":"http://arxiv.org/abs/1901.08256v1","url_pdf":"http://arxiv.org/pdf/1901.08256v1.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"large-batch-training-for-lstm-and-beyond","repo_url":"https://github.com/NUS-HPC-AI-Lab/LARS-ImageNet-PyTorch","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":null}],"tasks":[],"methods":[{"method_slug":"lstm","method_name":"LSTM"},{"method_slug":"sigmoid-activation","method_name":"Sigmoid Activation"},{"method_slug":"tanh-activation","method_name":"Tanh Activation"}],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1901.08256","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}