{"id":1069491,"date":"2022-04-29T21:03:00","date_gmt":"2022-04-30T04:03:00","guid":{"rendered":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/?post_type=msr-blog-post&p=1069491"},"modified":"2024-09-25T01:26:41","modified_gmt":"2024-09-25T08:26:41","slug":"on-the-connection-between-local-attention-and-dynamic-depth-wise-convolution","status":"publish","type":"msr-blog-post","link":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/articles\/on-the-connection-between-local-attention-and-dynamic-depth-wise-convolution\/","title":{"rendered":"ICLR 2022 Spotlight\uff1aTransformer\u4e0d\u6bd4CNN\u5f3a\uff01Local Attention\u548c\u52a8\u6001Depth-wise\u5377\u79ef\u7684\u524d\u4e16\u4eca\u751f"},"content":{"rendered":"\n

\u8fd1\u4e24\u5e74\u6765\uff0c\u6709\u5173Transformer\u7684\u6587\u7ae0\u53ef\u8c13\u662f\u4e95\u55b7\u5f0f\u7206\u53d1\uff0c\u8bb8\u591a\u7814\u7a76\u5458\u90fd\u5728\u8bbe\u8ba1\u5404\u79cd\u4efb\u52a1\u4e0a\u7684Transformer\u6a21\u578b\u3002\u7136\u800c\uff0cAttention\u4f5c\u4e3aTransformer\u7684\u6838\u5fc3\u6a21\u5757\uff0c\u771f\u7684\u6bd4\u5377\u79ef\u5f3a\u5417\uff1f\u4eca\u5929\u7684\u8fd9\u7bc7\u6587\u7ae0\u6216\u8bb8\u80fd\u5e26\u7ed9\u4f60\u65b0\u89c6\u89d2\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u4eceLocal Attention\u548cDynamic Depth-wise Convolution\u7684\u89c6\u89d2\u51fa\u53d1\u53d1\u73b0\uff0c\u8bbe\u8ba1\u597d\u7684\u5377\u79ef\u7ed3\u6784\u5e76\u4e0d\u6bd4Transformer\u5dee\uff01\u76f8\u5173\u8bba\u6587\u201cOn the Connection between Local Attention and Dynamic Depth-wise Convolution\u201d\u5df2\u88abICLR 2022\u6536\u5f55\u3002<\/p>\n\n\n\n

\u8bba\u6587\u94fe\u63a5: https:\/\/arxiv.org\/abs\/2106.04263<\/p>\n\n\n\n

\u4ee3\u7801\u94fe\u63a5: https:\/\/github.com\/Atten4Vis\/DemystifyLocalViT<\/p>\n\n\n\n

Local Attention\u662f\u4ec0\u4e48\uff1f<\/h2>\n\n\n\n

2020\u5e74\uff0cViT\u6a2a\u7a7a\u51fa\u4e16\uff0c\u5e2d\u5377\u4e86\u6a21\u578b\u8bbe\u8ba1\u9886\u57df\u3002\u5404\u79cd\u57fa\u4e8eTransformer\u7684\u7ed3\u6784\u5f00\u59cb\u88ab\u63d0\u51fa\uff0c\u4e00\u4e9b\u5728\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u4e2d\u53d6\u5f97\u6210\u529f\u7684\u5148\u9a8c\u77e5\u8bc6\uff0c\u5982Local Operation\u3001\u591a\u5c3a\u5ea6\u3001Shuffle\u7b49\u5404\u79cd\u64cd\u4f5c\u548cinductive bias\u88ab\u5f15\u5165Transformer\u4e4b\u4e2d\u3002\u5176\u4e2d\u6bd4\u8f83\u6210\u529f\u7684\u6848\u4f8b\u4fbf\u662f\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684Swin Transformer\uff0c\u901a\u8fc7\u5c06Local Operation\u5f15\u5165ViT\u5f53\u4e2d\uff0c\u5229\u7528shifted window\u4e00\u4e3e\u83b7\u5f97\u591a\u4e2a\u4efb\u52a1\u7684SOTA\u7ed3\u679c\uff0c\u5e76\u83b7\u5f97\u4e862021\u5e74 ICCV \u6700\u4f73\u8bba\u6587\u5956\u2014\u2014\u9a6c\u5c14\u5956\u3002\u90a3\u4e48Swin Transformer\u4e2d\u6700\u6838\u5fc3\u7684\u6a21\u5757\uff0cLocal Attention\u7a76\u7adf\u6709\u4ec0\u4e48\u795e\u79d8\u4e4b\u5904\u5462?<\/p>\n\n\n\n

Local Attention\u672c\u8d28\u4e0a\u662f\u5728\u4e00\u4e2a2D Local Window\u5185\u8fdb\u884c\u7279\u5f81\u805a\u5408\uff0c\u5176\u6bcf\u4e2a\u4f4d\u7f6e\u7684\u805a\u5408\u6743\u91cd\u53ef\u4ee5\u901a\u8fc7KQV\u4e4b\u95f4\u8ba1\u7b97Attention Similarity\u5f97\u5230\uff08\u4e3b\u8981\u5305\u62ecdot-production, scaling, softmax\uff09\uff0c\u662f\u4e00\u4e2a\u65e0\u53c2\u6570\u7684\u3001\u52a8\u6001\u8ba1\u7b97\u7684\u5c40\u90e8\u7279\u5f81\u8ba1\u7b97\u6a21\u5757\u3002<\/p>\n\n\n\n

\"aij\u4e3a\u805a\u5408\u6743\u91cd\uff0cxij\u4e3a\u5f85\u805a\u5408\u7684\u7279\u5f81\"
aij\u4e3a\u805a\u5408\u6743\u91cd\uff0cxij\u4e3a\u5f85\u805a\u5408\u7684\u7279\u5f81<\/em><\/figcaption><\/figure>\n\n\n\n

\u6b64\u6b21ICLR\u8fd9\u7bc7\u6587\u7ae0\u7684\u5148\u524d\u7248\u672c \uff08Demystifying Local Vision Transformer\uff09\u57282021\u5e746\u6708\u5c31\u9996\u6b21\u5728arXiv\u516c\u5e03\u5e76\u5206\u6790\u4e86Local Attention\u4e09\u4e2a\u5f3a\u5927\u7684\u8bbe\u8ba1\u539f\u5219\uff1a<\/p>\n\n\n\n

(1) \u7a00\u758f\u8fde\u63a5<\/strong>\u2014\u2014\u6307\u4e00\u4e9b\u8f93\u51fa\u53d8\u91cf\u548c\u4e00\u4e9b\u8f93\u5165\u53d8\u91cf\u4e4b\u95f4\u6ca1\u6709\u76f8\u4e92\u8fde\u63a5\u3002\u5b83\u6709\u6548\u5730\u51cf\u5c11\u4e86\u6a21\u578b\u7684\u590d\u6742\u5ea6\u800c\u4e0d\u51cf\u5c11\u8f93\u5165\u8f93\u51fa\u53d8\u91cf\u4e2a\u6570\u3002\u5728Local Attention\u5f53\u4e2d\uff0c\u7a00\u758f\u8fde\u63a5\u4f53\u73b0\u5728\u4e24\u4e2a\u65b9\u9762\uff1a\u4e00\u662fLocal Attention\u5728\u56fe\u50cf\u7a7a\u95f4\u4e0a\uff0c\u6bcf\u4e00\u4e2a\u8f93\u51fa\u503c\u4ec5\u4e0e\u5c40\u90e8\u7684Local Window\u5185\u7684\u8f93\u5165\u76f8\u8fde\u63a5\uff0c\u4e0eViT\u7684\u5168\u50cf\u7d20\uff08token\uff09\u8fde\u63a5\u4e0d\u540c\u3002\u4e8c\u662fLocal Attention\u5728\u901a\u9053\u4e0a\uff0c\u6bcf\u4e00\u4e2a\u8f93\u51fa\u901a\u9053\u4ec5\u4e0e\u4e00\u4e2a\u8f93\u5165\u901a\u9053\u8fde\u63a5\uff0c\u6ca1\u6709\u4ea4\u53c9\u8fde\u63a5\uff0c\u4e0d\u540c\u4e8egroup convolution\u4e0enormal convolution\u3002<\/p>\n\n\n\n

(2) \u6743\u91cd\u5171\u4eab<\/strong>\u2014\u2014\u610f\u5473\u7740\u6709\u4e00\u4e9b\u8fde\u63a5\u7684\u6743\u91cd\u662f\u76f8\u540c\u4e14\u5171\u4eab\u7684\uff0c\u5b83\u964d\u4f4e\u4e86\u6a21\u578b\u7684\u53c2\u6570\u91cf\uff0c\u540c\u65f6\u4e0d\u9700\u8981\u589e\u52a0\u8bad\u7ec3\u6570\u636e\u5373\u53ef\u589e\u5f3a\u6a21\u578b\u3002\u5728\u6a21\u578b\u4e2d\uff0c\u4e00\u4e2a\u6743\u91cd\u88ab\u5171\u4eab\u4f7f\u7528\u53ef\u4ee5\u88ab\u8ba4\u4e3a\u9488\u5bf9\u8be5\u6743\u91cd\u7684\u8bad\u7ec3\u6837\u672c\u589e\u52a0\uff0c\u6709\u52a9\u4e8e\u6a21\u578b\u7684\u4f18\u5316\u3002\u5728Local Attention\u4e2d\uff0c\u6743\u91cd\u5171\u4eab\u901a\u8fc7multi-head self-attention\u6765\u5b9e\u73b0\uff0c\u518d\u628a\u901a\u9053\u5206\u6210head\uff08group\uff09\uff0c\u5728\u540c\u4e00\u4e2ahead\u5185\u5171\u4eab\u4f7f\u7528\u4e00\u7ec4\u805a\u5408\u6743\u91cd\uff0c\u964d\u4f4e\u805a\u5408\u6743\u91cd\u7684\u53c2\u6570\u91cf\uff08\u975e\u6a21\u578b\u53c2\u6570\u91cf\uff09\u3002<\/p>\n\n\n\n

(3) \u52a8\u6001\u6743\u91cd<\/strong>\u2014\u2014\u6307\u6839\u636e\u4e0d\u540c\u6837\u672c\u7684\u7279\u5f81\uff0c\u52a8\u6001\u5730\u4ea7\u751f\u8fde\u63a5\u6743\u91cd\u3002\u5b83\u80fd\u591f\u589e\u52a0\u6a21\u578b\u7684\u5bb9\u91cf\u3002\u5982\u679c\u628a\u8fde\u63a5\u6743\u91cd\u770b\u4f5c\u662f\u9690\u5c42\u53d8\u91cf\uff0c\u90a3\u4e48\u8fd9\u79cd\u52a8\u6001\u6743\u91cd\u53ef\u4ee5\u770b\u4f5c\u662f\u589e\u52a0\u6a21\u578b\u5bb9\u91cf\u7684\u4e8c\u9636\u64cd\u4f5c\u3002Local Attention\u7684\u52a8\u6001\u6743\u91cd\u4f53\u73b0\u5728\u6bcf\u4e00\u4e2a\u8fde\u63a5\u7684\u805a\u5408\u6743\u91cd\u90fd\u662f\u6839\u636e\u6837\u672c\u7279\u5f81\u4f7f\u7528\u57fa\u4e8edot-product\u7684\u65b9\u5f0f\u8ba1\u7b97\u5f97\u5230\u7684\u3002<\/p>\n\n\n\n

\u901a\u8fc7\u4ee5\u4e0a\u4e09\u4e2a\u6a21\u578b\u8bbe\u8ba1\u539f\u5219\uff0cLocal Attention\u8868\u73b0\u51fa\u4f18\u5f02\u7684\u6548\u679c\u3002\u7136\u800c\uff0c\u8fd9\u4e9b\u7279\u6027\u4e5f\u5929\u7136\u5b58\u5728\u4e8eCNN\u7ed3\u6784\u5f53\u4e2d\uff0c\u5c24\u5176\u662f\uff08Dynamic\uff09Depth-wise\u5377\u79ef\u3002<\/p>\n\n\n\n

\uff08Dynamic\uff09Depth-wise\u5377\u79ef\u548cLocal Attention\u7684\u524d\u4e16\u4eca\u751f<\/h2>\n\n\n\n

\u9010\u6b65\u62c6\u89e3Local Attention\u7684\u64cd\u4f5c\uff0c\u53ef\u4ee5\u53d1\u73b0\u5728\u7a00\u758f\u8fde\u63a5\u3001\u6743\u91cd\u5171\u4eab\u3001\u52a8\u6001\u6743\u91cd\u4e09\u4e2a\u7ef4\u5ea6\u4e0a\uff0c\u90fd\u4e0e\u201c\u5386\u53f2\u9738\u4e3b\u201dCNN\u7ed3\u6784\u4e2d\u7684\uff08Dynamic\uff09Depth-wise\u5377\u79ef\u5f88\u76f8\u4f3c\u3002Depth-wise\u5377\u79ef\u53ef\u8c13\u662f\u4e00\u4e2a\u88ab\u957f\u671f\u4f7f\u7528\u53c8\u88ab\u6e10\u6e10\u9057\u5fd8\u5728\u5386\u53f2\u957f\u6cb3\u4e2d\u7684\u7ed3\u6784\uff0c\u90a3\u4e48\u5176\u5728\u6a21\u578b\u8bbe\u8ba1\u4e0a\u53c8\u6709\u54ea\u4e9b\u51c6\u5219\u5462\uff1f<\/p>\n\n\n\n

(1) \u7a00\u758f\u8fde\u63a5<\/strong>\u3002\u4e0d\u96be\u53d1\u73b0\uff0cDepth-wise\u5377\u79ef\u7684\u7a00\u758f\u8fde\u63a5\u7279\u6027\u4e0eLocal Attention\u5b8c\u5168\u76f8\u540c\uff0c\u5728\u56fe\u50cf\u7a7a\u95f4\u4e0a\u5c40\u90e8\u94fe\u63a5\uff0c\u5728\u901a\u9053\u4e0a\u7a00\u758f\u8fde\u63a5\u3002<\/p>\n\n\n\n

(2) \u6743\u91cd\u5171\u4eab<\/strong>\u3002\u6743\u91cd\u5171\u4eab\u7684\u6982\u5ff5\u6700\u521d\u5c31\u8bde\u751f\u4e8e\u5377\u79ef\u64cd\u4f5c\u4e4b\u4e2d\uff0cDepth-wise\u5377\u79ef\u540c\u6837\u5f97\u76ca\u4e8e\u6743\u91cd\u5171\u4eab\u64cd\u4f5c\uff0c\u4f46\u4e0eLocal Attention\u7565\u6709\u4e0d\u540c\uff0cDepth-wise\u5377\u79ef\u5728\u56fe\u50cf\u7a7a\u95f4\u4e0a\u5171\u4eab\u6743\u91cd\uff0c\u6bcf\u4e00\u4e2a\u7a7a\u95f4\u4f4d\u7f6e\u90fd\u662f\u7528\u76f8\u540c\u6743\u91cd\u7684\u5377\u79ef\u6838\u6765\u8fdb\u884c\u7279\u5f81\u805a\u5408\uff0c\u800c\u5728\u901a\u9053\u4e0a\uff0c\u6bcf\u4e00\u4e2a\u901a\u9053\u4f1a\u4f7f\u7528\u72ec\u7acb\u7684\u805a\u5408\u6743\u91cd\u3002<\/p>\n\n\n\n

(3) \u52a8\u6001\u6743\u91cd<\/strong>\u3002\u52a8\u6001\u6743\u91cd\u7684\u8bbe\u8ba1\u539f\u5219\u5728\u539f\u59cb\u7684Depth-wise\u5377\u79ef\u4e2d\u5e76\u6ca1\u6709\u88ab\u4f7f\u7528\uff0c\u7136\u800c\uff0c\u52a8\u6001\u5377\u79ef\u4f5c\u4e3a\u4e00\u4e2a\u88ab\u5e7f\u6cdb\u7814\u7a76\u7684\u9886\u57df\uff0c\u53ef\u4ee5\u8f7b\u6613\u5730\u5c06\u52a8\u6001\u7279\u6027\u5f15\u5165Depth-wise\u5377\u79ef\u4e2d\uff0c\u5f62\u6210\u7279\u5f81\u4f9d\u8d56\u7684\u805a\u5408\u6743\u91cd\u3002<\/p>\n\n\n\n

\u5c3d\u7ba1\u5728\u6743\u91cd\u5171\u4eab\u4e0a\u4e24\u8005\u7684\u5171\u4eab\u65b9\u5f0f\u4e0d\u540c\uff0c\u4f46\u7ecf\u8fc7\u5b9e\u9a8c\u53d1\u73b0\uff0c\u4ee5Local MLP\uff08\u53bb\u6389dynamic\u7279\u6027\u7684local attention\uff09\u4e3a\u4f8b\uff0c\u5728\u901a\u9053\u548c\u7a7a\u95f4\u7ef4\u5ea6\u4e0a\u5171\u4eab\u6743\u91cd\u7684\u5f71\u54cd\u5e76\u4e0d\u5927\uff0c\u5728\u4efb\u4f55\u4e00\u4e2a\u7ef4\u5ea6\u5171\u4eab\u6743\u91cd\uff0c\u5747\u53ef\u4ee5\u964d\u4f4e\u6a21\u578b\u7684\u53c2\u6570\u91cf\uff0c\u5e2e\u52a9\u6a21\u578b\u4f18\u5316\u3002\u800c\u5728\u52a8\u6001\u6743\u91cd\u4e0a\uff0c\u867d\u7136\u4e24\u8005\u4e0d\u540c\uff0c\u4f46Depth-wise\u5377\u79ef\u4ecd\u7136\u53ef\u4ee5\u8f7b\u6613\u5177\u5907\u52a8\u6001\u7279\u6027\u3002<\/p>\n\n\n\n

\"\u88681
\u88681 \u4e0d\u540c\u7ed3\u6784\u5728\u7a00\u758f\u8fde\u63a5\u3001\u6743\u91cd\u5171\u4eab\u3001\u52a8\u6001\u6743\u91cd\u4e0a\u7684\u5bf9\u6bd4\u3002 D-DW-Conv.\u8868\u793a\u52a8\u6001DW\u5377\u79ef<\/em><\/figcaption><\/figure>\n\n\n\n

Depth-wise\u5377\u79ef\u7684\u8868\u73b0\u529b<\/h2>\n\n\n\n

Depth-wise\u5377\u79ef\u548cLocal Attention\u7684\u8bbe\u8ba1\u539f\u5219\u5982\u6b64\u76f8\u4f3c\uff0c\u90a3\u4e48\u4e3a\u4ec0\u4e48Local Attention\u53d6\u5f97\u4e86\u5982\u6b64\u9ad8\u7684\u8868\u73b0\u529b\uff0c\u800cDepth-wise\u5377\u79ef\u5374\u6ca1\u6709\u5462\uff1f\u4e3a\u4e86\u9a8c\u8bc1\u8fd9\u4e00\u95ee\u9898\uff0c\u7814\u7a76\u5458\u4eec\u4f7f\u7528Depth-wise\u5377\u79ef\u66ff\u6362Swin Transformer\u4e2d\u7684\u6240\u6709Local Attention\u6a21\u5757\uff0c\u5176\u4ed6\u7ed3\u6784\u4fdd\u6301\u4e0d\u53d8\uff08pre-LN\u4fee\u6539\u4e3apost-BN\uff09\uff1b\u540c\u65f6\uff0c\u4e3a\u4e86\u9a8c\u8bc1\u52a8\u6001DW\u5377\u79ef\u7684\u6548\u679c\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u6784\u5efa\u4e86\u4e24\u79cd\u52a8\u6001\u7279\u6027\u7684Depth-wise\u5377\u79ef\uff1a<\/p>\n\n\n\n

\uff081\uff09D-DW-Conv\u3002\u7b2c\u4e00\u79cd\u52a8\u6001DW\u5377\u79ef\uff0c\u91c7\u7528\u4e86\u548c\u666e\u901aDW\u5377\u79ef\u76f8\u540c\u7684\u6743\u91cd\u5171\u4eab\u65b9\u5f0f\u2014\u2014\u56fe\u50cf\u7a7a\u95f4\u5171\u4eab\u5377\u79ef\u6838\u548c\u901a\u9053\u95f4\u72ec\u7acb\u5377\u79ef\u6838\u3002\u5e76\u4f7f\u7528\u4e86Global Average Pooling\u6765\u5904\u7406\u8f93\u5165\u7279\u6027\uff0c\u7136\u540e\u901a\u8fc7FC Layer\u6765\u52a8\u6001\u5730\u9884\u6d4b\u51fa\u52a8\u6001\u5377\u79ef\u6838\u3002<\/p>\n\n\n\n

\uff082\uff09I-D-DW-Conv\u3002\u7b2c\u4e8c\u79cd\u52a8\u6001DW\u5377\u79ef\uff0c\u91c7\u7528\u4e86\u548cLocal Attention\u76f8\u540c\u7684\u6743\u91cd\u5171\u4eab\u65b9\u5f0f\uff0c\u6bcf\u4e2a\u50cf\u7d20\uff08token\uff09\u91c7\u7528\u72ec\u7acb\u7684\u805a\u5408\u6743\u91cd\uff0c\u800c\u5728\u901a\u9053head\uff08group\uff09\u4e2d\u5171\u4eab\u6743\u91cd\u3002\u79f0\u4e3aInhomogeneous Dynamic DW\u5377\u79ef\u3002<\/p>\n\n\n\n

\u6765\u770b\u4e00\u770b\u5b9e\u9a8c\u7ed3\u679c\u5982\u4f55:<\/p>\n\n\n\n

\"\u88682
\u88682 ImageNet1k, COCO, ADE20K\u5bf9\u6bd4\u7ed3\u679c<\/em><\/figcaption><\/figure>\n\n\n\n

\u672c\u6587\u4f7f\u7528\u4e86\u4e0eSwin Transformer\u5b8c\u5168\u76f8\u540c\u7684\u8bad\u7ec3\u53c2\u6570\u548c\u7f51\u7edc\u7ed3\u6784\uff0c\u5728ImageNet\u5206\u7c7b\u3001COCO\u68c0\u6d4b\u3001ADE20K\u8bed\u4e49\u5206\u5272\u4e2d\uff0c\u57fa\u4e8eDepth-wise\u5377\u79ef\u7684\u7ed3\u6784\u53d6\u5f97\u4e86\u548cSwin Transformer\u76f8\u540c\u7684\u8868\u73b0\u529b\uff0c\u800cDepth-wise\u5377\u79ef\u7684\u8ba1\u7b97\u5f00\u9500\u8fd8\u66f4\u5c0f\u3002<\/p>\n\n\n\n

\u6240\u4ee5\uff0cDepth-wise\u5377\u79ef\u771f\u7684\u4e0d\u5dee\uff01\u6709\u4eba\u4f1a\u95ee\uff0c\u662f\u4e0d\u662f\u5728\u66f4\u5927\u7684\u6a21\u578b\u548c\u66f4\u5927\u7684\u6570\u636e\u96c6\u4e0a\uff0cLocal Attention\u4f1a\u66f4\u6709\u4f18\u52bf\u5462\uff1f\u7531\u4e8e\u8ba1\u7b97\u8d44\u6e90\u7684\u9650\u5236\uff0c\u7814\u7a76\u5458\u4eec\u4ec5\u5728Base\u6a21\u578b\u4e0a\u8fdb\u884c\u4e86\u90e8\u5206\u5b9e\u9a8c\uff1a<\/p>\n\n\n\n

\"\u88683
\u88683 ImageNet22k Pretraining<\/em><\/figcaption><\/figure>\n\n\n\n

\u5728\u66f4\u5927\u7684\u6570\u636e\u96c6ImageNet22k\u4e0a\u9884\u8bad\u7ec3\u7684\u7ed3\u679c\u8868\u660e\uff0c\u57fa\u4e8eDepth-wise\u5377\u79ef\u7684\u7ed3\u6784\u4ecd\u7136\u4e0eLocal Attention\u4e0d\u76f8\u4e0a\u4e0b\u3002\u8fd1\u671f\u7684ConvNext [1] \u548cRepLKNet [2] \u5de5\u4f5c\u66f4\u662f\u7ed9\u51fa\u4e86\u8bc1\u660e\u3002<\/p>\n\n\n\n

\u73b0\u4ee3\u5377\u79ef\u5f3a\u5728\u54ea\u91cc\uff0c\u5982\u4f55\u8bbe\u8ba1\u66f4\u597d\u7684\u6a21\u578b<\/h2>\n\n\n\n

\u65e2\u7136Depth-wise\u5377\u79ef\u7684\u8868\u73b0\u529b\u4e5f\u8fd9\u4e48\u597d\uff0c\u90a3\u4e3a\u4ec0\u4e48\u957f\u4e45\u4ee5\u6765\u6ca1\u6709\u5f15\u8d77\u4eba\u4eec\u7684\u5e7f\u6cdb\u5173\u6ce8\u5462\uff1f\u901a\u8fc7\u4e0e\u4f20\u7edf\u7684CNN\u5bf9\u6bd4\u53d1\u73b0\uff0c\u73b0\u4ee3\u5377\u79ef\u5728\u8bbe\u8ba1\u4e0a\u4e00\u822c\u6ee1\u8db3\u6587\u7ae0\u6307\u51fa\u7684\u4e09\u4e2a\u8bbe\u8ba1\u539f\u5219\u3002\u540c\u65f6\uff0c\u4e0e\u4f20\u7edfCNN\u76f8\u6bd4\uff0cSwin Transformer\u7b49\u7ed3\u6784\u4f7f\u7528\u4e86\u66f4\u5927\u7684Kernel Size\uff0c \u59827×7, 12×12\u7b49\uff0c\u8fdc\u5927\u4e8eCNN\u4e2d\u957f\u4e45\u4f7f\u7528\u76843×3\u5377\u79ef\u3002<\/p>\n\n\n\n

Depth-wise\u5377\u79ef\uff0c\u914d\u5408\u5408\u7406\u7684\u52a8\u6001\u7279\u6027\u4ee5\u53ca\u5927kernel size\uff0c\u518d\u52a0\u4e0a\u73b0\u4ee3\u7f51\u7edc\u8bad\u7ec3\uff08\u6570\u636e\u589e\u5f3a\u3001\u4f18\u5316\u3001\u6b63\u5219\u5316\uff09\u7b49\u7b56\u7565\uff0c\u5c06\u662f\u73b0\u4ee3\u5377\u79ef\u7684\u5f3a\u5927\u4e4b\u5904\u3002<\/p>\n\n\n\n

\u81f3\u4e8e\u5982\u4f55\u8bbe\u8ba1\u66f4\u597d\u7684\u6a21\u578b\uff0c\u9996\u5148\u8981\u5206\u6790\u73b0\u6709\u4f18\u79c0\u6a21\u578b\u7684\u5171\u6027\u3002\u5728\u7a00\u758f\u8fde\u63a5\u65b9\u9762\uff0c\u56fe1\u5c55\u793a\u4e86\u4e0d\u540c\u7ed3\u6784\u7684\u7a00\u758f\u7279\u6027\u3002\u8d8a\u7a00\u758f\u7684\u6a21\u578b\uff0c\u8d8a\u6709\u52a9\u4e8e\u8bad\u7ec3\u9636\u6bb5\u7684\u4f18\u5316\uff0c\u4ea7\u751f\u66f4\u597d\u7684\u6027\u80fd\uff0c\u540c\u65f6\u964d\u4f4e\u6a21\u578b\u7684\u8ba1\u7b97\u590d\u6742\u5ea6\uff0c\u4f7f\u7f51\u7edc\u53ef\u4ee5\u6784\u5efa\u3001\u8bbe\u8ba1\u66f4\u81ea\u7531\u7684\u6a21\u5757\u3002<\/p>\n\n\n\n

\"\u56fe1
\u56fe1 (a) \u5377\u79ef (b) global attention (c) local attention\u3001DW\u5377\u79ef (d) 1×1\u5377\u79ef (e) fully-connected MLP<\/em><\/figcaption><\/figure>\n\n\n\n

\u6b64\u5916\uff0c\u6587\u7ae0\u8fd8\u6784\u5efa\u4e86\u4e00\u4e2aRelation Graph\u6765\u9610\u8ff0\u6a21\u578b\u7ed3\u6784\u8bbe\u8ba1\u4e2d\u4ea7\u751f\u7684\u4e00\u4e9b\u8bbe\u8ba1\u539f\u5219\u4e0a\u7684\u6f14\u8fdb\u8fc7\u7a0b\uff1a<\/p>\n\n\n\n

\"\u56fe2\u4e2dViT\u548cLocal
\u56fe2\u4e2dViT\u548cLocal ViT\u6307\u7ed3\u6784\u4e2d\u7684Attention\u7ed3\u6784\uff0cPVT\u4e3a\u4f4e\u79e9\u5f62\u5f0f\u4e0b\u7684\u91d1\u5b57\u5854\u7ed3\u6784Transformer\uff0cDim. Sep.\u8868\u793a\u901a\u9053\u7ef4\u5ea6\u4e0a\u7a00\u758f\u5316\uff0cLocality Sep.\u8868\u793a\u7a7a\u95f4\u4f4d\u7f6e\u8fde\u63a5\u7a00\u758f\u5316\uff0cLR\u8868\u793a\u4f4e\u79e9\uff0cMS Conv.\u8868\u793a\u591a\u5c3a\u5ea6\u5377\u79ef\u3002<\/em><\/figcaption><\/figure>\n\n\n\n

\u5173\u7cfb\u56fe\u4e2d\uff0c\u4ece\u4e0a\u5230\u4e0b\u5448\u73b0\u4e86\u4f9d\u6b21\u589e\u5f3a\u7684\u6b63\u5219\u5316\u65b9\u5f0f\u6216\u5f15\u5165\u52a8\u6001\u6743\u91cd\uff0c\u968f\u7740\u6b63\u5219\u5316\u53ca\u52a8\u6001\u6743\u91cd\u7684\u589e\u52a0\uff0c\u7f51\u7edc\u7684\u4eba\u4e3a\u5148\u9a8c\u4e5f\u968f\u4e4b\u589e\u52a0\u3002\u8fd9\u79cd\u5f62\u5f0f\u5e26\u6765\u4e86\u4f18\u5316\u4e0a\u7684\u597d\u5904\uff0c\u4f7f\u5f97\u7f51\u7edc\u66f4\u5bb9\u6613\u8bad\u7ec3\u5e76\u5f97\u5230\u66f4\u597d\u7684\u7ed3\u679c\uff0c\u73b0\u6709\u5b9e\u9a8c\u7ed3\u679c\u4e5f\u540c\u6837\u9a8c\u8bc1\u4e86\u6b64\u9879\u7ed3\u8bba\u3002\u6700\u7ec8\u8fd9\u79cd\u7a00\u758f\u5316\u4e0e\u52a8\u6001\u6027\u4e0a\u7684\u6f14\u8fdb\uff0c\u4f1a\u8d70\u5411\u57fa\u4e8eDepth-wise\u5377\u79ef\u7684\u52a8\u6001\u5377\u79ef\uff0c\u7ed3\u5408\u73b0\u4ee3Large kernel\u7684\u8bad\u7ec3\u539f\u5219\uff0c\u5c06\u80fd\u591f\u5b9e\u73b0\u66f4\u597d\u7684\u8868\u73b0\u6548\u679c\u3002<\/p>\n\n\n\n

[1] Liu Z, Mao H, Wu C Y, et al. A ConvNet for the 2020s. arXiv preprint arXiv:2201.03545, 2022.<\/p>\n\n\n\n

[2] Ding X, Zhang X, Zhou Y, et al. Scaling Up Your Kernels to 31×31: Revisiting Large Kernel Design in CNNs. CVPR, 2022.<\/p>\n","protected":false},"excerpt":{"rendered":"

\u8fd1\u4e24\u5e74\u6765\uff0c\u6709\u5173Transformer\u7684\u6587\u7ae0\u53ef\u8c13\u662f\u4e95\u55b7\u5f0f\u7206\u53d1\uff0c\u8bb8\u591a\u7814\u7a76\u5458\u90fd\u5728\u8bbe\u8ba1\u5404\u79cd\u4efb\u52a1\u4e0a\u7684Transformer\u6a21\u578b\u3002\u7136\u800c\uff0cAttention\u4f5c\u4e3aTransformer\u7684\u6838\u5fc3\u6a21\u5757\uff0c\u771f\u7684\u6bd4\u5377\u79ef\u5f3a\u5417\uff1f\u4eca\u5929\u7684\u8fd9\u7bc7\u6587\u7ae0\u6216\u8bb8\u80fd\u5e26\u7ed9\u4f60\u65b0\u89c6\u89d2\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u4eceLocal Attention\u548cDynamic Depth-wise Convolution\u7684\u89c6\u89d2\u51fa\u53d1\u53d1\u73b0\uff0c\u8bbe\u8ba1\u597d\u7684\u5377\u79ef\u7ed3\u6784\u5e76\u4e0d\u6bd4Transformer\u5dee\uff01\u76f8\u5173\u8bba\u6587\u201cOn the Connection between Local Attention and Dynamic Depth-wise Convolution\u201d\u5df2\u88abICLR 2022\u6536\u5f55\u3002 \u8bba\u6587\u94fe\u63a5: https:\/\/arxiv.org\/abs\/2106.04263 \u4ee3\u7801\u94fe\u63a5: https:\/\/github.com\/Atten4Vis\/DemystifyLocalViT 2020\u5e74\uff0cViT\u6a2a\u7a7a\u51fa\u4e16\uff0c\u5e2d\u5377\u4e86\u6a21\u578b\u8bbe\u8ba1\u9886\u57df\u3002\u5404\u79cd\u57fa\u4e8eTransformer\u7684\u7ed3\u6784\u5f00\u59cb\u88ab\u63d0\u51fa\uff0c\u4e00\u4e9b\u5728\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u4e2d\u53d6\u5f97\u6210\u529f\u7684\u5148\u9a8c\u77e5\u8bc6\uff0c\u5982Local Operation\u3001\u591a\u5c3a\u5ea6\u3001Shuffle\u7b49\u5404\u79cd\u64cd\u4f5c\u548cinductive bias\u88ab\u5f15\u5165Transformer\u4e4b\u4e2d\u3002\u5176\u4e2d\u6bd4\u8f83\u6210\u529f\u7684\u6848\u4f8b\u4fbf\u662f\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684Swin Transformer\uff0c\u901a\u8fc7\u5c06Local Operation\u5f15\u5165ViT\u5f53\u4e2d\uff0c\u5229\u7528shifted window\u4e00\u4e3e\u83b7\u5f97\u591a\u4e2a\u4efb\u52a1\u7684SOTA\u7ed3\u679c\uff0c\u5e76\u83b7\u5f97\u4e862021\u5e74 ICCV \u6700\u4f73\u8bba\u6587\u5956\u2014\u2014\u9a6c\u5c14\u5956\u3002\u90a3\u4e48Swin Transformer\u4e2d\u6700\u6838\u5fc3\u7684\u6a21\u5757\uff0cLocal Attention\u7a76\u7adf\u6709\u4ec0\u4e48\u795e\u79d8\u4e4b\u5904\u5462? Local Attention\u672c\u8d28\u4e0a\u662f\u5728\u4e00\u4e2a2D Local Window\u5185\u8fdb\u884c\u7279\u5f81\u805a\u5408\uff0c\u5176\u6bcf\u4e2a\u4f4d\u7f6e\u7684\u805a\u5408\u6743\u91cd\u53ef\u4ee5\u901a\u8fc7KQV\u4e4b\u95f4\u8ba1\u7b97Attention Similarity\u5f97\u5230\uff08\u4e3b\u8981\u5305\u62ecdot-production, scaling, softmax\uff09\uff0c\u662f\u4e00\u4e2a\u65e0\u53c2\u6570\u7684\u3001\u52a8\u6001\u8ba1\u7b97\u7684\u5c40\u90e8\u7279\u5f81\u8ba1\u7b97\u6a21\u5757\u3002 \u6b64\u6b21ICLR\u8fd9\u7bc7\u6587\u7ae0\u7684\u5148\u524d\u7248\u672c \uff08Demystifying Local Vision Transformer\uff09\u57282021\u5e746\u6708\u5c31\u9996\u6b21\u5728arXiv\u516c\u5e03\u5e76\u5206\u6790\u4e86Local Attention\u4e09\u4e2a\u5f3a\u5927\u7684\u8bbe\u8ba1\u539f\u5219\uff1a (1) \u7a00\u758f\u8fde\u63a5\u2014\u2014\u6307\u4e00\u4e9b\u8f93\u51fa\u53d8\u91cf\u548c\u4e00\u4e9b\u8f93\u5165\u53d8\u91cf\u4e4b\u95f4\u6ca1\u6709\u76f8\u4e92\u8fde\u63a5\u3002\u5b83\u6709\u6548\u5730\u51cf\u5c11\u4e86\u6a21\u578b\u7684\u590d\u6742\u5ea6\u800c\u4e0d\u51cf\u5c11\u8f93\u5165\u8f93\u51fa\u53d8\u91cf\u4e2a\u6570\u3002\u5728Local Attention\u5f53\u4e2d\uff0c\u7a00\u758f\u8fde\u63a5\u4f53\u73b0\u5728\u4e24\u4e2a\u65b9\u9762\uff1a\u4e00\u662fLocal Attention\u5728\u56fe\u50cf\u7a7a\u95f4\u4e0a\uff0c\u6bcf\u4e00\u4e2a\u8f93\u51fa\u503c\u4ec5\u4e0e\u5c40\u90e8\u7684Local Window\u5185\u7684\u8f93\u5165\u76f8\u8fde\u63a5\uff0c\u4e0eViT\u7684\u5168\u50cf\u7d20\uff08token\uff09\u8fde\u63a5\u4e0d\u540c\u3002\u4e8c\u662fLocal Attention\u5728\u901a\u9053\u4e0a\uff0c\u6bcf\u4e00\u4e2a\u8f93\u51fa\u901a\u9053\u4ec5\u4e0e\u4e00\u4e2a\u8f93\u5165\u901a\u9053\u8fde\u63a5\uff0c\u6ca1\u6709\u4ea4\u53c9\u8fde\u63a5\uff0c\u4e0d\u540c\u4e8egroup convolution\u4e0enormal convolution\u3002 (2) \u6743\u91cd\u5171\u4eab\u2014\u2014\u610f\u5473\u7740\u6709\u4e00\u4e9b\u8fde\u63a5\u7684\u6743\u91cd\u662f\u76f8\u540c\u4e14\u5171\u4eab\u7684\uff0c\u5b83\u964d\u4f4e\u4e86\u6a21\u578b\u7684\u53c2\u6570\u91cf\uff0c\u540c\u65f6\u4e0d\u9700\u8981\u589e\u52a0\u8bad\u7ec3\u6570\u636e\u5373\u53ef\u589e\u5f3a\u6a21\u578b\u3002\u5728\u6a21\u578b\u4e2d\uff0c\u4e00\u4e2a\u6743\u91cd\u88ab\u5171\u4eab\u4f7f\u7528\u53ef\u4ee5\u88ab\u8ba4\u4e3a\u9488\u5bf9\u8be5\u6743\u91cd\u7684\u8bad\u7ec3\u6837\u672c\u589e\u52a0\uff0c\u6709\u52a9\u4e8e\u6a21\u578b\u7684\u4f18\u5316\u3002\u5728Local Attention\u4e2d\uff0c\u6743\u91cd\u5171\u4eab\u901a\u8fc7multi-head self-attention\u6765\u5b9e\u73b0\uff0c\u518d\u628a\u901a\u9053\u5206\u6210head\uff08group\uff09\uff0c\u5728\u540c\u4e00\u4e2ahead\u5185\u5171\u4eab\u4f7f\u7528\u4e00\u7ec4\u805a\u5408\u6743\u91cd\uff0c\u964d\u4f4e\u805a\u5408\u6743\u91cd\u7684\u53c2\u6570\u91cf\uff08\u975e\u6a21\u578b\u53c2\u6570\u91cf\uff09\u3002 (3) \u52a8\u6001\u6743\u91cd\u2014\u2014\u6307\u6839\u636e\u4e0d\u540c\u6837\u672c\u7684\u7279\u5f81\uff0c\u52a8\u6001\u5730\u4ea7\u751f\u8fde\u63a5\u6743\u91cd\u3002\u5b83\u80fd\u591f\u589e\u52a0\u6a21\u578b\u7684\u5bb9\u91cf\u3002\u5982\u679c\u628a\u8fde\u63a5\u6743\u91cd\u770b\u4f5c\u662f\u9690\u5c42\u53d8\u91cf\uff0c\u90a3\u4e48\u8fd9\u79cd\u52a8\u6001\u6743\u91cd\u53ef\u4ee5\u770b\u4f5c\u662f\u589e\u52a0\u6a21\u578b\u5bb9\u91cf\u7684\u4e8c\u9636\u64cd\u4f5c\u3002Local Attention\u7684\u52a8\u6001\u6743\u91cd\u4f53\u73b0\u5728\u6bcf\u4e00\u4e2a\u8fde\u63a5\u7684\u805a\u5408\u6743\u91cd\u90fd\u662f\u6839\u636e\u6837\u672c\u7279\u5f81\u4f7f\u7528\u57fa\u4e8edot-product\u7684\u65b9\u5f0f\u8ba1\u7b97\u5f97\u5230\u7684\u3002 \u901a\u8fc7\u4ee5\u4e0a\u4e09\u4e2a\u6a21\u578b\u8bbe\u8ba1\u539f\u5219\uff0cLocal Attention\u8868\u73b0\u51fa\u4f18\u5f02\u7684\u6548\u679c\u3002\u7136\u800c\uff0c\u8fd9\u4e9b\u7279\u6027\u4e5f\u5929\u7136\u5b58\u5728\u4e8eCNN\u7ed3\u6784\u5f53\u4e2d\uff0c\u5c24\u5176\u662f\uff08Dynamic\uff09Depth-wise\u5377\u79ef\u3002 \u9010\u6b65\u62c6\u89e3Local Attention\u7684\u64cd\u4f5c\uff0c\u53ef\u4ee5\u53d1\u73b0\u5728\u7a00\u758f\u8fde\u63a5\u3001\u6743\u91cd\u5171\u4eab\u3001\u52a8\u6001\u6743\u91cd\u4e09\u4e2a\u7ef4\u5ea6\u4e0a\uff0c\u90fd\u4e0e\u201c\u5386\u53f2\u9738\u4e3b\u201dCNN\u7ed3\u6784\u4e2d\u7684\uff08Dynamic\uff09Depth-wise\u5377\u79ef\u5f88\u76f8\u4f3c\u3002Depth-wise\u5377\u79ef\u53ef\u8c13\u662f\u4e00\u4e2a\u88ab\u957f\u671f\u4f7f\u7528\u53c8\u88ab\u6e10\u6e10\u9057\u5fd8\u5728\u5386\u53f2\u957f\u6cb3\u4e2d\u7684\u7ed3\u6784\uff0c\u90a3\u4e48\u5176\u5728\u6a21\u578b\u8bbe\u8ba1\u4e0a\u53c8\u6709\u54ea\u4e9b\u51c6\u5219\u5462\uff1f […]<\/p>\n","protected":false},"author":34512,"featured_media":0,"template":"","meta":{"msr-url-field":"","msr-podcast-episode":"","msrModifiedDate":"","msrModifiedDateEnabled":false,"ep_exclude_from_search":false,"_classifai_error":"","msr-content-parent":1012650,"msr_hide_image_in_river":0,"footnotes":""},"research-area":[],"msr-locale":[268881],"msr-post-option":[],"class_list":["post-1069491","msr-blog-post","type-msr-blog-post","status-publish","hentry","msr-locale-zh_cn"],"msr_assoc_parent":{"id":1012650,"type":"lab"},"_links":{"self":[{"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1069491","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post"}],"about":[{"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/types\/msr-blog-post"}],"author":[{"embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/users\/34512"}],"version-history":[{"count":3,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1069491\/revisions"}],"predecessor-version":[{"id":1070991,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1069491\/revisions\/1070991"}],"wp:attachment":[{"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/media?parent=1069491"}],"wp:term":[{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=1069491"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=1069491"},{"taxonomy":"msr-post-option","embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-post-option?post=1069491"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}