SaiT: Sparse Vision Transformers through Adaptive Token Pruning

While vision transformers have achieved impressive results, effectively and efficiently accelerating these models can further boost performances. In this work, we propose a dense/sparse training framework to obtain a unified model, enabling weight sharing across various token densities. Thus one mod...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Li, Ling, Thorsley, David, Hassoun, Joseph
Format: Artikel
Sprache:eng
Schlagworte:
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 Li, Ling
Thorsley, David
Hassoun, Joseph
description While vision transformers have achieved impressive results, effectively and efficiently accelerating these models can further boost performances. In this work, we propose a dense/sparse training framework to obtain a unified model, enabling weight sharing across various token densities. Thus one model offers a range of accuracy and throughput tradeoffs for different applications. Besides, we introduce adaptive token pruning to optimize the patch token sparsity based on the input image. In addition, we investigate knowledge distillation to enhance token selection capability in early transformer modules. Sparse adaptive image Transformer (SaiT) offers varying levels of model acceleration by merely changing the token sparsity on the fly. Specifically, SaiT reduces the computation complexity (FLOPs) by 39% - 43% and increases the throughput by 67% - 91% with less than 0.5% accuracy loss for various vision transformer models. Meanwhile, the same model also provides the zero accuracy drop option by skipping the sparsification step. SaiT achieves better accuracy and computation tradeoffs than state-of-the-art transformer and convolutional models.
doi_str_mv 10.48550/arxiv.2210.05832
format Article
fullrecord <record><control><sourceid>arxiv_GOX</sourceid><recordid>TN_cdi_arxiv_primary_2210_05832</recordid><sourceformat>XML</sourceformat><sourcesystem>PC</sourcesystem><sourcerecordid>2210_05832</sourcerecordid><originalsourceid>FETCH-LOGICAL-a672-1283a86c9fae8cd61cf1c3302c8c4c6db6bef42de0afad05942030f623b4112d3</originalsourceid><addsrcrecordid>eNotz81KAzEUhuFsXEj1AlyZG5ianGRi6kZK8Q8KCg1uhzP5aYM2M5y0Re9ebV198C4-eBi7kmKqbduKG6SvfJgC_AbRWgXn7H6F2d3x1YhUI3_PNQ-FO8JS00DbSJXvNjTs1xs-Dzju8iFyN3zEwt9oX3JZX7CzhJ81Xv7vhLnHB7d4bpavTy-L-bJBcwuNBKvQGj9LGK0PRvokvVICvPXam9CbPiYNIQpMGEQ70yCUSAZUr6WEoCbs-nR7FHQj5S3Sd_cn6Y4S9QPPVkOv</addsrcrecordid><sourcetype>Open Access Repository</sourcetype><iscdi>true</iscdi><recordtype>article</recordtype></control><display><type>article</type><title>SaiT: Sparse Vision Transformers through Adaptive Token Pruning</title><source>arXiv.org</source><creator>Li, Ling ; Thorsley, David ; Hassoun, Joseph</creator><creatorcontrib>Li, Ling ; Thorsley, David ; Hassoun, Joseph</creatorcontrib><description>While vision transformers have achieved impressive results, effectively and efficiently accelerating these models can further boost performances. In this work, we propose a dense/sparse training framework to obtain a unified model, enabling weight sharing across various token densities. Thus one model offers a range of accuracy and throughput tradeoffs for different applications. Besides, we introduce adaptive token pruning to optimize the patch token sparsity based on the input image. In addition, we investigate knowledge distillation to enhance token selection capability in early transformer modules. Sparse adaptive image Transformer (SaiT) offers varying levels of model acceleration by merely changing the token sparsity on the fly. Specifically, SaiT reduces the computation complexity (FLOPs) by 39% - 43% and increases the throughput by 67% - 91% with less than 0.5% accuracy loss for various vision transformer models. Meanwhile, the same model also provides the zero accuracy drop option by skipping the sparsification step. SaiT achieves better accuracy and computation tradeoffs than state-of-the-art transformer and convolutional models.</description><identifier>DOI: 10.48550/arxiv.2210.05832</identifier><language>eng</language><subject>Computer Science - Computer Vision and Pattern Recognition</subject><creationdate>2022-10</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>$$Tsyndetics_thumb_exl</thumbnail><link.rule.ids>228,230,780,885</link.rule.ids><linktorsrc>$$Uhttps://arxiv.org/abs/2210.05832$$EView_record_in_Cornell_University$$FView_record_in_$$GCornell_University$$Hfree_for_read</linktorsrc><backlink>$$Uhttps://doi.org/10.48550/arXiv.2210.05832$$DView paper in arXiv$$Hfree_for_read</backlink></links><search><creatorcontrib>Li, Ling</creatorcontrib><creatorcontrib>Thorsley, David</creatorcontrib><creatorcontrib>Hassoun, Joseph</creatorcontrib><title>SaiT: Sparse Vision Transformers through Adaptive Token Pruning</title><description>While vision transformers have achieved impressive results, effectively and efficiently accelerating these models can further boost performances. In this work, we propose a dense/sparse training framework to obtain a unified model, enabling weight sharing across various token densities. Thus one model offers a range of accuracy and throughput tradeoffs for different applications. Besides, we introduce adaptive token pruning to optimize the patch token sparsity based on the input image. In addition, we investigate knowledge distillation to enhance token selection capability in early transformer modules. Sparse adaptive image Transformer (SaiT) offers varying levels of model acceleration by merely changing the token sparsity on the fly. Specifically, SaiT reduces the computation complexity (FLOPs) by 39% - 43% and increases the throughput by 67% - 91% with less than 0.5% accuracy loss for various vision transformer models. Meanwhile, the same model also provides the zero accuracy drop option by skipping the sparsification step. SaiT achieves better accuracy and computation tradeoffs than state-of-the-art transformer and convolutional models.</description><subject>Computer Science - Computer Vision and Pattern Recognition</subject><fulltext>true</fulltext><rsrctype>article</rsrctype><creationdate>2022</creationdate><recordtype>article</recordtype><sourceid>GOX</sourceid><recordid>eNotz81KAzEUhuFsXEj1AlyZG5ianGRi6kZK8Q8KCg1uhzP5aYM2M5y0Re9ebV198C4-eBi7kmKqbduKG6SvfJgC_AbRWgXn7H6F2d3x1YhUI3_PNQ-FO8JS00DbSJXvNjTs1xs-Dzju8iFyN3zEwt9oX3JZX7CzhJ81Xv7vhLnHB7d4bpavTy-L-bJBcwuNBKvQGj9LGK0PRvokvVICvPXam9CbPiYNIQpMGEQ70yCUSAZUr6WEoCbs-nR7FHQj5S3Sd_cn6Y4S9QPPVkOv</recordid><startdate>20221011</startdate><enddate>20221011</enddate><creator>Li, Ling</creator><creator>Thorsley, David</creator><creator>Hassoun, Joseph</creator><scope>AKY</scope><scope>GOX</scope></search><sort><creationdate>20221011</creationdate><title>SaiT: Sparse Vision Transformers through Adaptive Token Pruning</title><author>Li, Ling ; Thorsley, David ; Hassoun, Joseph</author></sort><facets><frbrtype>5</frbrtype><frbrgroupid>cdi_FETCH-LOGICAL-a672-1283a86c9fae8cd61cf1c3302c8c4c6db6bef42de0afad05942030f623b4112d3</frbrgroupid><rsrctype>articles</rsrctype><prefilter>articles</prefilter><language>eng</language><creationdate>2022</creationdate><topic>Computer Science - Computer Vision and Pattern Recognition</topic><toplevel>online_resources</toplevel><creatorcontrib>Li, Ling</creatorcontrib><creatorcontrib>Thorsley, David</creatorcontrib><creatorcontrib>Hassoun, Joseph</creatorcontrib><collection>arXiv Computer Science</collection><collection>arXiv.org</collection></facets><delivery><delcategory>Remote Search Resource</delcategory><fulltext>fulltext_linktorsrc</fulltext></delivery><addata><au>Li, Ling</au><au>Thorsley, David</au><au>Hassoun, Joseph</au><format>journal</format><genre>article</genre><ristype>JOUR</ristype><atitle>SaiT: Sparse Vision Transformers through Adaptive Token Pruning</atitle><date>2022-10-11</date><risdate>2022</risdate><abstract>While vision transformers have achieved impressive results, effectively and efficiently accelerating these models can further boost performances. In this work, we propose a dense/sparse training framework to obtain a unified model, enabling weight sharing across various token densities. Thus one model offers a range of accuracy and throughput tradeoffs for different applications. Besides, we introduce adaptive token pruning to optimize the patch token sparsity based on the input image. In addition, we investigate knowledge distillation to enhance token selection capability in early transformer modules. Sparse adaptive image Transformer (SaiT) offers varying levels of model acceleration by merely changing the token sparsity on the fly. Specifically, SaiT reduces the computation complexity (FLOPs) by 39% - 43% and increases the throughput by 67% - 91% with less than 0.5% accuracy loss for various vision transformer models. Meanwhile, the same model also provides the zero accuracy drop option by skipping the sparsification step. SaiT achieves better accuracy and computation tradeoffs than state-of-the-art transformer and convolutional models.</abstract><doi>10.48550/arxiv.2210.05832</doi><oa>free_for_read</oa></addata></record>
fulltext fulltext_linktorsrc
identifier DOI: 10.48550/arxiv.2210.05832
ispartof
issn
language eng
recordid cdi_arxiv_primary_2210_05832
source arXiv.org
subjects Computer Science - Computer Vision and Pattern Recognition
title SaiT: Sparse Vision Transformers through Adaptive Token Pruning
url https://sfx.bib-bvb.de/sfx_tum?ctx_ver=Z39.88-2004&ctx_enc=info:ofi/enc:UTF-8&ctx_tim=2025-01-04T17%3A31%3A06IST&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=SaiT:%20Sparse%20Vision%20Transformers%20through%20Adaptive%20Token%20Pruning&rft.au=Li,%20Ling&rft.date=2022-10-11&rft_id=info:doi/10.48550/arxiv.2210.05832&rft_dat=%3Carxiv_GOX%3E2210_05832%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