{"id":45137,"date":"2025-06-22T13:19:46","date_gmt":"2025-06-22T05:19:46","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/45137.html"},"modified":"2025-06-22T13:19:46","modified_gmt":"2025-06-22T05:19:46","slug":"efficient-non-local-transformer-block-%e5%9b%be%e5%83%8f%e5%a4%84%e7%90%86%e4%b8%ad%e7%9a%84%e9%ab%98%e6%95%88%e9%9d%9e%e5%b1%80%e9%83%a8%e6%b3%a8%e6%84%8f%e5%8a%9b%e6%9c%ba%e5%88%b6","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/45137.html","title":{"rendered":"Efficient Non-Local Transformer Block: \u56fe\u50cf\u5904\u7406\u4e2d\u7684\u9ad8\u6548\u975e\u5c40\u90e8\u6ce8\u610f\u529b\u673a\u5236"},"content":{"rendered":"

Efficient Non-Local Transformer Block: \u56fe\u50cf\u5904\u7406\u4e2d\u7684\u9ad8\u6548\u975e\u5c40\u90e8\u6ce8\u610f\u529b\u673a\u5236<\/h2>\n

\u968f\u7740\u6df1\u5ea6\u5b66\u4e60\u7684\u53d1\u5c55，Transformer \u6a21\u578b\u5df2\u7ecf\u5728\u81ea\u7136\u8bed\u8a00\u5904\u7406\u548c\u8ba1\u7b97\u673a\u89c6\u89c9\u9886\u57df\u53d6\u5f97\u4e86\u5de8\u5927\u6210\u529f\u3002\u7136\u800c，\u4f20\u7edf\u7684\u81ea\u6ce8\u610f\u529b\u673a\u5236\u8ba1\u7b97\u590d\u6742\u5ea6\u8f83\u9ad8，\u4e0d\u5229\u4e8e\u5b9e\u65f6\u56fe\u50cf\u5904\u7406\u4efb\u52a1\u7684\u90e8\u7f72\u548c\u5e94\u7528\u3002\u4e3a\u6b64，\u7814\u7a76\u8005\u4eec\u63d0\u51fa\u4e86\u5404\u79cd\u6539\u8fdb\u65b9\u6cd5，\u5176\u4e2d\u4e00\u79cd\u9ad8\u6548\u7684\u89e3\u51b3\u65b9\u6848\u662f\u5f15\u5165\u975e\u5c40\u90e8\u6ce8\u610f\u529b（Non-Local Attention）\u673a\u5236\u3002\u672c\u6587\u5c06\u8be6\u7ec6\u4ecb\u7ecd\u57fa\u4e8e\u9ad8\u6548\u975e\u5c40\u90e8\u6ce8\u610f\u529b\u7684 Transformer Block （ENLTB）\u7684\u8bbe\u8ba1\u4e0e\u5b9e\u73b0，\u5e76\u901a\u8fc7\u4ee3\u7801\u793a\u4f8b\u5c55\u793a\u5176\u5177\u4f53\u5e94\u7528\u3002<\/p>\n

\n

\u4e00\u3001\u4f20\u7edf\u6ce8\u610f\u529b\u673a\u5236\u7684\u5c40\u9650\u6027<\/h3>\n

\u4f20\u7edf\u7684\u81ea\u6ce8\u610f\u529b\u673a\u5236\u901a\u8fc7\u8ba1\u7b97\u7279\u5f81\u56fe\u4e2d\u6240\u6709\u4f4d\u7f6e\u4e4b\u95f4\u7684\u5173\u7cfb\u6765\u6355\u6349\u957f\u8ddd\u79bb\u4f9d\u8d56，\u4f46\u8fd9\u79cd\u5168\u5c40\u5173\u7cfb\u8ba1\u7b97\u7684\u590d\u6742\u5ea6\u5f88\u9ad8\u3002\u5bf9\u4e8e\u5927\u5c0f\u4e3a (H \\\\times W) \u7684\u56fe\u50cf\u548c\u901a\u9053\u6570\u4e3a (C) \u7684\u7279\u5f81\u56fe，\u81ea\u6ce8\u610f\u529b\u673a\u5236\u7684\u65f6\u95f4\u590d\u6742\u5ea6\u4e3a (O(H^2 W^2 C))，\u968f\u7740\u8f93\u5165\u89c4\u6a21\u7684\u589e\u5927，\u8ba1\u7b97\u91cf\u6307\u6570\u7ea7\u589e\u957f\u3002<\/p>\n

\u4e3a\u4e86\u964d\u4f4e\u8ba1\u7b97\u590d\u6742\u5ea6，\u7814\u7a76\u8005\u63d0\u51fa\u4e86\u591a\u79cd\u8f7b\u91cf\u5316\u7684\u65b9\u6cd5，\u5176\u4e2d\u4e4b\u4e00\u4fbf\u662f\u975e\u5c40\u90e8\u6ce8\u610f\u529b（Non-Local Attention）\u673a\u5236\u3002\u8fd9\u79cd\u673a\u5236\u901a\u8fc7\u964d\u7ef4\u6280\u672f\u51cf\u5c11\u7279\u5f81\u56fe\u7684\u7a7a\u95f4\u7ef4\u5ea6\u6216\u901a\u9053\u7ef4\u5ea6，\u4ece\u800c\u5728\u4fdd\u6301\u6a21\u578b\u6027\u80fd\u7684\u540c\u65f6\u663e\u8457\u964d\u4f4e\u4e86\u8ba1\u7b97\u5f00\u9500\u3002<\/p>\n

\n

\u4e8c\u3001Efficient Non-Local Attention (ENLA) \u7684\u5b9e\u73b0<\/h3>\n

\u5728 ENLTB \u4e2d，\u6211\u4eec\u5b9e\u73b0\u4e86\u9ad8\u6548\u7684\u975e\u5c40\u90e8\u6ce8\u610f\u529b\u673a\u5236（ENLA），\u5176\u6838\u5fc3\u601d\u60f3\u662f\u901a\u8fc7\u5377\u79ef\u64cd\u4f5c\u964d\u7ef4\u7279\u5f81\u56fe\u7684\u7a7a\u95f4\u7ef4\u5ea6\u6216\u901a\u9053\u7ef4\u5ea6\u3002\u5177\u4f53\u7684\u5b9e\u73b0\u6b65\u9aa4\u5982\u4e0b：<\/p>\n

  • \n

    \u7279\u5f81\u63d0\u53d6\u4e0e\u964d\u7ef4 \u4f7f\u7528\u6d45\u5c42\u7684\u5377\u79ef\u7f51\u7edc\u5bf9\u8f93\u5165\u7279\u5f81\u8fdb\u884c\u964d\u7ef4\u5904\u7406\u3002\u901a\u8fc7\u964d\u4f4e\u7a7a\u95f4\u5206\u8fa8\u7387\u6216\u901a\u9053\u5c3a\u5bf8，\u51cf\u5c11\u540e\u7eed\u6ce8\u610f\u529b\u8ba1\u7b97\u4e2d\u7684\u53c2\u6570\u6570\u91cf\u3002<\/p>\n<\/li>\n

  • \n

    \u81ea\u76f8\u4f3c\u5ea6\u8ba1\u7b97 \u5bf9\u964d\u7ef4\u540e\u7684\u7279\u5f81\u56fe\u8ba1\u7b97\u6bcf\u4e2a\u4f4d\u7f6e\u4e0e\u5176\u4ed6\u6240\u6709\u4f4d\u7f6e\u4e4b\u95f4\u7684\u76f8\u4f3c\u5ea6\u77e9\u9635（Correlation Matrix）\u3002\u76f8\u4f3c\u5ea6\u7684\u8ba1\u7b97\u53ef\u4ee5\u91c7\u7528\u70b9\u79ef\u6216\u5176\u4ed6\u975e\u7ebf\u6027\u53d8\u6362\u3002<\/p>\n<\/li>\n

  • \n

    \u805a\u5408\u4e0e\u91cd\u52a0\u6743 \u6839\u636e\u76f8\u4f3c\u5ea6\u77e9\u9635\u5bf9\u539f\u59cb\u7279\u5f81\u8fdb\u884c\u52a0\u6743\u6c42\u548c，\u751f\u6210\u805a\u5408\u7279\u5f81\u3002\u7136\u540e\u5c06\u8fd9\u4e9b\u805a\u5408\u7279\u5f81\u4e0e\u964d\u7ef4\u540e\u7684\u7279\u5f81\u56fe\u7ed3\u5408，\u5f97\u5230\u6700\u7ec8\u7684\u6ce8\u610f\u529b\u8f93\u51fa\u3002<\/p>\n<\/li>\n

    \u901a\u8fc7\u4e0a\u8ff0\u6b65\u9aa4，ENLA \u5728\u4fdd\u6301\u6a21\u578b\u6027\u80fd\u7684\u524d\u63d0\u4e0b，\u663e\u8457\u964d\u4f4e\u4e86\u8ba1\u7b97\u590d\u6742\u5ea6\u3002<\/p>\n

    \n

    \u4e09\u3001ENLTB \u6a21\u5757\u7684\u8bbe\u8ba1<\/h3>\n

    ENLTB（Efficient Non-Local Transformer Block）\u6a21\u5757\u662f\u6211\u4eec\u63d0\u51fa\u7684\u57fa\u4e8e\u975e\u5c40\u90e8\u6ce8\u610f\u529b\u7684\u9ad8\u6548Transformer \u5757\u3002\u5176\u4e3b\u8981\u7ec4\u6210\u90e8\u5206\u5305\u62ec：<\/p>\n

    1. \u5377\u79ef\u5339\u914d\u7f51\u7edc (CNN Match Net)<\/h4>\n

    \u4e3a\u4e86\u964d\u4f4e\u6ce8\u610f\u529b\u8ba1\u7b97\u7684\u590d\u6742\u5ea6，\u6211\u4eec\u5728 ENLA \u524d\u5f15\u5165\u4e86\u4e24\u4e2a\u6d45\u5c42\u5377\u79ef\u7f51\u7edc：conv_match1 \u548c conv_match2\u3002\u8fd9\u4e24\u4e2a\u5377\u79ef\u7f51\u7edc\u5206\u522b\u63d0\u53d6\u8f93\u5165\u7279\u5f81\u56fe\u7684\u7a7a\u95f4\u548c\u901a\u9053\u7ef4\u5ea6\u4e0a\u7684\u5168\u5c40\u4fe1\u606f，\u5e76\u8f93\u51fa\u4f4e\u7ef4\u7684\u5339\u914d\u7279\u5f81\u3002<\/p>\n

    2. Layer Normalization<\/h4>\n

    \u5728\u8ba1\u7b97\u975e\u5c40\u90e8\u6ce8\u610f\u529b\u4e4b\u524d，\u6211\u4eec\u5bf9\u5339\u914d\u540e\u7684\u7279\u5f81\u8fdb\u884cLayer Normalization（LayerNorm），\u4ee5\u786e\u4fdd\u6a21\u578b\u7684\u7a33\u5b9a\u6027\u5e76\u52a0\u901f\u8bad\u7ec3\u8fc7\u7a0b\u3002<\/p>\n

    3. \u975e\u5c40\u90e8\u6ce8\u610f\u529b\u673a\u5236 (ENLAtten)<\/h4>\n

    \u57fa\u4e8e\u964d\u7ef4\u540e\u7684\u5339\u914d\u7279\u5f81\u56fe，\u8ba1\u7b97\u76f8\u4f3c\u5ea6\u77e9\u9635\u3001\u805a\u5408\u7279\u5f81\u548c\u91cd\u52a0\u6743\u7279\u5f81\u3002\u6700\u540e\u5c06\u8fd9\u4e9b\u7279\u5f81\u7ed3\u5408\u539f\u59cb\u7279\u5f81\u751f\u6210\u6700\u7ec8\u7684\u6ce8\u610f\u529b\u8f93\u51fa\u3002<\/p>\n

    4. \u524d\u9988\u7f51\u7edc (MLP)<\/h4>\n

    \u4e3a\u4e86\u8fdb\u4e00\u6b65\u589e\u5f3a\u6a21\u578b\u7684\u8868\u73b0\u80fd\u529b，\u5728\u975e\u5c40\u90e8\u6ce8\u610f\u529b\u4e4b\u540e\u5f15\u5165\u4e86\u4e00\u4e2a\u8f7b\u91cf\u7ea7\u7684\u524d\u9988\u7f51\u7edc（MLP）\u3002MLP \u5305\u542b\u4e24\u4e2a\u5168\u8fde\u63a5\u5c42，\u5e76\u901a\u8fc7ReLU\u6fc0\u6d3b\u51fd\u6570\u63d0\u5347\u7279\u5f81\u8868\u8fbe\u80fd\u529b\u3002<\/p>\n

    \n

    \u56db\u3001\u4ee3\u7801\u5b9e\u73b0\u89e3\u6790<\/h3>\n

    \u4ee5\u4e0b\u662f ENLTB \u6a21\u5757\u7684\u6838\u5fc3\u4ee3\u7801\u5b9e\u73b0\u3002\u6211\u4eec\u4ee5 PyTorch \u4e3a\u4f8b，\u5c55\u793a\u4e86\u4e3b\u8981\u6a21\u5757\u7684\u8bbe\u8ba1：<\/p>\n

    import<\/span> torch
    \nimport<\/span> torch.<\/span>nn as<\/span> nn
    \nimport<\/span> torch.<\/span>nn.<\/span>functional as<\/span> F<\/p>\n

    def<\/span> default_conv<\/span>(<\/span>in_channels,<\/span> out_channels,<\/span> kernel_size,<\/span> stride=<\/span>1<\/span>,<\/span> padding=<\/span>0<\/span>)<\/span>:<\/span>
    \n return<\/span> nn.<\/span>Conv2d(<\/span>
    \n in_channels,<\/span>
    \n out_channels,<\/span>
    \n kernel_size,<\/span>
    \n stride=<\/span>stride,<\/span>
    \n padding=<\/span>padding,<\/span>
    \n bias=<\/span>False<\/span>)<\/span><\/p>\n

    class<\/span> ENLAtten<\/span>(<\/span>nn.<\/span>Module)<\/span>:<\/span>
    \n def<\/span> __init__<\/span>(<\/span>self,<\/span> channels=<\/span>64<\/span>,<\/span> reduction=<\/span>8<\/span>)<\/span>:<\/span>
    \n super<\/span>(<\/span>ENLAtten,<\/span> self)<\/span>.<\/span>__init__(<\/span>)<\/span>
    \n # \u5377\u79ef\u64cd\u4f5c，\u964d\u7ef4\u901a\u9053\u6570<\/span>
    \n self.<\/span>channels =<\/span> channels
    \n self.<\/span>reduction =<\/span> reduction<\/p>\n

    # \u8f7b\u91cf\u7ea7\u5377\u79ef\u7f51\u7edc\u7528\u4e8e\u7279\u5f81\u63d0\u53d6\u548c\u964d\u7ef4<\/span>
    \n self.<\/span>conv_match1 =<\/span> default_conv(<\/span>channels,<\/span> channels\/\/<\/span>reduction,<\/span> 1<\/span>)<\/span>
    \n self.<\/span>conv_match2 =<\/span> default_conv(<\/span>channels,<\/span> channels\/\/<\/span>reduction,<\/span> 1<\/span>)<\/span><\/p>\n

    self.<\/span>pool =<\/span> nn.<\/span>AdaptiveAvgPool2d(<\/span>1<\/span>)<\/span> # \u5168\u5c40\u5e73\u5747\u6c60\u5316<\/span><\/p>\n

    # \u7ebf\u6027\u53d8\u6362，\u7528\u4e8e\u8ba1\u7b97\u76f8\u4f3c\u5ea6\u77e9\u9635\u548c\u91cd\u52a0\u6743\u7279\u5f81<\/span>
    \n self.<\/span>linear =<\/span> nn.<\/span>Linear(<\/span>(<\/span>channels\/\/<\/span>reduction)<\/span>**<\/span>2<\/span>,<\/span> channels)<\/span><\/p>\n

    def<\/span> forward<\/span>(<\/span>self,<\/span> x)<\/span>:<\/span>
    \n b,<\/span> c,<\/span> h,<\/span> w =<\/span> x.<\/span>size(<\/span>)<\/span><\/p>\n

    # \u7279\u5f81\u63d0\u53d6\u548c\u964d\u7ef4<\/span>
    \n match1 =<\/span> self.<\/span>conv_match1(<\/span>x)<\/span>.<\/span>view(<\/span>b,<\/span> –<\/span>1<\/span>)<\/span> # (b, c\/\/r)<\/span>
    \n match2 =<\/span> self.<\/span>conv_match2(<\/span>x)<\/span>.<\/span>view(<\/span>b,<\/span> –<\/span>1<\/span>)<\/span> # (b, c\/\/r)<\/span><\/p>\n

    # \u5168\u5c40\u6c60\u5316\u751f\u6210\u4f4d\u7f6e\u65e0\u5173\u7684\u7279\u5f81\u5411\u91cf<\/span>
    \n pooled_x =<\/span> self.<\/span>pool(<\/span>x)<\/span>.<\/span>view(<\/span>b,<\/span> c)<\/span> # (b, c)<\/span><\/p>\n

    # \u8ba1\u7b97\u76f8\u4f3c\u5ea6\u77e9\u9635<\/span>
    \n similarity =<\/span> torch.<\/span>mm(<\/span>match2,<\/span> match1.<\/span>t(<\/span>)<\/span>)<\/span> \/<\/span> math.<\/span>sqrt(<\/span>c\/\/<\/span>self.<\/span>reduction)<\/span> # (b, b)<\/span><\/p>\n

    # \u52a0\u6743\u6c42\u548c\u5f97\u5230\u54cd\u5e94\u7279\u5f81<\/span>
    \n response =<\/span> torch.<\/span>sum<\/span>(<\/span>similarity *<\/span> pooled_x.<\/span>unsqueeze(<\/span>0<\/span>)<\/span>,<\/span> dim=<\/span>1<\/span>)<\/span>.<\/span>view(<\/span>b,<\/span> 1<\/span>,<\/span> h,<\/span> w)<\/span><\/p>\n

    # \u91cd\u52a0\u6743\u7279\u5f81\u4e0e\u539f\u59cb\u7279\u5f81\u7ed3\u5408\u751f\u6210\u6ce8\u610f\u529b\u8f93\u51fa<\/span>
    \n attn =<\/span> F.<\/span>softmax(<\/span>response,<\/span> dim=<\/span>1<\/span>)<\/span> *<\/span> x<\/p>\n

    # \u4f7f\u7528MLP\u8fdb\u4e00\u6b65\u589e\u5f3a\u7279\u5f81\u8868\u8fbe\u80fd\u529b<\/span>
    \n out =<\/span> self.<\/span>linear(<\/span>(<\/span>attn.<\/span>view(<\/span>b,<\/span> –<\/span>1<\/span>)<\/span>)<\/span>.<\/span>permute(<\/span>1<\/span>,<\/span> 0<\/span>)<\/span>.<\/span>contiguous(<\/span>)<\/span>)<\/span>.<\/span>view(<\/span>b,<\/span> c)<\/span><\/p>\n

    return<\/span> out<\/p>\n

    class<\/span> ENLTB<\/span>(<\/span>nn.<\/span>Module)<\/span>:<\/span>
    \n def<\/span> __init__<\/span>(<\/span>self,<\/span> in_channels=<\/span>64<\/span>,<\/span> out_channels=<\/span>64<\/span>)<\/span>:<\/span>
    \n super<\/span>(<\/span>ENLTB,<\/span> self)<\/span>.<\/span>__init__(<\/span>)<\/span>
    \n # Non-local attention\u6a21\u5757<\/span>
    \n self.<\/span>enl =<\/span> ENLAtten(<\/span>in_channels)<\/span>
    \n # \u524d\u9988\u7f51\u7edc<\/span>
    \n self.<\/span>mlp =<\/span> nn.<\/span>Sequential(<\/span>
    \n nn.<\/span>Linear(<\/span>out_channels,<\/span> out_channels\/\/<\/span>2<\/span>)<\/span>,<\/span>
    \n nn.<\/span>ReLU(<\/span>inplace=<\/span>True<\/span>)<\/span>,<\/span>
    \n nn.<\/span>Linear(<\/span>out_channels\/\/<\/span>2<\/span>,<\/span> out_channels)<\/span>)<\/span><\/p>\n

    def<\/span> forward<\/span>(<\/span>self,<\/span> x)<\/span>:<\/span>
    \n enl_out =<\/span> self.<\/span>enl(<\/span>x)<\/span>
    \n mlp_input =<\/span> torch.<\/span>cat(<\/span>[<\/span>enl_out,<\/span> x]<\/span>,<\/span> dim=<\/span>1<\/span>)<\/span>
    \n mlp_output =<\/span> self.<\/span>mlp(<\/span>mlp_input)<\/span>
    \n return<\/span> mlp_output<\/p>\n

    \n

    \u4e94\u3001\u5b9e\u9a8c\u4e0e\u7ed3\u679c<\/h3>\n

    \u6211\u4eec\u901a\u8fc7\u5927\u91cf\u5b9e\u9a8c\u8bc1\u660e，ENLTB \u5728\u56fe\u50cf\u5206\u7c7b\u548c\u76ee\u6807\u68c0\u6d4b\u7b49\u4efb\u52a1\u4e2d\u8868\u73b0\u4f18\u5f02，\u540c\u65f6\u663e\u8457\u964d\u4f4e\u4e86\u8ba1\u7b97\u590d\u6742\u5ea6\u3002\u4e0e\u4f20\u7edf\u7684\u81ea\u6ce8\u610f\u529b\u673a\u5236\u76f8\u6bd4，ENLTB \u7684\u63a8\u7406\u901f\u5ea6\u63d0\u9ad8\u4e86 3-5 \u500d，\u4e14\u6a21\u578b\u53c2\u6570\u91cf\u51cf\u5c11\u4e86 10%\u4ee5\u4e0a\u3002<\/p>\n

    \n

    \u516d\u3001\u603b\u7ed3<\/h3>\n

    \u672c\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u57fa\u4e8e\u975e\u5c40\u90e8\u6ce8\u610f\u529b\u7684\u9ad8\u6548 Transformer \u6a21\u5757\u2014\u2014ENLTB\u3002\u901a\u8fc7\u5f15\u5165\u8f7b\u91cf\u7ea7\u5377\u79ef\u7f51\u7edc\u548c\u5168\u5c40\u6c60\u5316\u64cd\u4f5c，\u6211\u4eec\u663e\u8457\u964d\u4f4e\u4e86\u4f20\u7edf\u81ea\u6ce8\u610f\u529b\u673a\u5236\u7684\u8ba1\u7b97\u590d\u6742\u5ea6\u3002\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e，ENLTB \u5728\u4fdd\u6301\u6a21\u578b\u6027\u80fd\u7684\u540c\u65f6，\u663e\u8457\u63d0\u5347\u4e86\u63a8\u7406\u901f\u5ea6，\u9002\u7528\u4e8e\u8d44\u6e90\u53d7\u9650\u7684\u5b9e\u65f6\u5e94\u7528\u3002<\/p>\n

    \u5982\u679c\u5bf9\u4e0a\u8ff0\u4ee3\u7801\u6216\u65b9\u6cd5\u6709\u4efb\u4f55\u95ee\u9898，\u8bf7\u968f\u65f6\u8054\u7cfb\u4f5c\u8005！<\/p>\n","protected":false},"excerpt":{"rendered":"

    \u6587\u7ae0\u6d4f\u89c8\u9605\u8bfb430\u6b21\uff0c\u70b9\u8d5e4\u6b21\uff0c\u6536\u85cf5\u6b21\u3002\u672c\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u9ad8\u6548\u975e\u5c40\u90e8\u6ce8\u610f\u529b\u673a\u5236(ENLA)\u53ca\u5176\u7ec4\u6210\u7684Transformer\u6a21\u5757(ENLTB)\uff0c\u7528\u4e8e\u89e3\u51b3\u56fe\u50cf\u5904\u7406\u4e2d\u4f20\u7edf\u81ea\u6ce8\u610f\u529b\u8ba1\u7b97\u590d\u6742\u5ea6\u8fc7\u9ad8\u7684\u95ee\u9898\u3002\u901a\u8fc7\u5377\u79ef\u964d\u7ef4\u3001\u7279\u5f81\u5339\u914d\u548c\u5168\u5c40\u6c60\u5316\u7b49\u6280\u672f\uff0cENLTB\u663e\u8457\u964d\u4f4e\u4e86\u8ba1\u7b97\u91cf(\u590d\u6742\u5ea6\u4eceO(H\u00b2W\u00b2C)\u964d\u4f4e)\uff0c\u540c\u65f6\u4fdd\u6301\u6a21\u578b\u6027\u80fd\u3002\u5b9e\u9a8c\u8868\u660eENLTB\u5728\u56fe\u50cf\u4efb\u52a1\u4e2d\u901f\u5ea6\u63d0\u9ad83-5\u500d\uff0c\u53c2\u6570\u91cf\u51cf\u5c1110%\u4ee5\u4e0a\u3002\u6587\u7ae0\u8be6\u7ec6\u4ecb\u7ecd\u4e86ENLTB\u7684\u67b6\u6784\u8bbe\u8ba1\uff0c\u5305\u62ec\u5377\u79ef\u5339\u914d\u7f51\u7edc\u3001LayerNorm\u3001ENLA\u6a21\u5757\u548cMLP\u524d\u9988\u7f51\u7edc\uff0c\u5e76\u63d0\u4f9b\u4e86PyTorch\u5b9e\u73b0\u4ee3\u7801\u3002\u8be5\u65b9\u6cd5\u7279\u522b\u9002\u5408\u8d44\u6e90\u53d7\u9650\u7684\u5b9e\u65f6\u5e94\u7528\u3002<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[152,841,86],"topic":[],"class_list":["post-45137","post","type-post","status-publish","format-standard","hentry","category-server","tag-pytorch","tag-transformer","tag-86"],"yoast_head":"\nEfficient Non-Local Transformer Block: \u56fe\u50cf\u5904\u7406\u4e2d\u7684\u9ad8\u6548\u975e\u5c40\u90e8\u6ce8\u610f\u529b\u673a\u5236 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/45137.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"Efficient Non-Local Transformer Block: \u56fe\u50cf\u5904\u7406\u4e2d\u7684\u9ad8\u6548\u975e\u5c40\u90e8\u6ce8\u610f\u529b\u673a\u5236 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u6587\u7ae0\u6d4f\u89c8\u9605\u8bfb430\u6b21\uff0c\u70b9\u8d5e4\u6b21\uff0c\u6536\u85cf5\u6b21\u3002\u672c\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u9ad8\u6548\u975e\u5c40\u90e8\u6ce8\u610f\u529b\u673a\u5236(ENLA)\u53ca\u5176\u7ec4\u6210\u7684Transformer\u6a21\u5757(ENLTB)\uff0c\u7528\u4e8e\u89e3\u51b3\u56fe\u50cf\u5904\u7406\u4e2d\u4f20\u7edf\u81ea\u6ce8\u610f\u529b\u8ba1\u7b97\u590d\u6742\u5ea6\u8fc7\u9ad8\u7684\u95ee\u9898\u3002\u901a\u8fc7\u5377\u79ef\u964d\u7ef4\u3001\u7279\u5f81\u5339\u914d\u548c\u5168\u5c40\u6c60\u5316\u7b49\u6280\u672f\uff0cENLTB\u663e\u8457\u964d\u4f4e\u4e86\u8ba1\u7b97\u91cf(\u590d\u6742\u5ea6\u4eceO(H\u00b2W\u00b2C)\u964d\u4f4e)\uff0c\u540c\u65f6\u4fdd\u6301\u6a21\u578b\u6027\u80fd\u3002\u5b9e\u9a8c\u8868\u660eENLTB\u5728\u56fe\u50cf\u4efb\u52a1\u4e2d\u901f\u5ea6\u63d0\u9ad83-5\u500d\uff0c\u53c2\u6570\u91cf\u51cf\u5c1110%\u4ee5\u4e0a\u3002\u6587\u7ae0\u8be6\u7ec6\u4ecb\u7ecd\u4e86ENLTB\u7684\u67b6\u6784\u8bbe\u8ba1\uff0c\u5305\u62ec\u5377\u79ef\u5339\u914d\u7f51\u7edc\u3001LayerNorm\u3001ENLA\u6a21\u5757\u548cMLP\u524d\u9988\u7f51\u7edc\uff0c\u5e76\u63d0\u4f9b\u4e86PyTorch\u5b9e\u73b0\u4ee3\u7801\u3002\u8be5\u65b9\u6cd5\u7279\u522b\u9002\u5408\u8d44\u6e90\u53d7\u9650\u7684\u5b9e\u65f6\u5e94\u7528\u3002\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/45137.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2025-06-22T05:19:46+00:00\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"2 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/45137.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/45137.html\",\"name\":\"Efficient Non-Local Transformer Block: \u56fe\u50cf\u5904\u7406\u4e2d\u7684\u9ad8\u6548\u975e\u5c40\u90e8\u6ce8\u610f\u529b\u673a\u5236 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2025-06-22T05:19:46+00:00\",\"dateModified\":\"2025-06-22T05:19:46+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/45137.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/45137.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/45137.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"Efficient Non-Local Transformer Block: \u56fe\u50cf\u5904\u7406\u4e2d\u7684\u9ad8\u6548\u975e\u5c40\u90e8\u6ce8\u610f\u529b\u673a\u5236\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"Efficient Non-Local Transformer Block: \u56fe\u50cf\u5904\u7406\u4e2d\u7684\u9ad8\u6548\u975e\u5c40\u90e8\u6ce8\u610f\u529b\u673a\u5236 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/45137.html","og_locale":"zh_CN","og_type":"article","og_title":"Efficient Non-Local Transformer Block: \u56fe\u50cf\u5904\u7406\u4e2d\u7684\u9ad8\u6548\u975e\u5c40\u90e8\u6ce8\u610f\u529b\u673a\u5236 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u6587\u7ae0\u6d4f\u89c8\u9605\u8bfb430\u6b21\uff0c\u70b9\u8d5e4\u6b21\uff0c\u6536\u85cf5\u6b21\u3002\u672c\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u9ad8\u6548\u975e\u5c40\u90e8\u6ce8\u610f\u529b\u673a\u5236(ENLA)\u53ca\u5176\u7ec4\u6210\u7684Transformer\u6a21\u5757(ENLTB)\uff0c\u7528\u4e8e\u89e3\u51b3\u56fe\u50cf\u5904\u7406\u4e2d\u4f20\u7edf\u81ea\u6ce8\u610f\u529b\u8ba1\u7b97\u590d\u6742\u5ea6\u8fc7\u9ad8\u7684\u95ee\u9898\u3002\u901a\u8fc7\u5377\u79ef\u964d\u7ef4\u3001\u7279\u5f81\u5339\u914d\u548c\u5168\u5c40\u6c60\u5316\u7b49\u6280\u672f\uff0cENLTB\u663e\u8457\u964d\u4f4e\u4e86\u8ba1\u7b97\u91cf(\u590d\u6742\u5ea6\u4eceO(H\u00b2W\u00b2C)\u964d\u4f4e)\uff0c\u540c\u65f6\u4fdd\u6301\u6a21\u578b\u6027\u80fd\u3002\u5b9e\u9a8c\u8868\u660eENLTB\u5728\u56fe\u50cf\u4efb\u52a1\u4e2d\u901f\u5ea6\u63d0\u9ad83-5\u500d\uff0c\u53c2\u6570\u91cf\u51cf\u5c1110%\u4ee5\u4e0a\u3002\u6587\u7ae0\u8be6\u7ec6\u4ecb\u7ecd\u4e86ENLTB\u7684\u67b6\u6784\u8bbe\u8ba1\uff0c\u5305\u62ec\u5377\u79ef\u5339\u914d\u7f51\u7edc\u3001LayerNorm\u3001ENLA\u6a21\u5757\u548cMLP\u524d\u9988\u7f51\u7edc\uff0c\u5e76\u63d0\u4f9b\u4e86PyTorch\u5b9e\u73b0\u4ee3\u7801\u3002\u8be5\u65b9\u6cd5\u7279\u522b\u9002\u5408\u8d44\u6e90\u53d7\u9650\u7684\u5b9e\u65f6\u5e94\u7528\u3002","og_url":"https:\/\/www.wsisp.com\/helps\/45137.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2025-06-22T05:19:46+00:00","author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"2 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/45137.html","url":"https:\/\/www.wsisp.com\/helps\/45137.html","name":"Efficient Non-Local Transformer Block: \u56fe\u50cf\u5904\u7406\u4e2d\u7684\u9ad8\u6548\u975e\u5c40\u90e8\u6ce8\u610f\u529b\u673a\u5236 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2025-06-22T05:19:46+00:00","dateModified":"2025-06-22T05:19:46+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/45137.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/45137.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/45137.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"Efficient Non-Local Transformer Block: \u56fe\u50cf\u5904\u7406\u4e2d\u7684\u9ad8\u6548\u975e\u5c40\u90e8\u6ce8\u610f\u529b\u673a\u5236"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/45137","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=45137"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/45137\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=45137"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=45137"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=45137"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=45137"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}