{"id":928,"date":"2023-08-29T19:52:22","date_gmt":"2023-08-29T11:52:22","guid":{"rendered":"http:\/\/www.tamanegi.xyz\/?p=928"},"modified":"2023-08-29T19:54:49","modified_gmt":"2023-08-29T11:54:49","slug":"omni-scale-feature-learning","status":"publish","type":"post","link":"http:\/\/tamanegi.xyz\/?p=928","title":{"rendered":"Omni-Scale Feature Learning"},"content":{"rendered":"

Methods<\/h1>\n

OSNet\u7528\u4e8e\u7ad9\u4eec\u5b66\u4e60\u884c\u4eba\u91cd\u8bc6\u522b\u4efb\u52a1\u7684\u5168\u5c3a\u5ea6\u7279\u5f81\u8868\u793a\u3002\u8be5\u7ae0\u8282\u4ece\u6df1\u5ea6\u53ef\u5206\u79bb\u5377\u79ef\u5f00\u59cb\uff0c\u7136\u540e\u5f15\u5165omni-scale residual block\u548cunified aggregation gate\u3002<\/p>\n

\u6df1\u5ea6\u53ef\u5206\u79bb\u5377\u79ef(DWConv)<\/h2>\n

\u76ee\u7684<\/h4>\n

\u51cf\u5c11\u53c2\u6570\u91cf\uff0c\u964d\u4f4e\u8ba1\u7b97\u6210\u672c<\/p>\n

\u65b9\u6cd5<\/h4>\n

\u5c06\u5377\u79ef\u5206\u4e3a\u6df1\u5ea6\u5377\u79ef\u548c\u9010\u70b9\u5377\u79ef\u3002
\n\u672c\u6587\u5b9e\u9a8c\u8868\u660e\u5bf9\u4e8e Omni-scale feature learning\u5148\u9010\u70b9\u5377\u79ef\u518d\u6df1\u5ea6\u5377\u79ef\u6548\u679c\u6bd4\u5148\u6df1\u5ea6\u5377\u79ef\u518d\u9010\u70b9\u5377\u79ef\u597d<\/font><\/p>\n

Link\uff5c\u5bf9\u6bd4<\/h4>\n

Run, Don\u2019t Walk: Chasing Higher FLOPS for Faster Neural Networks [CVPR 2023]->\u63d0\u51fa\u4e86\u66ff\u4ee3 DWConv \u7684\u5377\u79ef\u65b9\u5f0f\uff0c\u547d\u540d\u4e3a Partial \u5377\u79ef<\/p>\n

\u5907\u6ce8<\/h4>\n

\u8be5\u6a21\u5757\u88ab\u547d\u540d\u4e3a Lite 3X3<\/p>\n

Omni-Scale Residual Block<\/h2>\n

Baseline bottleneck:
\n\\begin{align}&y=x+\\tilde x \\&s.t.\\tilde x =F(x)\\end{align}
\nF: Lite 3X3 \u5b66\u4e60 scale=3 \u7684\u5355\u5c3a\u5ea6\u7279\u5f81\u3002<\/p>\n

\u5907\u6ce8<\/h4>\n

\u524d\u540e\u7684 1*1Layer\u53ea\u7528\u4e8e\u4fee\u6539\u7ef4\u5ea6\uff0c\u5bf9\u7a7a\u95f4\u4fe1\u606f\u7684\u805a\u5408\u6ca1\u6709\u5e2e\u52a9\u3002<\/p>\n

Multi-scale feature learning<\/h3>\n

\u76ee\u7684<\/h4>\n

\u5b9e\u73b0\u591a\u5c3a\u5ea6\u7279\u5f81\u5b66\u4e60<\/p>\n

\u65b9\u6cd5<\/h4>\n

\u5f15\u5165\u7ef4\u5ea6\u6307\u6570 t,\u5bf9\u4e8e\u6b8b\u5dee\u51fd\u6570 F\uff0ct \u7ef4\u5ea6\u5c31\u5806\u53e0 t \u4e2aLite 3X3->\u4ea7\u751f\u5927\u5c0f\u4e3a(2t+1)*(2t+1)\u7684\u611f\u53d7\u91ce\u3002
\n\u5b66\u4e60\u7684\u6b8b\u5dee\u53d8\u4e3a
\n\\begin{align}\\tilde x=\\sum^T_{t=1}F^t(x),\\quad s.t. T\\ge 1\\end{align}
\nOSNet \u8bbe\u7f6e\u6700\u5927 t =4\uff0c\u611f\u53d7\u91ce\u4e3a 9*9<\/p>\n

Unified aggregation gate<\/h2>\n

\u76ee\u7684<\/h4>\n

\u66f4\u597d\u7684\u5b66\u4e60\u4e0d\u540c\u5c3a\u5ea6\u7684\u7279\u5f81<\/p>\n

\u65b9\u6cd5<\/h4>\n

\u901a\u8fc7 Unified aggregation gate\u4ee5\u52a8\u6001\u7684\u65b9\u5f0f\u7ec4\u5408\u4e0d\u540c\u6d41\u7684\u7279\u5f81
\n\\begin{align}
\n&\\tilde x = \\sum^T{t=1}G(x^t)\\odot x^t,\\
\ns.t.&x^t\\triangleq F^t(x)
\n\\end{align}
\nG(x^t)\u662f\u957f\u5ea6\u8de8\u8d8ax^t\u6574\u4e2a\u901a\u9053\u7ef4\u5ea6\u7684\u5411\u91cf\uff0c\\odot\u662f\u70b9\u4e58
\n\u5728\u4ee3\u7801\u4e2d\uff0c\u7528\u4e00\u4e2a\u8ff7\u4f60\u7f51\u7edc\u5b9e\u73b0 G\uff0c\u7531\u975e\u53c2\u6570\u5e73\u5747\u6c60\u5316\u5c42\uff0c\u5e26\u6709\u4e00\u4e2a ReLU \u6fc0\u6d3b\u9690\u85cf\u5c42\u7684\u591a\u5c42\u611f\u77e5\u5668\u548c\u4e00\u4e2a sigmoid \u7ec4\u6210
\n\u8fd9\u91cc\u63d0\u5230\u4e86\u6839\u636e\u53e6\u5916\u4e24\u7bc7\u6587\u7ae0\u5c06\u6bd4\u7387\u8bbe\u4e3a 16 \u6765\u51cf\u5c11 MLP \u9690\u85cf\u7ef4\u5ea6<\/em>
\n\u6743\u91cd\u662f\u6839\u636e\u8f93\u5165\u6570\u636e\u52a8\u6001\u8ba1\u7b97\u7684\u3002\u8fd9\u5bf9\u4e8eRe-ID\u81f3\u5173\u91cd\u8981\uff0c\u56e0\u4e3a\u6d4b\u8bd5\u56fe\u50cf\u5305\u542b\u4e0e\u8bad\u7ec3\u4e2d\u8eab\u4efd\u4e0d\u540c\u7684\u4eba\uff1b\u56e0\u6b64\uff0c\u66f4\u9700\u8981\u81ea\u9002\u5e94\/\u8f93\u5165\u76f8\u5173\u7684\u7279\u5f81\u5c3a\u5ea6\u878d\u5408\u7b56\u7565\u3002 <\/font>
\nAggregation Gate \u5728\u6240\u6709 feature stream\u4e2d\u5171\u4eab\uff0c\u7531\u6b64\u5e26\u6765\u7684\u635f\u5931\u51fd\u6570\u4e3a
\n\\begin{align} \\frac{\\partial\\mathrm L}{\\partial G}=\\frac{\\partial L}{\\partial\\tilde x}\\frac{\\partial\\tilde x}{\\partial G}=\\frac{\\partial L}{\\partial\\tilde x} (\\sum^T<\/em>{t=1}x^t) \\end{align} <\/p>\n

Network Architecture<\/h1>\n

\"\"<\/p>\n","protected":false},"excerpt":{"rendered":"

Methods OSNet\u7528\u4e8e\u7ad9\u4eec\u5b66\u4e60\u884c\u4eba\u91cd\u8bc6\u522b\u4efb\u52a1\u7684\u5168\u5c3a\u5ea6\u7279\u5f81\u8868\u793a\u3002\u8be5\u7ae0\u8282\u4ece\u6df1\u5ea6\u53ef\u5206\u79bb\u5377\u79ef\u5f00\u59cb\uff0c\u7136\u540e\u5f15\u5165om […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[105],"tags":[],"_links":{"self":[{"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=\/wp\/v2\/posts\/928"}],"collection":[{"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=928"}],"version-history":[{"count":4,"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=\/wp\/v2\/posts\/928\/revisions"}],"predecessor-version":[{"id":940,"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=\/wp\/v2\/posts\/928\/revisions\/940"}],"wp:attachment":[{"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=928"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=928"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=928"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}