[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-184299-en":3,"doc-seo-184299-105":31,"detail-sidebar-cat-0-en-105":97},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":20,"is_deleted":4,"is_public":20,"is_downloadable":20,"audit_status":20,"page_count":21,"language":22,"language_code":23,"site_id":24,"html_lang":25,"table_of_contents":26,"faqs":27,"seo_title":28,"seo_description":14,"update_tm":29,"read_time":30},184299,1099523882182,"Alex Sinclair","https://ap-avatar.wpscdn.com/davatar_6f874abed73319feea01a86fa6f0fab8",8,"Research & Report","Liu2022A - RegNetY/EffNet/DeiT/Swin/ConvNeXt 预训练与检测结果对比","文档汇总并对比多种视觉模型在不同预训练设置下的规模与效果表现，包括ImageNet-1K与ImageNet-22K预训练、以及下游微调后的检测性能。内容涵盖RegNetY、EfficientNet（EffNet）、DeiT、Swin Transformer、ConvNeXt等模型的参数量与计算量，同时给出Mask-RCNN与Cascade Mask-RCNN在不同训练日程下的指标结果。另包含微调与优化器的训练配置，如学习率、权重衰减、余弦学习率衰减及数据增强策略等。","| ImageNet-1K trained models\u003Cbr>􀀏 RegNetY-4G [51] 2242 21M 4 .0G\u003Cbr>􀀏 RegNetY-8G [51] 2242 39M 8 .0G\u003Cbr>􀀏 RegNetY-16G [51] 2242 84M 16 .0G\u003Cbr>􀀏 EffNet-B3 [67] 3002 12M 1 . 8G\u003Cbr>􀀏 EffNet-B4 [67] 3802 19M 4 .2G\u003Cbr>􀀏 EffNet-B5 [67] 4562 30M 9 .9G\u003Cbr>􀀏 EffNet-B6 [67] 5282 43M 19 .0G\u003Cbr>􀀏 EffNet-B7 [67] 6002 66M 37 .0G |  | 1156.7\u003Cbr>591.6\u003Cbr>334.7\u003Cbr>732.1\u003Cbr>349.4\u003Cbr>169.1\u003Cbr>96.9\u003Cbr>55.1 | 80.0\u003Cbr>81.7\u003Cbr>82.9\u003Cbr>81.6\u003Cbr>82.9\u003Cbr>83.6\u003Cbr>84.0\u003Cbr>84.3 |\n| --- | --- | --- | --- |\n| 􀀎 DeiT-S [68]\u003Cbr>􀀎 DeiT-B [68] | 2242 22M 4 .6G\u003Cbr>2242 87M 17 .6G | 978.5\u003Cbr>302.1 | 79.8\u003Cbr>81.8 |\n| 􀀎 Swin-T | 2242 28M 4 .5G | 757.9 | 81.3 |\n| 􀀏 ConvNeXt-T |  2242 29M 4.5G  | 774.7 | 82.1 |\n| 􀀎 Swin-S | 2242 50M 8 .7G | 436.7 | 83.0 |\n| 􀀏 ConvNeXt-S |  2242 50M 8.7G  | 447.1 | 83.1 |\n| 􀀎 Swin-B | 2242 88M 15 .4G | 286.6 | 83.5 |\n| 􀀏 ConvNeXt-B |  2242 89M 15.4G  | 292.1 | 83.8 |\n| 􀀎 Swin-B | 3842 88M 47 . 1G | 85.1 | 84.5 |\n| 􀀏 ConvNeXt-B |  3842 89M 45.0G  | 95.7 | 85.1 |\n| 􀀏 ConvNeXt-L | \u003Cbr>2242 198M 34.4G | 146.8 | 84.3 |\n| 􀀏 ConvNeXt-L |  3842 198M 101.0G | 50.4 | 85.5 |\n| 􀀏 R-101x3 [36]\u003Cbr>􀀏 R-152x4 [36] | ImageNet-22K pre-trained models 3842 388M 204 .6G 4802 937M 840 .5G | -\u003Cbr>- | 84.4\u003Cbr>85.4 |\n| 􀀎 ViT-B/16 [18]\u003Cbr>􀀎 ViT-L/16 [18] | 3842 87M 55 .5G\u003Cbr>3842 305M 191 . 1G | 93.1\u003Cbr>28.5 | 84.0\u003Cbr>85.2 |\n| 􀀎 Swin-B | 2242 88M 15 .4G | 286.6 | 85.2 |\n| 􀀏 ConvNeXt-B |  2242 89M 15.4G  | 292.1 | 85.8 |\n| 􀀎 Swin-B | 3842 88M 47 .0G | 85.1 | 86.4 |\n| 􀀏 ConvNeXt-B |  3842 89M 45.1G  | 95.7 | 86.8 |\n| 􀀎 Swin-L | 2242 197M 34 .5G | 145.0 | 86.3 |\n| 􀀏 ConvNeXt-L |  2242 198M 34.4G  | 146.8 | 86.6 |\n| 􀀎 Swin-L | 3842 197M 103 .9G | 46.0 | 87.3 |\n| 􀀏 ConvNeXt-L |  3842 198M 101.0G | 50.4 | 87.5 |\n| 􀀏 ConvNeXt-XL\u003Cbr>􀀏 ConvNeXt-XL | \u003Cbr>2242 350M 60.9G 3842 350M 179.0G | 89.3 30.2 | 87.0 87.8 |\n\n| Mask-RCNN 3 􀀂 schedule |  |  |  |  |\n| --- | --- | --- | --- | --- |\n| 􀀎 Swin-T | 267G 23.1 46.0 68.1 50.3 | 41.6 | 65.1 | 44.9 |\n| 􀀏 ConvNeXt-T | 262G 25.6 46.2 67.9 50.8 | 41.7 | 65.0 | 44.9 |\n|  | Cascade Mask-RCNN 3 􀀂 schedule |  |  |  |\n| 􀀏 ResNet-50 | 739G 11.4 46.3 64.3 50.5 | 40.1 | 61.7 | 43.4 |\n| 􀀏 X101-32 | 819G 9.2 48.1 66.5 52.4 | 41.6 | 63.9 | 45.2 |\n| 􀀏 X101-64 | 972G 7.1 48.3 66.4 52.3 | 41.7 | 64.0 | 45.1 |\n| 􀀎 Swin-T | 745G 12.2 50.4 69.2 54.7 | 43.7 | 66.6 | 47.3 |\n| 􀀏 ConvNeXt-T | 741G 13.5 50.4 69.1 54.8 | 43.7 | 66.5 | 47.3 |\n| 􀀎 Swin-S | 838G 11.4 51.9 70.7 56.3 | 45.0 | 68.2 | 48.8 |\n| 􀀏 ConvNeXt-S | 827G 12.0 51.9 70.8 56.5 | 45.0 | 68.4 | 49.1 |\n| 􀀎 Swin-B | 982G 10.7 51.9 70.5 56.4 | 45.0 | 68.1 | 48.9 |\n| 􀀏 ConvNeXt-B | 964G 11.4 52.7 71.3 57.2 | 45.6 | 68.9 | 49.5 |\n\n| ImageNet-1K pre-trained\u003Cbr>􀀎 Swin-T 5122 45.8 60M 945G |\n| --- |\n| 􀀏 ConvNeXt-T 5122 46.7 60M 939G |\n| 􀀎 Swin-S 5122 49.5 81M 1038G |\n| 􀀏 ConvNeXt-S 5122 49.6 82M 1027G |\n| 􀀎 Swin-B 5122 49.7 121M 1188G |\n| 􀀏 ConvNeXt-B 5122 49.9 122M 1170G |\n| ImageNet-22K pre-trained\u003Cbr>􀀎 Swin-Bz 6402 51.7 121M 1841G |\n| 􀀏 ConvNeXt-Bz 6402 53.1 122M 1828G |\n| 􀀎 Swin-Lz 6402 53.5 234M 2468G |\n| 􀀏 ConvNeXt-Lz\u003Cbr>􀀏 ConvNeXt-XLz\u003Cbr>64026402\u003Cbr>53.7 54.0\u003Cbr>235M 391M\u003Cbr>2458G 3335G |\n\n\n| pre-training config | ConvNeXt-B/L\u003Cbr>ImageNet-1K\u003Cbr>2242 | ConvNeXt-B/L/XL\u003Cbr>ImageNet-22K\u003Cbr>2242 |\n| --- | --- | --- |\n| fine-tuning config | ImageNet-1K\u003Cbr>3842 | ImageNet-1K\u003Cbr>2242 and 3842 |\n| optimizer | AdamW | AdamW |\n| base learning rate | 5e-5 | 5e-5 |\n| weight decay | 1e-8 | 1e-8 |\n| optimizer momentum | 􀀌 1; 􀀌 2=0:9 ; 0:999 | 􀀌 1; 􀀌 2=0:9 ; 0:999 |\n| batch size | 512 | 512 |\n| training epochs | 30 | 30 |\n| learning rate schedule | cosine decay | cosine decay |\n| layer-wise lr decay | 0.7 | 0.8 |\n| warmup epochs | None | None |\n| warmup schedule | N/A | N/A |\n| randaugment | (9, 0.5) | (9, 0.5) |\n| label smoothing | 0.1 | 0.1 |\n| mixup | None | None |\n| cutmix | None | None |\n| stochastic depth | 0.8/0.95 | 0.2/0.3/0.4 |\n| layer scale | pre-trained | pre-trained |\n| gradient clip | None | None |\n| [exp. mov. avg](exp. mov. avg). (EMA) | None | None/None/0.9999 |\n\n\n| ImageNet-1K pre-trained |  ","cbCaigUfg7bXBlkz","https://ap.wps.com/l/cbCaigUfg7bXBlkz","pdf",742649,1,14,"Chinese","zh",105,"en","# 模型预训练规模与效果对比\n## ImageNet-1K预训练模型\n## ImageNet-22K预训练模型\n# 检测任务：Mask-RCNN与Cascade Mask-RCNN\n## 不同日程（schedule）下的结果\n# 微调与训练配置\n## 优化器与学习率策略\n## 数据增强与正则化参数","[{\"question\":\"文档比较了哪些主流视觉模型？\",\"answer\":\"包含RegNetY、EffNet（EfficientNet）、DeiT、Swin Transformer、ConvNeXt，以及部分ResNet与R-101/R-152等对照模型。\"},{\"question\":\"对比使用了哪些预训练数据集？\",\"answer\":\"主要使用ImageNet-1K预训练与ImageNet-22K预训练，并在相应设置下给出模型规模与结果指标。\"},{\"question\":\"下游任务采用了哪些检测框架与训练日程？\",\"answer\":\"包含Mask-RCNN的schedule对比，以及Cascade Mask-RCNN在不同schedule下的指标结果。\"},{\"question\":\"微调训练配置包含哪些关键超参数？\",\"answer\":\"文档列出优化器（如AdamW）、学习率与权重衰减、余弦学习率衰减、层级学习率衰减、stochastic depth，以及数据增强与标签平滑等参数。\"}]","Liu2022A - RegNetY/EffNet/DeiT/Swin/ConvNeXt 预训练与检测结果对比 | PDF",1788358660,35,{"code":4,"msg":32,"data":33},"ok",{"site_id":24,"language":23,"slug":34,"title":13,"keywords":35,"description":14,"schema_data":36,"social_meta":92,"head_meta":94,"extra_data":96,"updated_unix":29},"liu2022a-comparison-of-regnetyeffnetdeitswinconvnext-pretraining-and-detection-results","",{"@graph":37,"@context":91},[38,55,70],{"@type":39,"itemListElement":40},"BreadcrumbList",[41,45,49,52],{"item":42,"name":43,"@type":44,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":46,"name":47,"@type":44,"position":48},"https://docshare.wps.com/document/","Document",2,{"item":50,"name":12,"@type":44,"position":51},"https://docshare.wps.com/document/research-report/",3,{"item":53,"name":13,"@type":44,"position":54},"https://docshare.wps.com/document/liu2022a-comparison-of-regnetyeffnetdeitswinconvnext-pretraining-and-detection-results/184299/",4,{"url":53,"name":13,"@type":56,"author":57,"headline":13,"publisher":59,"fileFormat":62,"inLanguage":23,"description":14,"dateModified":63,"datePublished":64,"encodingFormat":62,"isAccessibleForFree":65,"interactionStatistic":66},"DigitalDocument",{"name":9,"@type":58},"Person",{"url":42,"name":60,"@type":61},"DocShare","Organization","application/pdf","2026-09-05","2026-09-02",true,{"@type":67,"interactionType":68,"userInteractionCount":20},"InteractionCounter",{"@type":69},"ViewAction",{"@type":71,"mainEntity":72},"FAQPage",[73,79,83,87],{"name":74,"@type":75,"acceptedAnswer":76},"文档比较了哪些主流视觉模型？","Question",{"text":77,"@type":78},"包含RegNetY、EffNet（EfficientNet）、DeiT、Swin Transformer、ConvNeXt，以及部分ResNet与R-101/R-152等对照模型。","Answer",{"name":80,"@type":75,"acceptedAnswer":81},"对比使用了哪些预训练数据集？",{"text":82,"@type":78},"主要使用ImageNet-1K预训练与ImageNet-22K预训练，并在相应设置下给出模型规模与结果指标。",{"name":84,"@type":75,"acceptedAnswer":85},"下游任务采用了哪些检测框架与训练日程？",{"text":86,"@type":78},"包含Mask-RCNN的schedule对比，以及Cascade Mask-RCNN在不同schedule下的指标结果。",{"name":88,"@type":75,"acceptedAnswer":89},"微调训练配置包含哪些关键超参数？",{"text":90,"@type":78},"文档列出优化器（如AdamW）、学习率与权重衰减、余弦学习率衰减、层级学习率衰减、stochastic depth，以及数据增强与标签平滑等参数。","https://schema.org",{"og:url":53,"og:type":93,"og:title":13,"og:site_name":60,"og:description":14},"article",{"robots":95,"canonical":53},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":98},[99,103,107,111,116,121,126,129,134,137,141],{"id":20,"doc_module":4,"doc_module_name":47,"category_name":100,"show_sort_weight":101,"slug":102},"Story & Novel",90,"story-novel",{"id":48,"doc_module":4,"doc_module_name":47,"category_name":104,"show_sort_weight":105,"slug":106},"Literature",80,"literature",{"id":54,"doc_module":4,"doc_module_name":47,"category_name":108,"show_sort_weight":109,"slug":110},"Exam",70,"exam",{"id":112,"doc_module":4,"doc_module_name":47,"category_name":113,"show_sort_weight":114,"slug":115},5,"Comic",60,"comic",{"id":117,"doc_module":4,"doc_module_name":47,"category_name":118,"show_sort_weight":119,"slug":120},6,"Technology",50,"technology",{"id":122,"doc_module":4,"doc_module_name":47,"category_name":123,"show_sort_weight":124,"slug":125},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":47,"category_name":12,"show_sort_weight":127,"slug":128},30,"research-report",{"id":130,"doc_module":4,"doc_module_name":47,"category_name":131,"show_sort_weight":132,"slug":133},9,"Religion & Spirituality",20,"religion-spirituality",{"id":132,"doc_module":4,"doc_module_name":47,"category_name":135,"show_sort_weight":132,"slug":136},"World Cup","world-cup",{"id":138,"doc_module":4,"doc_module_name":47,"category_name":139,"show_sort_weight":138,"slug":140},10,"Lifestyle","lifestyle",{"id":142,"doc_module":4,"doc_module_name":47,"category_name":143,"show_sort_weight":112,"slug":144},19,"General","general"]