{"id":396,"date":"2021-08-23T17:19:20","date_gmt":"2021-08-23T09:19:20","guid":{"rendered":"https:\/\/aif.amtbbs.org\/?p=396"},"modified":"2021-08-23T17:19:20","modified_gmt":"2021-08-23T09:19:20","slug":"%e5%9b%be%e7%a5%9e%e7%bb%8f%e7%bd%91%e7%bb%9c%ef%bc%8c%e5%a6%82%e4%bd%95%e5%8f%98%e6%b7%b1%ef%bc%9f","status":"publish","type":"post","link":"https:\/\/aif.amtbbs.org\/index.php\/2021\/08\/23\/396\/","title":{"rendered":"\u56fe\u795e\u7ecf\u7f51\u7edc\uff0c\u5982\u4f55\u53d8\u6df1\uff1f"},"content":{"rendered":"
\"\u56fe\u7247\"<\/section>\n

\u6765\u6e90\uff1a\u667a\u6e90\u793e\u533a<\/strong><\/p>\n

\u7248\u5f0f\u5236\u4f5c\uff1aCamel<\/strong><\/p>\n

\u64b0\u6587\uff1a\u4faf\u632f\u5b87
\n<\/strong><\/p>\n

\u7f16\u8f91\uff1a\u8d3e \u00a0 \u4f1f<\/strong><\/p>\n

\u73b0\u5b9e\u4e16\u754c\u4e2d\u5f88\u591a\u5f88\u591a\u4efb\u52a1\u53ef\u4ee5\u63cf\u8ff0\u4e3a\u56fe(Graph<\/em>)\u95ee\u9898\uff0c\u6bd4\u5982\u793e\u4ea4\u7f51\u7edc\uff0c\u86cb\u767d\u8d28\u7ed3\u6784\uff0c\u4ea4\u901a\u8def\u7f51\u6570\u636e\uff0c\u56fe\u662f\u4e00\u4e2a\u5f88\u503c\u5f97\u7814\u7a76\u7684\u9886\u57df\u3002<\/p>\n

\u8fd1\u5e74\u6765\uff0c\u968f\u7740\u6df1\u5ea6\u5b66\u4e60\u7684\u53d1\u5c55\uff0c\u7814\u7a76\u4eba\u5458\u501f\u9274\u4e86CNN\u7b49\u795e\u7ecf\u7f51\u7edc\u7684\u601d\u60f3\uff0c\u5b9a\u4e49\u548c\u8bbe\u8ba1\u4e86\u7528\u4e8e\u5904\u7406\u56fe\u6570\u636e\u7684\u795e\u7ecf\u7f51\u7edc\u7ed3\u6784\u2014\u2014\u56fe\u795e\u7ecf\u7f51\u7edc\uff08Graph Neural Networks\uff0cGNN<\/em>\uff09\u3002<\/p>\n

\u4e0d\u540c\u4e8eCNN\u53ef\u4ee5\u901a\u8fc7\u5806\u53e0\u975e\u5e38\u591a\u5c42\u6570\u7684\u795e\u7ecf\u7f51\u7edc\u6765\u53d6\u5f97\u66f4\u597d\u7684\u6a21\u578b\u8868\u73b0\uff0c\u81ea\u4ece\u300c\u56fe\u795e\u7ecf\u7f51\u7edc\u300d\u7684\u6982\u5ff5\u88ab\u63d0\u51fa\u540e\uff0c\u5927\u90e8\u5206\u56fe\u795e\u7ecf\u7f51\u7edc\u4ec5\u4ec5\u5806\u53e0\u4e24\u5c42\u5c31\u80fd\u591f\u5728\u4e0b\u6e38\u4efb\u52a1\u4e0a\u53d6\u5f97\u8f83\u597d\u7684\u6548\u679c\u3002<\/p>\n

\u90a3\u4e48\u81ea\u7136\u63d0\u51fa\u4e00\u4e2a\u95ee\u9898\uff1a\u6df1\u5c42\u7684\u56fe\u795e\u7ecf\u7f51\u7edc\u5fc5\u8981\u5417\uff1f<\/p>\n

\u5f53\u7f51\u7edc\u5c42\u6570\u52a0\u6df1\u65f6\uff0c\u6309\u7167\u5e38\u89c1\u7684\u89e3\u91ca\uff0c\u4f1a\u53d1\u751f\u8fc7\u5e73\u6ed1(over-smoothing<\/em>)\u6216\u8005\u8fc7\u538b\u7f29(over-squashing<\/em>)\u73b0\u8c61\uff0c\u5bfc\u81f4\u4e0d\u540c\u8282\u70b9\u4e4b\u95f4\u7684\u533a\u5206\u5ea6\u548c\u8282\u70b9\u81ea\u8eab\u7684\u7279\u5f81\u8868\u8fbe\u4e0b\u964d\uff0c\u4ece\u800c\u4f7f\u5f97\u6df1\u5c42\u7684\u56fe\u795e\u7ecf\u7f51\u7edc\u6548\u679c\u53cd\u800c\u4e0d\u5982\u6d45\u5c42\u7f51\u7edc\u3002<\/p>\n

\u4f46\u9274\u4e8eCV\/NLP\u4e2d\u7684\u76f8\u5173\u7ecf\u9a8c\uff0c\u6d45\u5c42\u7f51\u7edc\u5bf9\u6570\u636e\u7684\u62df\u5408\u80fd\u529b\u4f1a\u5f31\u4e00\u4e9b\uff0c\u56e0\u6b64\u5bf9\u5982\u4f55\u6784\u5efa\u6df1\u5c42\u56fe\u795e\u7ecf\u7f51\u7edc\u7684\u63a2\u7d22\u4ece\u672a\u505c\u6b62\uff0c\u4f8b\u5982\u6df1\u5c42\u56fe\u795e\u7ecf\u7f51\u7edcGCNII\u5728Cora\/PPI\u7b49\u6570\u636e\u96c6\uff0cRevGEN\u5728ogbn-proteins\u6570\u636e\u96c6\u4e0a\u90fd\u53d6\u5f97\u4e86state-of-art\u7684\u7ed3\u679c\uff0c\u6211\u4eec\u4ecd\u7136\u5e0c\u671b\u770b\u5230\u6df1\u5c42\u56fe\u795e\u7ecf\u7f51\u7edc\u672a\u6765\u7684\u6f5c\u529b\u3002<\/p>\n

\u6df1\u5c42\u7684\u56fe\u795e\u7ecf\u7f51\u7edc\u4e00\u65b9\u9762\u589e\u52a0\u4e86\u6bcf\u4e2a\u8282\u70b9\u7684\u611f\u53d7\u91ce\uff0c\u4f7f\u5f97\u8282\u70b9\u53ef\u4ee5\u611f\u77e5\u5230\u66f4\u5927\u90bb\u57df\u7684\u62d3\u6251\u548c\u8282\u70b9\u7279\u5f81\u4fe1\u606f\uff1b\u53e6\u4e00\u65b9\u9762\uff0c\u6df1\u5c42\u7f51\u7edc\u589e\u5927\u4e86\u6a21\u578b\u7684\u590d\u6742\u5ea6\u548c\u6a21\u578b\u5bb9\u91cf\uff0c\u4f7f\u5f97\u7f51\u7edc\u7684\u62df\u5408\u80fd\u529b\u66f4\u5f3a\u3002<\/p>\n

<\/section>\n
\u00a0<\/section>\n
\n
\n
\n
\n
\n
\n
\n

01<\/strong><\/p>\n<\/section>\n<\/section>\n<\/section>\n

\n
<\/section>\n
<\/section>\n<\/section>\n
\n
\n
\u5982\u4f55\u8bbe\u8ba1\u6df1\u5c42GNN\uff1f<\/strong><\/p>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n
<\/section>\n

\u90a3\u4e48\u5982\u4f55\u8bbe\u8ba1\u4e00\u4e2a\u6df1\u5c42\u7684GNN\u5462\uff1f<\/p>\n

\u56fe\u795e\u7ecf\u7f51\u7edc\u53d8\u6df1\uff0c\u9700\u8981\u9762\u4e34\u7684\u6700\u4e3b\u8981\u7684\u95ee\u9898\u5728\u4e8e\uff1a\u5982\u4f55\u514b\u670dover-smoothing\u548cover-squashing\u95ee\u9898\uff0c\u4ee5\u53ca\u5982\u4f55\u8bad\u7ec3\u597d\u4e00\u4e2a\u6df1\u5c42\u7684\u56fe\u795e\u7ecf\u7f51\u7edc\u6a21\u578b\u3002<\/p>\n

\u5f53\u7ed3\u5408\u6240\u6709\u5c42\u7684\u8f93\u51fa\uff0c\u56fe\u795e\u7ecf\u7f51\u7edc\u5c42\u6570\u8fc7\u6df1\u3001\u4f20\u64ad\u6b21\u6570\u8fc7\u591a\u4f1a\u5e26\u6765\u5927\u90e8\u5206\u8282\u70b9\u7684\u8fc7\u5e73\u6ed1\uff0c\u800c\u6211\u4eec\u53c8\u5e0c\u671b\u901a\u8fc7\u6355\u83b7\u9ad8\u9636\u90bb\u5c45\u7684\u4fe1\u606f\u6765\u5e2e\u52a9\u83b7\u5f97\u66f4\u597d\u7684\u8282\u70b9\u8868\u793a\uff0cJKNet\uff08Jumping Knowledge Networks<\/em>\uff09[2]\u7ed9\u51fa\u7684\u89e3\u51b3\u65b9\u6cd5\u662f\uff1a\u8003\u8651\u6bcf\u4e00\u5c42\u7f51\u7edc\u7684\u8f93\u51fa\u7ed3\u679c\uff0c\u5c06\u6bcf\u4e00\u5c42\u7684\u8282\u70b9\u8f93\u51fa\u8fdb\u884c\u878d\u5408\uff0c\u4f7f\u5f97\u7f51\u7edc\u80fd\u591f\u81ea\u9002\u5e94\u5730\u5b66\u4e60\u6574\u5408\u4e0d\u540c\u5c42\u7684\u4fe1\u606f\uff0c\u5f97\u5230\u6700\u7ec8\u7684\u8282\u70b9\u8868\u793a\uff1a<\/p>\n

\"\u56fe\u7247\"<\/p>\n

\"\u56fe\u7247\"<\/p>\n

4\u5c42JKNet<\/p>\n

\u53e6\u5916GRAND[10]\u662f\u5c06\u6240\u6709\u4f20\u64ad\u7684\u7ed3\u679c\u53e0\u52a0\u5728\u4e00\u8d77:<\/p>\n

\"\u56fe\u7247\"<\/p>\n

\u4e0d\u8fc7\uff0cJKNet\u8fd9\u7c7b\u6a21\u578b\u5e76\u4e0d\u5c5e\u4e8e\u6df1\u5c42GNN\uff0c\u4e00\u65b9\u9762\u5e76\u6ca1\u6709\u514b\u670dGNN\u52a0\u6df1\u65f6\u5b58\u5728\u7684\u5404\u79cd\u95ee\u9898\uff0c\u53e6\u4e00\u65b9\u9762\u4e5f\u65e0\u6cd5\u5806\u53e0\u5230\u975e\u5e38\u6df1\uff0c\u6a21\u578b\u201c\u81ea\u9002\u5e94\u5b66\u4e60\u878d\u5408\u4e0d\u540c\u5c42\u201d\u7684\u4fe1\u606f\u5f88\u96be\u5b9e\u73b0\u3002\u6240\u4ee5JKNet\u5b9e\u9645\u4e0a\u4ecd\u7136\u5c5e\u4e8e\u6d45\u5c42GNN\uff0c\u53ea\u4e0d\u8fc7\u5728\u5f15\u5165\u4e86\u805a\u5408\u66f4\u9ad8\u9636\u90bb\u5c45\u7684\u4fe1\u606f\u7684\u65b9\u5f0f\u3002<\/p>\n

\u6211\u4eec\u9488\u5bf9\u8fd9\u4e9b\u95ee\u9898\uff0c\u5728\u6700\u65b0\u53d1\u8868\u7684\u8bba\u6587\u300aCogDL: An Extensive Toolkit for Deep Learning on Graphs\u300b\u4e2d\uff0c\u63d0\u51fa\u4e09\u79cd\u7b56\u7565\uff0c\u5305\u62ec\uff1a\u521d\u59cb\u6b8b\u5dee\u8fde\u63a5\u3001\u5f52\u4e00\u5316+\u8df3\u8dc3\u8fde\u63a5\u4ee5\u53ca\u53ef\u9006\u7f51\u7edc\uff08\u514b\u670d\u5185\u5b58\u74f6\u9888<\/em>\uff09\u3002<\/p>\n

\u8bba\u6587\u9898\u76ee\uff1aCogDL: An Extensive Toolkit for Deep Learning on Graphs<\/p>\n

\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/www.aminer.cn\/pub\/603e2f2191e01129ef28fecc<\/p>\n

1\u3001\u521d\u59cb\u6b8b\u5dee\u8fde\u63a5<\/strong><\/p>\n

APPNP \u63d0\u51fa\u5229\u7528PageRank\u6765\u7684\u6a21\u62df\u4f20\u64ad\u8fc7\u7a0b\uff08\u5b9e\u9645\u4e0a\u662f\u4e00\u4e2a\u56fa\u5b9a\u7684\u56fe\u6ee4\u6ce2\u5668<\/em>\uff09\u6765\u53d6\u4ee3\u9010\u5c42\u5806\u53e0\u7684GCN\u3002\u7531\u4e8ePageRank\u7684\u5c40\u90e8\u6027\uff0cAPPNP\u80fd\u591f\u5229\u7528\u56fe\u7684\u5c40\u90e8\u6027\uff0c\u5728\u8282\u70b9\u5206\u7c7b\u4efb\u52a1\u4e0a\u53d6\u5f97\u4e86\u4e0d\u9519\u7684\u8868\u73b0\u3002\u8bb0\u5f52\u4e00\u5316\u540e\u7684\u90bb\u63a5\u77e9\u9635\u00c3\uff0cAPPNP\u7684\u8ba1\u7b97\u53ef\u4ee5\u8868\u793a\u4e3a\uff1a<\/p>\n

\"\u56fe\u7247\"<\/p>\n

\u5728\u5c06\u4fe1\u606f\u4f20\u64ad\u548c\u7279\u5f81\u8f6c\u6362\u89e3\u8026\u7684\u540c\u65f6\uff0c\u5728\u6bcf\u6b21\u4f20\u64ad\u7684\u8fc7\u7a0b\u4e2d\uff0cAPPNP\u90fd\u5c06\u521d\u59cb\u8282\u70b9\u7279\u5f81\u4ee5\u4e00\u5b9a\u7684\u6743\u91cd \u03b1\u00a0\u589e\u53e0\u5230\u6bcf\u6b21\u4f20\u64ad\u7684\u7ed3\u679c\u4e0a\uff0c\u4ece\u800c\u4f7f\u5f97\u5728\u805a\u5408\u90bb\u57df\u8282\u70b9\u4fe1\u606f\u7684\u540c\u65f6\uff0c\u4fdd\u6301\u8282\u70b9\u672c\u8eab\u7684\u7279\u5f81\uff0c\u907f\u514d\u8fc7\u5e73\u6ed1\u73b0\u8c61\u3002\u4f46\u7531\u4e8eAPPNP\u5728\u6bcf\u4e00\u6b21\u4f20\u64ad\u65f6\u53ea\u6709\u7ebf\u6027\u53d8\u6362\u800c\u4e14\u5e76\u65e0\u53ef\u5b66\u4e60\u7684\u6743\u91cd\uff0c\u6240\u4ee5\u4ecd\u7136\u53ea\u80fd\u7b97\u662f\u4e00\u4e2a\u6d45\u5c42\u6a21\u578b\u3002\u5173\u4e8eAPPNP\u7684\u5b9e\u73b0\u53ef\u4ee5\u53c2\u89c1\uff1a<\/p>\n

https:\/\/github.com\/THUDM\/cogdl\/blob\/master\/cogdl\/models\/nn\/ppnp.py<\/p>\n

\u4e3a\u4e86\u6784\u5efa\u771f\u6b63\u7684\u6df1\u5c42\u56fe\u795e\u7ecf\u7f51\u7edc\uff0cGCNII \u5728\u6b64APPNP\u57fa\u7840\u4e0a\uff0c\u63d0\u51fa\u4e86\u4e24\u4e2a\u6280\u5de7\uff1a\u521d\u59cb\u6b8b\u5dee\u8fde\u63a5\uff08Initial residual connection<\/em>\uff09\u548c\u6052\u7b49\u6620\u5c04\uff08Identity Mapping<\/em>\uff09\u3002<\/p>\n

Initial Residual Connection(\u521d\u59cb\u6b8b\u5dee\u8fde\u63a5)\u00a0<\/strong>\u4e0eAPPNP\u4e2d\u76f8\u540c\uff0cGCNII \u6bcf\u6b21\u5c06\u521d\u59cb\u7279\u5f81$H^{(0)}$ \u4ee5\u4e00\u5b9a\u7684\u6743\u91cd\u589e\u52a0\u5230\u6bcf\u5c42\u7684\u4f20\u64ad\u7ed3\u679c\u4e0a\uff0c\u6765\u4f7f\u5f97\u4f20\u64ad\u7684\u6bcf\u4e00\u5c42\u90fd\u80fd\u4fdd\u6301\u81f3\u5c11\u4e00\u5b9a\u6bd4\u4f8b\u7684\u8282\u70b9\u7684\u521d\u59cb\u7279\u5f81\uff1a(1-\u03b1\u03b9)\u00c3H(\u03b9)+\u03b1\u03b9H(0)\u5728\u4ee3\u7801\u5b9e\u73b0\u4e2d\uff0c\u03b1\u03b9,\u2200\u03b9=0,1,…,L\u53d6\u76f8\u540c\u4e14\u56fa\u5b9a\u7684\u503c\u3002<\/p>\n

Identity Mapping(\u6052\u7b49\u6620\u5c04<\/em>)\u00a0\u5728ResNet\u4e2d\uff0c\u6b8b\u5dee\u7684\u5b58\u5728\u4f7f\u5f97\u7f51\u7edc\u5728\u6df1\u5ea6\u589e\u52a0\u65f6\uff0c\u81f3\u5c11\u53ef\u4ee5\u4fdd\u6301\u201c\u6a21\u578b\u8868\u73b0\u4e0d\u4e0b\u964d\u201d\uff0c\u53ea\u8981\u5728\u7f51\u7edc\u6df1\u5ea6\u4e0d\u65ad\u52a0\u6df1\u65f6\u5b66\u4e60\u5230\u4e00\u4e2a\u6052\u7b49\u6620\u5c04\u5373\u53ef\u3002\u5728GCNII\u4f5c\u8005\u7684\u8bba\u8ff0\u4e2d\uff0cGCNII\u501f\u7528\u76f8\u4f3c\u7684\u601d\u8def\uff0c\u8bbe\u8ba1\u4e86\u7279\u5f81\u53d8\u6362\u7684\u6052\u7b49\u6620\u5c04\uff1a<\/p>\n

\"\u56fe\u7247\"<\/p>\n

\u5728\u5b9e\u9645\u5b9e\u73b0\u4e2d\uff0c\\beta_l\u4e3a\u56fa\u5b9a\u503c(0, 0.5)\uff0c\u968f\u7740\u5c42\u6570l\u7684\u589e\u5927\u800c\u51cf\u5c0f\u3002\u5728\u5173\u4e8e\u8fc7\u5e73\u6ed1\u7684\u5206\u6790\u4e2d\uff0c\u7531\u4e8e\u4e00\u822cW(\u03b9)\u7684\u6700\u5927\u7279\u5f81\u503c\uff1a<\/p>\n

\"\u56fe\u7247\"<\/p>\n

\u56e0\u6b64W(\u03b9)\u7684\u5b58\u5728\u4f1a\u52a0\u5267\u8fc7\u5e73\u6ed1\u7684\u51fa\u73b0\uff0cIdentity Mapping\u5219\u589e\u5927\u4e86\u201c\u7279\u5f81\u53d8\u6362\u201d\u77e9\u9635\u7684\u7279\u5f81\u503c\uff1a<\/p>\n

\"\u56fe\u7247\"<\/p>\n

\u56e0\u6b64\u6709\u52a9\u4e8e\u7f13\u89e3\u8fc7\u5e73\u6ed1\u73b0\u8c61\u3002<\/p>\n

\u56e0\u6b64\uff0cGCNII \u7684\u5f62\u5f0f\u4e3a\uff1a<\/p>\n

\"\u56fe\u7247\"<\/p>\n

\u5176\u4e2d\u03c3(\u00b7)\u4e3a\u975e\u7ebf\u6027\u53d8\u6362\u3002<\/p>\n

\u4e00\u822c\u5730\uff0c\u7531\u4e8e\u03b1\u03b9\u22600\uff0c\u03b2\u03b9\u22600\uff0c\u6240\u4ee5\u6bcf\u4e00\u5c42GCNII\u5b9e\u9645\u4e0a\u5e76\u4e0d\u80fd\u5b66\u4e60\u5230\u6052\u7b49\u6620\u5c04\uff0c\u5728\u6bcf\u4e00\u5c42\u4ecd\u7136\u4f1a\u4e0d\u65ad\u805a\u5408\u66f4\u9ad8\u9636\u90bb\u5c45\u7684\u4fe1\u606f\u3002<\/p>\n

\u800c\u6d88\u878d\u5b9e\u9a8c\u4e5f\u8bf4\u660e\uff0cInitial Residual Connection \u5728\u62d3\u5c55\u7f51\u7edc\u6df1\u5ea6\u3001\u907f\u514d\u8fc7\u5e73\u6ed1\u73b0\u8c61\u4e2d\u53d1\u6325\u4e86\u4e3b\u8981\u4f5c\u7528\uff0c\u800cIdentity Mapping\u5728\u6709\u52a9\u4e8e\u7f13\u89e3\u8fc7\u5e73\u6ed1\u73b0\u8c61\u7684\u540c\u65f6\uff0c\u80fd\u591f\u8fdb\u4e00\u6b65\u589e\u5f3a\u6a21\u578b\u7684\u6548\u679c\u3002<\/p>\n

GCNII\u7684\u5b9e\u73b0\u53ef\u4ee5\u53c2\u89c1\uff1a<\/p>\n

https:\/\/github.com\/THUDM\/cogdl\/blob\/master\/cogdl\/models\/nn\/gcnii.py<\/p>\n

\"\u56fe\u7247\"<\/p>\n

GCNII\u6d88\u878d\u5b9e\u9a8c<\/p>\n

2\u3001\u5f52\u4e00\u5316 + \u8df3\u8dc3\u8fde\u63a5<\/strong><\/p>\n

CNN\u5229\u7528BatchNorm\u548c\u8df3\u8dc3\u8fde\u63a5(Skip-connection<\/em>)\uff0c\u80fd\u591f\u5c06\u7f51\u7edc\u6df1\u5ea6\u505a\u5230\u4e0a\u524d\u5c42\u3002BatchNorm\u6709\u6548\u6539\u5584\u4e86\u6df1\u5c42\u7f51\u7edc\u8bad\u7ec3\u4e2d\u51fa\u73b0\u7684\u68af\u5ea6\u5f25\u6563\u73b0\u8c61\uff0cskip-connection\u5219\u4f7f\u5f97CNN\u80fd\u591f\u5728\u589e\u52a0\u7f51\u7edc\u6df1\u5ea6\u7684\u540c\u65f6\u80fd\u591f\u53d6\u5f97\u4e0d\u5f31\u4e8e\u6d45\u5c42\u7f51\u7edc\u7684\u8868\u73b0\u3002<\/p>\n

\"\u56fe\u7247\"<\/p>\n

\"\u56fe\u7247\"<\/p>\n

\u8df3\u8dc3\u8fde\u63a5\/\u6b8b\u5dee<\/p>\n

\u540c\u6837\u5730\uff0c\u5f52\u4e00\u5316\u548c\u8df3\u8dc3\u8fde\u63a5\u5728\u8bbe\u8ba1\u6df1\u5c42\u56fe\u795e\u7ecf\u7f51\u7edc\u65f6\u4e5f\u80fd\u591f\u6709\u6548\u53d1\u6325\u4f5c\u7528\u3002DeeperGCN\u5c06\u5f52\u4e00\u5316\u548c\u8df3\u8dc3\u8fde\u63a5\u6574\u5408\u5230\u4e86\u56fe\u795e\u7ecf\u7f51\u7edc\u4e2d\uff0c\u8bbe\u8ba1\u4e86DeeperGCN\uff0c\u540c\u65f6\u8bbe\u8ba1\u4e86\u201c\u63d0\u524d\u6fc0\u6d3b(pre-activation<\/em>)”\u5f62\u5f0f\u7684\u6b8b\u5dee\u56fe\u795e\u7ecf\u7f51\u7edc\uff0c\u5728\u5b9e\u9a8c\u4e2d\uff0c\u5f53\u5c42\u6570\u589e\u52a0\u5230100+\u5c42\u65f6\uff0cDeeperGCN\u4ecd\u7136\u80fd\u591f\u5728ogbn-proteins\u6570\u636e\u96c6\u4e0a\u53d6\u5f97\u975e\u5e38\u4f18\u5f02\u7684\u8868\u73b0\u3002<\/p>\n

ResGCN:\u00a0<\/strong><\/p>\n

\"\u56fe\u7247\"<\/p>\n

ResGCN+:\u00a0<\/strong><\/p>\n

\"\u56fe\u7247\"<\/p>\n

\u4e0d\u8fc7\uff0c\u5728DeeperGCN\u8bba\u6587\u4e2d\uff0c\u4f5c\u8005\u5e76\u6ca1\u6709\u6d4b\u8bd5\u5f53\u7f51\u7edc\u5c42\u6570\u52a0\u6df1\u65f6\uff0c\u6a21\u578b\u5728Cora\/Citeseer\/PubMed\u8fd9\u4e9b\u5e38\u7528\u534a\u76d1\u7763\u8282\u70b9\u5206\u7c7b\u6570\u636e\u96c6\u4e0a\u7684\u6548\u679c\u3002\u56e0\u4e3a\u5177\u6709\u201c\u7a00\u758f\u6027\u9ad8\uff0c\u56fe\u89c4\u6a21\u3001\u76f4\u5f84\u8f83\u5c0f\u201d\u7b49\u7279\u70b9\uff0c\u5173\u4e8e\u8fc7\u5e73\u6ed1\u73b0\u8c61\u5e38\u89c1\u7684\u8ba8\u8bba\u4e3b\u8981\u5173\u6ce8\u5728\u8fd9\u4e09\u4e2a\u6570\u636e\u96c6\u4e0a\u3002\u6211\u4eec\u5229\u7528CogDL\u4e2d\u590d\u73b0\u7684DeeperGCN\u4ee3\u7801\u53d1\u73b0\uff0c\u6df1\u5c42\u7684DeeperGCN\u867d\u7136\u672a\u53d1\u751f\u8fc7\u5e73\u6ed1\uff0c\u4f46\u5728\u4e09\u4e2a\u534a\u76d1\u7763\u7684\u6570\u636e\u96c6\u4e0a\u7684\u8868\u73b0\u7565\u5fae\u900a\u8272\u4e00\u4e9b\u3002<\/p>\n

\u5728DeeperGCN\u4e2d\uff0c\u5f52\u4e00\u5316\u548c\u8df3\u8dc3\u8fde\u63a5\u662f\u4e00\u8d77\u4f7f\u7528\u7684\uff0c\u9664\u4e86\u51cf\u7f13\u8fc7\u5e73\u6ed1\u73b0\u8c61\u5916\uff0c\u5f52\u4e00\u5316\u548c\u6b8b\u5dee\u64cd\u4f5c\u90fd\u80fd\u591f\u63d0\u5347\u8bad\u7ec3\u7684\u7a33\u5b9a\u6027\u548c\u6536\u655b\u6027\uff0c\u4f7f\u5f97\u6df1\u5c42\u56fe\u795e\u7ecf\u7f51\u7edc\u7684\u8bad\u7ec3\u66f4\u52a0\u5bb9\u6613\uff0c\u635f\u5931\u66f2\u7ebf\u66f4\u52a0\u5e73\u6ed1\u3002\u7279\u522b\u5730\uff0c\u5f52\u4e00\u5316\u80fd\u591f\u63d0\u9ad8\u6df1\u5ea6\u7f51\u7edc\u8fd0\u7b97\u8fc7\u7a0b\u4e2d\u7684\u6570\u503c\u7a33\u5b9a\u6027\uff0c\u907f\u514d\u8fd0\u7b97\u8fc7\u7a0b\u4e2d\u51fa\u73b0\u6570\u503c\u7206\u70b8\uff0c\u800c\u6b8b\u5dee\u8fde\u63a5\u80fd\u591f\u63d0\u9ad8\u6df1\u5c42\u56fe\u795e\u7ecf\u7f51\u7edc\u7684\u6a21\u578b\u8bad\u7ec3\u8868\u73b0\u3002<\/p>\n

\u6709\u8da3\u7684\u662f\uff0cGCNII\u5728\u5b9e\u73b0\u4e0a\u9664\u4e86\u5229\u7528Initial Residual Connection\u4e4b\u5916\uff0c\u6709\u65f6\u4e5f\u4f1a\u4f7f\u7528Residual Connection\u6765\u63d0\u5347\u6a21\u578b\u7684\u8868\u73b0\uff0c\u8fd9\u8868\u660e\u4e86Residual Connection\u5728\u52a0\u6df1\u56fe\u795e\u7ecf\u7f51\u7edc\u65f6\u662f\u4e00\u4e2a\u975e\u5e38\u91cd\u8981\u7684\u6280\u5de7\u3002<\/p>\n

CogDL\u4e2d\u5b9e\u73b0\u4e86\u4ee5\u4e0a\u63d0\u5230\u7684GCNII\u548cDeeperGCN\uff0c\u5305\u62ec\u66f4\u591a\u7684\u53c2\u6570\u6a21\u5757\uff0c\u53ef\u4ee5\u8fdb\u884c\u66f4\u591a\u5730\u6d88\u878d\u5b9e\u9a8c\u6765\u63a2\u7d22\u6bcf\u4e00\u90e8\u5206\u7684\u5f71\u54cd\u548c\u4f5c\u7528\u3002<\/p>\n

3\u3001\u514b\u670d\u5185\u5b58\u74f6\u9888<\/strong><\/p>\n

\u5f53\u7f51\u7edc\u5c42\u6570\u589e\u52a0\u65f6\uff0c\u4e00\u65b9\u9762\u6a21\u578b\u7684\u8ba1\u7b97\u901f\u5ea6\u4f1a\u964d\u4f4e\uff0c\u56e0\u4e3a\u9700\u8981\u8fdb\u884c\u66f4\u591a\u6b21\u7684\u4f20\u64ad\u548c\u8fd0\u7b97\uff1b\u53e6\u4e00\u65b9\u9762\uff0c\u7f51\u7edc\u5c42\u6570\u7684\u589e\u52a0\u4e5f\u4f1a\u5e26\u6765\u5185\u5b58\u5360\u7528\u7684\u7ebf\u6027\u589e\u52a0\uff0c\u5728GPU\u663e\u5b58\u6709\u9650\u7684\u60c5\u51b5\u4e0b\uff0c\u786c\u4ef6\u8d44\u6e90\u5f80\u5f80\u4e5f\u662f\u7f51\u7edc\u5c42\u6570\u589e\u52a0\u7684\u4e00\u4e2a\u91cd\u8981\u9650\u5236\u3002\u7f51\u7edc\u5c42\u6570\u589e\u52a0\u5e26\u6765\u7684\u5185\u5b58\u5360\u7528\u589e\u52a0\u4e3b\u8981\u5305\u62ec2\u4e2a\u90e8\u5206\uff1a<\/p>\n

    \n
  • \n
    \u6a21\u578b\u53c2\u6570\u7684\u589e\u52a0\uff0c\u8fd9\u4e00\u90e8\u5206\u5f80\u5f80\u975e\u5e38\u5c0fO(d\u00b2)\uff0c\u6240\u4ee5\u5f80\u5f80\u53ef\u4ee5\u5ffd\u7565\uff1b<\/em><\/section>\n<\/li>\n
  • \n
    \u9700\u8981\u4fdd\u5b58\u6bcf\u4e00\u5c42\u524d\u5411\u8ba1\u7b97\u7684\u7ed3\u679cO(Nd)\u00a0\u00a0\uff0c\u7528\u4e8e\u53cd\u5411\u4f20\u64ad\u65f6\u68af\u5ea6\u7684\u8ba1\u7b97\uff0c\u8fd9\u4e00\u90e8\u5206\u5360\u7528\u5185\u5b58\u8f83\u9ad8\u3002<\/em><\/section>\n<\/li>\n<\/ul>\n

    \u4e3a\u4e86\u7f13\u89e3\u6a21\u578b\u5c42\u6570\u589e\u52a0\u5e26\u6765\u7684\u5185\u5b58\u7ebf\u6027\u589e\u957f\uff0c[7]\u63d0\u51fa\u4e86\u53ef\u9006\u7f51\u7edc\uff08RevNets<\/em>\uff09\uff0c\u5373\u5f53\u524d\u5c42\u7684\u6fc0\u6d3b\u8ba1\u7b97\u7ed3\u679c\u53ef\u4ee5\u7531\u4e0b\u4e00\u5c42\u7684\u8ba1\u7b97\u7ed3\u679c\u8ba1\u7b97\u5f97\u51fa\u3002\u4e5f\u5c31\u662f\u8bf4\uff0c\u5982\u679c\u6211\u4eec\u77e5\u9053\u7f51\u7edc\u6700\u540e\u4e00\u5c42\u7684\u7ed3\u679c\uff0c\u5c31\u53ef\u4ee5\u53cd\u63a8\u51fa\u4e2d\u95f4\u6bcf\u4e00\u5c42\u7684\u7ed3\u679c\u3002\u56e0\u6b64\uff0c\u5728\u795e\u7ecf\u7f51\u7edc\u8fdb\u884c\u524d\u5411\u4f20\u64ad\u7684\u8fc7\u7a0b\u4e2d\uff0c\u4e0d\u518d\u9700\u8981\u4fdd\u5b58\u4e2d\u95f4\u8ba1\u7b97\u7ed3\u679c\uff0c\u4ece\u800c\u6781\u5927\u5730\u8282\u7701GPU\u663e\u5b58\u7684\u5360\u7528\u3002<\/p>\n

    \"\u56fe\u7247\"<\/p>\n

    \u4e3a\u4e86\u5b9e\u73b0\u53ef\u9006\u7f51\u7edc\uff0c\u6211\u4eec\u5c06\u8f93\u5165X\u5207\u5206\u62102\u90e8\u5206(X\u2081,X\u2082)\uff0c\u56e0\u6b64\u4f1a\u5f97\u5230\u4e24\u4e2a\u8f93\u51fa(Y\u2081,Y\u2082)\u3002\u5728\u8ba1\u7b97\u4e2d\uff0c\u6211\u4eec\u6709\uff1a<\/p>\n

    \"\u56fe\u7247\"<\/p>\n

    F \u548c G \u8868\u793a\u7f51\u7edc\u524d\u5411\u4f20\u64ad\u7684\u8ba1\u7b97\u8fc7\u7a0b\uff0c\u5728\u5b9e\u73b0\u4e2d\u5f80\u5f80\u53d6 F = G\u3002<\/p>\n

    \u5728\u540e\u5411\u4f20\u64ad\u65f6\uff0c\u6211\u4eec\u53ef\u4ee5\u5229\u7528\u8ba1\u7b97\u7ed3\u679c\u53cd\u63a8\u51fa\u8f93\u5165\uff0c\u5373\u6bcf\u4e00\u5c42\u7684\u8f93\u51fa\u53ef\u4ee5\u7531\u4e0b\u4e00\u5c42\u7684\u8f93\u51fa\u8ba1\u7b97\u5f97\u5230\uff1a<\/p>\n

    \"\u56fe\u7247\"<\/p>\n

    \u56e0\u6b64\uff0c\u7f51\u7edc\u5728\u5b9e\u9645\u8ba1\u7b97\u65f6\uff0c\u5728\u53cd\u5411\u4f20\u64ad\u65f6\u53ea\u9700\u8981\u6839\u636e\u4e0b\u4e00\u5c42\u7684\u7ed3\u679c\u8ba1\u7b97\u4e0a\u4e00\u5c42\u7684\u7ed3\u679c\u5373\u53ef\uff0c\u4e0d\u518d\u5728\u9700\u8981\u4fdd\u5b58\u4e2d\u95f4\u7684\u6fc0\u6d3b\u7ed3\u679c\u3002\u8fd9\u867d\u7136\u5e26\u6765\u4e86\u4e00\u5b9a\u7684\u989d\u5916\u7684\u8ba1\u7b97\u7684\u5f00\u9500\uff08\u7ea6\u989d\u5916\u4e00\u500d\u7684\u524d\u5411\u4f20\u64ad\u8ba1\u7b97\u5f00\u9500<\/em>\uff09\uff0c\u4f46\u6781\u5927\u5730\u8282\u7701\u4e86\u663e\u5b58\u7684\u5360\u7528\uff0c\u4f7f\u5f97\u6a21\u578b\u8ba1\u7b97\u7684\u663e\u5b58\u5360\u7528\u4e0e\u5c42\u6570\u65e0\u5173\u3002<\/p>\n

    \u6240\u4ee5\uff0c\u5982\u679c\u4e0d\u8003\u8651\u6a21\u578b\u53c2\u6570\u5360\u7528\u7684\u663e\u5b58\uff0c\u7406\u8bba\u4e0a\u53ef\u9006\u7f51\u7edc\u53ef\u4ee5\u62d3\u5c55\u5230\u4efb\u610f\u6df1\u5ea6\u3002[6] \u5c06\u53ef\u9006\u7f51\u7edc\u5e94\u7528\u5230\u4e86\u56fe\u795e\u7ecf\u7f51\u7edc\u4e0a\uff0c\u4ece\u800c\u6784\u5efa\u4e86\u4e0a\u767e\u5c42\u751a\u81f3\u4e0a\u5343\u5c42\u7684\u56fe\u795e\u7ecf\u7f51\u7edc\uff0c\u5e76\u5229\u7528448\u5c42\u7684GNN\u5728ogbn-proteins\u6570\u636e\u96c6\u4e0a\u53d6\u5f97\u4e86state-of-art\u7684\u6548\u679c\u3002\u8fd9\u8868\u660e\u6df1\u5c42\u56fe\u795e\u7ecf\u7f51\u7edc\u5728\u4ecd\u7136\u5177\u6709\u6f5c\u529b\u3002<\/p>\n

    <\/section>\n
    <\/section>\n
    \n
    \n
    \n
    \n
    \n
    \n
    \n

    02<\/strong><\/p>\n<\/section>\n<\/section>\n<\/section>\n

    \n
    <\/section>\n
    <\/section>\n<\/section>\n
    \n
    \n
    \u6a21\u578b\u5b9e\u73b0<\/strong><\/p>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n
    <\/section>\n

    \u5728CogDL\u00a0\u4e2d\uff0c\u6211\u4eec\u5b9e\u73b0\u4e86\u6211\u4eec\u5b9e\u73b0\u4e86\u4e0a\u8ff0\u63d0\u5230\u7684\u6240\u6709\u56fe\u795e\u7ecf\u7f51\u7edc\u6a21\u578b\uff0c\u5e76\u653e\u5728Models\u4e0b\uff0c\u53ef\u4ee5\u5c1d\u8bd5\u6a21\u578b\u5728\u4e0d\u540c\u6570\u636e\u96c6\u4e0a\u7684\u6548\u679c\u548c\u8fdb\u884c\u4e0d\u540c\u53c2\u6570\u7684\u8c03\u8282\u5b9e\u9a8c\uff0c\u6a21\u578b\u7684\u8c03\u7528\u53ef\u4ee5\u76f4\u63a5\u5229\u7528\u547d\u4ee4\u884c\u6216\u8005\u00a0cogdl.experiment<\/em>\u00a0<\/em>\u63a5\u53e3\u5b9e\u73b0\uff0c\u540c\u65f6\u5355\u72ec\u7684GNN\u5c42\u653e\u5728Layers\u4e0b\uff0c\u65b9\u4fbf\u8fdb\u884c\u4f7f\u7528\u548c\u6784\u5efa\u65b0\u6a21\u578b\u3002<\/p>\n

    <\/section>\n
    python scripts\/train.py --dataset cora --model gcnii --task node_classification<\/code><\/section>\n
    <\/section>\n

    \u6216\u8005\u76f4\u63a5\u5728\u4ee3\u7801\u4e2d\u4f7f\u7528<\/p>\n

    <\/section>\n
    from<\/strong> cogdl.experiment import<\/strong> experiment
    \nexperiment(task=\"node_classification\", model=\"gcnii\", dataset=\"cora\")<\/em><\/code><\/section>\n
    <\/section>\n

    \u540c\u65f6\uff0c\u6211\u4eec\u5c06DeeperGCN\u7684\u6b8b\u5dee\u8fde\u63a5\u5f62\u5f0f\u548cRevNet\u7684\u53ef\u9006\u5b9e\u73b0\u5c01\u88c5\u6210\u7edf\u4e00\u7684\u63a5\u53e3\uff0c\u4ece\u800c\u53ef\u4ee5\u5c06\u4efb\u610f\u56fe\u795e\u7ecf\u7f51\u7edc\u8fdb\u884c\u5c01\u88c5\u6210\u6df1\u5c42\u7684\u56fe\u795e\u7ecf\u7f51\u7edc\uff0c\u5177\u4f53\u4ee3\u7801\u53ef\u4ee5\u53c2\u89c1\uff1a<\/p>\n

    from<\/strong> cogdl.layers import<\/strong> RevGNNLayer, ResGNNLayer
    \nfrom<\/strong> cogdl.layers import<\/strong> GCNLayer
    \nfrom<\/strong> cogdl.datasets import<\/strong> build_dataset_from_namedata = build_dataset_from_name(\"cora\").data<\/p>\n

    data = build_dataset_from_name(\"cora\").data<\/p>\n

    # \u901a\u8fc7CogDL\u4e2d\u7684\u56fe\u795e\u7ecf\u7f51\u7edc\u5c42GCNLayer\/ResGNNLayer\/RevGNNLayer\u6784\u5efa\u4e00\u4e2a\u6df1\u5c42GNN
    \n<\/em>class<\/strong> RevGCN(BaseModel):
    \ndef __init__<\/strong>(self, in_feats, out_feats, nhidden, nlayer):
    \nself.input_fc = nn.Linear(in_feats, nhidden)
    \nself.output_fc = nn.Lineare(nhidden, out_feats)<\/p>\n

    self.layers = nn.ModuleList()
    \nfor<\/strong> i in<\/strong> range(nlayer):
    \ngcn_layer = GCNLayer(data.num_features, data.num_classes)
    \n# \u5f97\u5230Pre-activation\u7684\u6b8b\u5dee\u8fde\u63a5GCN\u5c42
    \nresgcn_layer = ResGCNLayer(gcn_layer)
    \n# \u5f97\u5230Reversible\u7684GCN\u5c42
    \nrevgcn_layer = RevGNNLayer(resgcn_layer)
    \nself.layers.append(revgcn_layer) def forward(self, graph):<\/p>\n

    def forward<\/strong>(self, graph):
    \nx = graph.x
    \nh = self.input_fc(x)
    \nfor<\/strong> layer in<\/strong> self.layers:
    \nh = layer(graph, h)
    \nreturn<\/strong> self.output_fc(h)<\/p>\n

    deepgnn = RevGNN(data.num_features, data.num_classes, 64, nlayer=100)<\/code><\/section>\n

    <\/section>\n

    \u6211\u4eec\u76ee\u524d\u5728CogDL\u4e2d\u5b9e\u73b0\u4e86Deeerpgcn<\/em>\uff0c\u53ef\u9006\u7f51\u7edc\u5305\u62ec\u00a0revgcn<\/em>\u3001revgat<\/em>\u3001revgen<\/em>\u7b49\uff0c\u5f53\u7136\u7528\u6237\u4e5f\u53ef\u4ee5\u6839\u636e\u4e2a\u4eba\u9700\u6c42\u5229\u7528CogDL\u63d0\u4f9b\u7684\u63a5\u53e3\u4fbf\u6377\u5730\u5b9e\u73b0\u81ea\u5b9a\u4e49\u7684\u6a21\u578b\u3002<\/p>\n

    <\/section>\n
    \n
    \n
    \n
    \n
    \n
    \n
    \n

    03<\/strong><\/p>\n<\/section>\n<\/section>\n<\/section>\n

    \n
    <\/section>\n
    <\/section>\n<\/section>\n
    \n
    \n
    \u6df1\u5c42GNN\u8868\u73b0\u5982\u4f55\uff1f<\/strong><\/p>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n
    <\/section>\n

    GCNII\u5728Cora\/PPI\u4e0a\uff0c\u4ee5\u53caRevGEN\u5728ogbn-proteins\u7b49\u6570\u636e\u96c6\u4e0a\u90fd\u53d6\u5f97\u975e\u5e38\u4f18\u79c0\u751a\u81f3state-of-art\u7684\u6548\u679c\u3002<\/p>\n

    \u6211\u4eec\u5728Flickr\u3001Reddit\u7b49\u5e38\u7528\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u4e86\u6d4b\u8bd5\uff0c\u7ed3\u679c\u5982\u4e0b\u8868\u6240\u793a\uff0c\u53ef\u4ee5\u770b\u5230\uff0c\u901a\u8fc7\u589e\u52a0\u5c42\u6570\uff0c\u6a21\u578b\u53ef\u4ee5\u53d6\u5f97\u66f4\u4f18\u7684\u8868\u73b0\u3002\u8fd9\u4e9b\u8bf4\u660e\u6df1\u5c42\u56fe\u795e\u7ecf\u7f51\u7edc\u8fd8\u6709\u5f85\u8fdb\u884c\u66f4\u6df1\u5165\u7684\u7814\u7a76\uff0c\u4ee5\u53ca\u5f00\u53d1\u6df1\u5c42\u56fe\u795e\u7ecf\u7f51\u7edc\u7684\u6f5c\u529b\u3002<\/p>\n

    \"\u56fe\u7247\"<\/p>\n

    \u5173\u4e8e\u66f4\u591a\u6df1\u5c42\u56fe\u795e\u7ecf\u7f51\u7edc\u548c\u5176\u4ed6GNN\u6a21\u578b\uff0c\u5305\u62ec\u5728\u5f02\u6784\u56fe[9]\u3001\u56fe\u5bf9\u6297[12]\u3001\u63a8\u8350[13]\u7b49\u65b9\u9762\u7684\u6a21\u578b\u548c\u5e94\u7528\uff0c\u53ef\u4ee5\u53c2\u7167 CogDL\u4e2d\u7684\u5b9e\u73b0\u3002<\/p>\n

    CogDL\u00a0\u9879\u76ee\u4f20\u9001\u95e8:<\/p>\n

    https:\/\/github.com\/THUDM\/cogdl CogDL\u00a00.1.1\u00a0release<\/p>\n

    \n

    \u5185\u5bb9\u6765\u6e90\uff1a\u667a\u6e90\u793e\u533a<\/p>\n

    <\/div>\n

    <\/g><\/svg><\/i> \"Loading\"<\/p>\n

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

    \u6765\u6e90\uff1a\u667a\u6e90\u793e\u533a \u7248\u5f0f\u5236\u4f5c\uff1aCamel \u64b0\u6587\uff1a\u4faf\u632f\u5b87 \u7f16\u8f91\uff1a\u8d3e \u00a0 \u4f1f \u73b0\u5b9e\u4e16\u754c\u4e2d\u5f88\u591a\u5f88\u591a\u4efb\u52a1\u53ef\u4ee5\u63cf\u8ff0\u4e3a\u56fe(Gr […]<\/p>\n

    <\/div>\n

    <\/g><\/svg><\/i> \"Loading\"<\/p>\n

    <\/div>\n","protected":false},"author":1,"featured_media":404,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[3],"tags":[],"class_list":["post-396","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai"],"_links":{"self":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/396","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/comments?post=396"}],"version-history":[{"count":1,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/396\/revisions"}],"predecessor-version":[{"id":412,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/396\/revisions\/412"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media\/404"}],"wp:attachment":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media?parent=396"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/categories?post=396"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/tags?post=396"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}