Massively Distributed SGD: ImageNet/ResNet-50 Training in a Flash

Scaling the distributed deep learning to a massive GPU cluster level is challenging due to the instability of the large mini-batch training and the overhead of the gradient synchronization. We address the instability of the large mini-batch training with batch-size control and label smoothing. We ad...

Ausführliche Beschreibung

Gespeichert in:

Bibliographische Detailangaben
Hauptverfasser:	Mikami, Hiroaki, Suganuma, Hisahiro, U-chupala, Pongsakorn, Tanaka, Yoshiki, Kageyama, Yuichi
Format:	Artikel
Sprache:	eng
Schlagworte:	Computer Science - Computer Vision and Pattern Recognition Computer Science - Learning
Online-Zugang:	Volltext bestellen
Tags:	Tag hinzufügen Keine Tags, Fügen Sie den ersten Tag hinzu!

container_end_page
container_issue
container_start_page
container_title
container_volume
creator	Mikami, Hiroaki Suganuma, Hisahiro U-chupala, Pongsakorn Tanaka, Yoshiki Kageyama, Yuichi
description	Scaling the distributed deep learning to a massive GPU cluster level is challenging due to the instability of the large mini-batch training and the overhead of the gradient synchronization. We address the instability of the large mini-batch training with batch-size control and label smoothing. We address the overhead of the gradient synchronization with 2D-Torus all-reduce. Specifically, 2D-Torus all-reduce arranges GPUs in a logical 2D grid and performs a series of collective operation in different orientations. These two techniques are implemented with Neural Network Libraries (NNL). We have successfully trained ImageNet/ResNet-50 in 122 seconds without significant accuracy loss on ABCI cluster.
doi_str_mv	10.48550/arxiv.1811.05233
format	Article
fullrecord	<record><control><sourceid>arxiv_GOX</sourceid><recordid>TN_cdi_arxiv_primary_1811_05233</recordid><sourceformat>XML</sourceformat><sourcesystem>PC</sourcesystem><sourcerecordid>1811_05233</sourcerecordid><originalsourceid>FETCH-LOGICAL-a673-d77b7813c395085d359f3573ab1b35c4e8d63e9946c3b8af3135761784ac2c423</originalsourceid><addsrcrecordid>eNotz0FPwjAYxvFeOBj0A3iyX2Cj3dt37bwRECRBSHT35W3XYZOxkHYS-fYievofnuRJfow9SpErgyhmFL_DOZdGylxgAXDH5m-UUjj7_sKXIY0x2K_Rt_xjvXzmmyMd_M6Ps3efrslQ8DpSGMJw4GHgxFc9pc97NumoT_7hv1NWr17qxWu23a83i_k2o1JD1mpttZHgoEJhsAWsOkANZKUFdMqbtgRfVap0YA11IK9rKbVR5AqnCpiyp7_bm6E5xXCkeGl-Lc3NAj8n2kGT</addsrcrecordid><sourcetype>Open Access Repository</sourcetype><iscdi>true</iscdi><recordtype>article</recordtype></control><display><type>article</type><title>Massively Distributed SGD: ImageNet/ResNet-50 Training in a Flash</title><source>arXiv.org</source><creator>Mikami, Hiroaki ; Suganuma, Hisahiro ; U-chupala, Pongsakorn ; Tanaka, Yoshiki ; Kageyama, Yuichi</creator><creatorcontrib>Mikami, Hiroaki ; Suganuma, Hisahiro ; U-chupala, Pongsakorn ; Tanaka, Yoshiki ; Kageyama, Yuichi</creatorcontrib><description>Scaling the distributed deep learning to a massive GPU cluster level is challenging due to the instability of the large mini-batch training and the overhead of the gradient synchronization. We address the instability of the large mini-batch training with batch-size control and label smoothing. We address the overhead of the gradient synchronization with 2D-Torus all-reduce. Specifically, 2D-Torus all-reduce arranges GPUs in a logical 2D grid and performs a series of collective operation in different orientations. These two techniques are implemented with Neural Network Libraries (NNL). We have successfully trained ImageNet/ResNet-50 in 122 seconds without significant accuracy loss on ABCI cluster.</description><identifier>DOI: 10.48550/arxiv.1811.05233</identifier><language>eng</language><subject>Computer Science - Computer Vision and Pattern Recognition ; Computer Science - Learning</subject><creationdate>2018-11</creationdate><rights>http://arxiv.org/licenses/nonexclusive-distrib/1.0</rights><oa>free_for_read</oa><woscitedreferencessubscribed>false</woscitedreferencessubscribed></display><links><openurl>$$Topenurl_article</openurl><openurlfulltext>$$Topenurlfull_article</openurlfulltext><thumbnail>$$Tsyndetics_thumb_exl</thumbnail><link.rule.ids>228,230,781,886</link.rule.ids><linktorsrc>$$Uhttps://arxiv.org/abs/1811.05233$$EView_record_in_Cornell_University$$FView_record_in_$$GCornell_University$$Hfree_for_read</linktorsrc><backlink>$$Uhttps://doi.org/10.48550/arXiv.1811.05233$$DView paper in arXiv$$Hfree_for_read</backlink></links><search><creatorcontrib>Mikami, Hiroaki</creatorcontrib><creatorcontrib>Suganuma, Hisahiro</creatorcontrib><creatorcontrib>U-chupala, Pongsakorn</creatorcontrib><creatorcontrib>Tanaka, Yoshiki</creatorcontrib><creatorcontrib>Kageyama, Yuichi</creatorcontrib><title>Massively Distributed SGD: ImageNet/ResNet-50 Training in a Flash</title><description>Scaling the distributed deep learning to a massive GPU cluster level is challenging due to the instability of the large mini-batch training and the overhead of the gradient synchronization. We address the instability of the large mini-batch training with batch-size control and label smoothing. We address the overhead of the gradient synchronization with 2D-Torus all-reduce. Specifically, 2D-Torus all-reduce arranges GPUs in a logical 2D grid and performs a series of collective operation in different orientations. These two techniques are implemented with Neural Network Libraries (NNL). We have successfully trained ImageNet/ResNet-50 in 122 seconds without significant accuracy loss on ABCI cluster.</description><subject>Computer Science - Computer Vision and Pattern Recognition</subject><subject>Computer Science - Learning</subject><fulltext>true</fulltext><rsrctype>article</rsrctype><creationdate>2018</creationdate><recordtype>article</recordtype><sourceid>GOX</sourceid><recordid>eNotz0FPwjAYxvFeOBj0A3iyX2Cj3dt37bwRECRBSHT35W3XYZOxkHYS-fYievofnuRJfow9SpErgyhmFL_DOZdGylxgAXDH5m-UUjj7_sKXIY0x2K_Rt_xjvXzmmyMd_M6Ps3efrslQ8DpSGMJw4GHgxFc9pc97NumoT_7hv1NWr17qxWu23a83i_k2o1JD1mpttZHgoEJhsAWsOkANZKUFdMqbtgRfVap0YA11IK9rKbVR5AqnCpiyp7_bm6E5xXCkeGl-Lc3NAj8n2kGT</recordid><startdate>20181113</startdate><enddate>20181113</enddate><creator>Mikami, Hiroaki</creator><creator>Suganuma, Hisahiro</creator><creator>U-chupala, Pongsakorn</creator><creator>Tanaka, Yoshiki</creator><creator>Kageyama, Yuichi</creator><scope>AKY</scope><scope>GOX</scope></search><sort><creationdate>20181113</creationdate><title>Massively Distributed SGD: ImageNet/ResNet-50 Training in a Flash</title><author>Mikami, Hiroaki ; Suganuma, Hisahiro ; U-chupala, Pongsakorn ; Tanaka, Yoshiki ; Kageyama, Yuichi</author></sort><facets><frbrtype>5</frbrtype><frbrgroupid>cdi_FETCH-LOGICAL-a673-d77b7813c395085d359f3573ab1b35c4e8d63e9946c3b8af3135761784ac2c423</frbrgroupid><rsrctype>articles</rsrctype><prefilter>articles</prefilter><language>eng</language><creationdate>2018</creationdate><topic>Computer Science - Computer Vision and Pattern Recognition</topic><topic>Computer Science - Learning</topic><toplevel>online_resources</toplevel><creatorcontrib>Mikami, Hiroaki</creatorcontrib><creatorcontrib>Suganuma, Hisahiro</creatorcontrib><creatorcontrib>U-chupala, Pongsakorn</creatorcontrib><creatorcontrib>Tanaka, Yoshiki</creatorcontrib><creatorcontrib>Kageyama, Yuichi</creatorcontrib><collection>arXiv Computer Science</collection><collection>arXiv.org</collection></facets><delivery><delcategory>Remote Search Resource</delcategory><fulltext>fulltext_linktorsrc</fulltext></delivery><addata><au>Mikami, Hiroaki</au><au>Suganuma, Hisahiro</au><au>U-chupala, Pongsakorn</au><au>Tanaka, Yoshiki</au><au>Kageyama, Yuichi</au><format>journal</format><genre>article</genre><ristype>JOUR</ristype><atitle>Massively Distributed SGD: ImageNet/ResNet-50 Training in a Flash</atitle><date>2018-11-13</date><risdate>2018</risdate><abstract>Scaling the distributed deep learning to a massive GPU cluster level is challenging due to the instability of the large mini-batch training and the overhead of the gradient synchronization. We address the instability of the large mini-batch training with batch-size control and label smoothing. We address the overhead of the gradient synchronization with 2D-Torus all-reduce. Specifically, 2D-Torus all-reduce arranges GPUs in a logical 2D grid and performs a series of collective operation in different orientations. These two techniques are implemented with Neural Network Libraries (NNL). We have successfully trained ImageNet/ResNet-50 in 122 seconds without significant accuracy loss on ABCI cluster.</abstract><doi>10.48550/arxiv.1811.05233</doi><oa>free_for_read</oa></addata></record>
fulltext	fulltext_linktorsrc
identifier	DOI: 10.48550/arxiv.1811.05233
ispartof
issn
language	eng
recordid	cdi_arxiv_primary_1811_05233
source	arXiv.org
subjects	Computer Science - Computer Vision and Pattern Recognition Computer Science - Learning
title	Massively Distributed SGD: ImageNet/ResNet-50 Training in a Flash
url	https://sfx.bib-bvb.de/sfx_tum?ctx_ver=Z39.88-2004&ctx_enc=info:ofi/enc:UTF-8&ctx_tim=2024-12-12T22%3A34%3A52IST&url_ver=Z39.88-2004&url_ctx_fmt=infofi/fmt:kev:mtx:ctx&rfr_id=info:sid/primo.exlibrisgroup.com:primo3-Article-arxiv_GOX&rft_val_fmt=info:ofi/fmt:kev:mtx:journal&rft.genre=article&rft.atitle=Massively%20Distributed%20SGD:%20ImageNet/ResNet-50%20Training%20in%20a%20Flash&rft.au=Mikami,%20Hiroaki&rft.date=2018-11-13&rft_id=info:doi/10.48550/arxiv.1811.05233&rft_dat=%3Carxiv_GOX%3E1811_05233%3C/arxiv_GOX%3E%3Curl%3E%3C/url%3E&disable_directlink=true&sfx.directlink=off&sfx.report_link=0&rft_id=info:oai/&rft_id=info:pmid/&rfr_iscdi=true