{"id":10899,"date":"2022-12-17T19:30:42","date_gmt":"2022-12-17T11:30:42","guid":{"rendered":"http:\/\/139.9.1.231\/?p=10899"},"modified":"2022-12-20T20:00:18","modified_gmt":"2022-12-20T12:00:18","slug":"vilt","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2022\/12\/17\/vilt\/","title":{"rendered":"\u591a\u6a21\u6001\u9884\u8bad\u7ec3 | ViLT"},"content":{"rendered":"\n

paper: https:\/\/arxiv.org\/abs\/2102.03334<\/a><\/strong> ICML 2021<\/em><\/strong><\/p>\n\n\n\n

code: https:\/\/github.com\/dandelin\/ViLT<\/a><\/strong> <\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n
\"\"
\u56fe1 Visual comparison of conventional VLP architectures
and our proposed ViLT.<\/em><\/strong><\/figcaption><\/figure>\n\n\n\n

\u89c6\u89c9\u6587\u672c\u591a\u6a21\u6001\u4efb\u52a1\uff0c\u6781\u5176\u7b80\u5355\u7684\u591a\u6a21\u6001\u7ed3\u6784\u3002\u6a21\u6001\u7684\u7279\u5f81\u62bd\u53d6\u505a\u5230\u4e86\u6781\u5c0f\u5316\uff0c\u4e3b\u8981\u7684\u8ba1\u7b97\u91cf\u653e\u5728\u540e\u8fb9\u7684\u6a21\u6001\u878d\u5408\u4e0a\uff0c\u63d0\u9ad8\u4e86\u63a8\u7406\u901f\u5ea6\u3002<\/strong>\u591a\u6a21\u6001\u9886\u57df\u91cc\u7a0b\u7891\u5f0f\u5de5\u4f5c\u3002\u5c06\u533a\u57df\u7279\u5f81\uff0cregion \u4ece\u591a\u6a21\u6001\u6846\u67b6\u4e2d\u79fb\u9664\u3002<\/p>\n\n\n\n

Vision and Language Pre-training(VLP)\u5df2\u7ecf\u5df2\u7ecf\u5728\u89c6\u89c9\u8bed\u8a00\u7684\u591a\u6a21\u6001\u4e0b\u6e38\u4efb\u52a1\u4e2d\u53d1\u5c55\u7684\u5f88\u597d\u3002\u7136\u800c\uff0c\u5f53\u524dVLP\u7684\u5de5\u4f5c\u4e3b\u8981\u96c6\u4e2d\u5728\u56fe\u50cf\u7279\u5f81\u62bd\u53d6\u4e0a\uff0c\u4e00\u822c\u6765\u8bb2\uff0c\u56fe\u50cf\u7279\u5f81\u62bd\u53d6\u7684\u8d8a\u597d\uff0c\u4e0b\u6e38\u4efb\u52a1\u4e2d\u7684\u8868\u73b0\u5c31\u8d8a\u597d\u3002\u4f46\u662f\uff0c\u73b0\u5728\u4e3b\u8981\u6709\u4e24\u4e2a\u95ee\u9898\uff0c\u4e00\u662f\u6548\u7387\u592a\u4f4e\uff0c\u901f\u5ea6\u592a\u6162\uff0c\u62bd\u53d6\u56fe\u50cf\u7279\u5f81\u82b1\u8d39\u5927\u91cf\u65f6\u95f4\uff0c\u6bd4\u591a\u6a21\u6001\u878d\u5408\u90fd\u591a\u3002\u6211\u4eec\u5e94\u8be5\u82b1\u8d39\u66f4\u591a\u65f6\u95f4\u5728\u878d\u5408\u4e0a\u3002\u7b2c\u4e8c\u4e2a\u662f\uff0c\u4f60\u7528\u4e00\u4e2a\u9884\u8bad\u7ec3\u597d\u7684\u6a21\u578b\u53bb\u62bd\u53d6\u7279\u5f81\uff0c\u8868\u8fbe\u80fd\u529b\u53d7\u9650\u3002\u76ee\u6807\u68c0\u6d4b\u6570\u636e\u96c6\u4e0d\u591f\u5927\uff0c\u89c4\u6a21\u4e0d\u591f\u5927\u3002\u5982\u679c\u6a21\u578b\u4e0d\u662f\u7aef\u5230\u7aef\u5b66\u4e60\uff0c\u53ea\u662f\u4ece\u9884\u8bad\u7ec3\u6a21\u578b\u62bd\u53d6\u7279\u5f81\uff0c\u5927\u6982\u7387\u6765\u8bf4\u4e0d\u662f\u6700\u4f18\u89e3\u3002<\/p>\n\n\n\n

Motivation<\/h3>\n\n\n\n

\u76ee\u524d\u53c2\u6570\u91cf\u6700\u5c0f\u7684\u591a\u6a21\u6001Transformer\u65b9\u6cd5\u3002ViLT\u4f7f\u7528\u9884\u8bad\u7ec3\u7684ViT\u6765\u521d\u59cb\u5316\u4ea4\u4e92\u7684transformer\uff0c\u8fd9\u6837\u5c31\u53ef\u4ee5\u76f4\u63a5\u5229\u7528\u4ea4\u4e92\u5c42\u6765\u5904\u7406\u89c6\u89c9\u7279\u5f81\uff0c\u4e0d\u9700\u8981\u989d\u5916\u589e\u52a0\u4e00\u4e2a\u89c6\u89c9encoder\uff08\u5982Faster-RCNN\uff09\u3002<\/p>\n\n\n\n

Contribution<\/h3>\n\n\n\n
  1. \u7b2c\u4e00\u4e2a\u57fa\u4e8epatch projection\u7684\u591a\u6a21\u6001\u9884\u8bad\u7ec3\u6a21\u578b\uff0c<\/strong>\u5176\u662f\u9996\u4e2a\u4f7f\u7528patch projection<\/strong>\u6765\u505avisual embedding<\/strong>\u7684\u65b9\u6cd5\u3002<\/li>
  2. \u8bc1\u660e\u4e86\u53ef\u4ee5\u5c06BERT\u7684\u65b9\u6cd5\u548cVison Transformer\u7ed3\u5408\u8d77\u6765\u7528\u4e8e\u591a\u6a21\u6001transformer<\/strong>\u3002<\/li>
  3. \u4f53\u73b0\u4e86\u5168\u8bcd\u63a9\u7801\u5728\u9884\u8bad\u7ec3\u65f6\u4ee5\u53ca\u56fe\u50cf\u589e\u5f3a\u5728\u5fae\u8c03\u65f6\u7684\u91cd\u8981\u6027\u3002<\/li><\/ol>\n\n\n\n

    Method<\/h3>\n\n\n\n

    \u73b0\u6709\u7684\u89c6\u89c9\u8bed\u8a00\u6a21\u578b\u7684\u4e09\u79cd\u7ed3\u6784\u7c7b\u522b\uff1a<\/p>\n\n\n\n

    VE = Vision Embedding<\/p>\n\n\n\n

    TE = Text Embedding<\/p>\n\n\n\n

    MI = Modality Interaction<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    \u4e0a\u56fe\u662f4\u79cd\u4e0d\u540c\u7c7b\u578b\u7684VLP\u6a21\u578b\u793a\u610f\u56fe\u3002\u5176\u4e2d\u6bcf\u4e2a\u77e9\u5f62\u7684\u9ad8\u8868\u793a\u76f8\u5bf9\u8ba1\u7b97\u91cf\u5927\u5c0f\uff0cVE\u3001TE\u548cMI\u5206\u522b\u662fvisual embedding\u3001text embedding\u548cmodality interaction\u7684\u7b80\u5199\u3002<\/p>\n\n\n\n

    \u4f5c\u8005\u63d0\u51fa\u8fd94\u79cd\u7c7b\u578b\u7684\u4e3b\u8981\u4f9d\u636e\u6709\u4e24\u70b9\uff1a<\/strong><\/p>\n\n\n\n

    1.\u5728\u53c2\u6570\u6216\u8005\u8ba1\u7b97\u4e0a\uff0c\u4e24\u79cd\u6a21\u6001\u662f\u5426\u4fdd\u6301\u5e73\u8861\u3002<\/p>\n\n\n\n

    2.\u5728\u7f51\u7edc\u6df1\u5c42\u4e2d\uff0c\u4e24\u79cd\u6a21\u6001\u662f\u5426\u76f8\u4e92\u4f5c\u7528\u3002<\/p>\n\n\n\n

    VSE\u3001VSE++\u548cSCAN\u5c5e\u4e8e(a)\u7c7b\u578b\u3002<\/strong>\u5bf9\u56fe\u50cf\u548c\u6587\u672c\u72ec\u7acb\u4f7f\u7528encoder\uff0c\u56fe\u50cf\u7684\u66f4\u91cd\uff0c\u6587\u672c\u7684\u66f4\u8f7b\uff0c\u4f7f\u7528\u7b80\u5355\u7684\u70b9\u79ef\u6216\u8005\u6d45\u5c42attention\u5c42\u6765\u8868\u793a\u4e24\u79cd\u6a21\u6001\u7279\u5f81\u7684\u76f8\u4f3c\u6027\u3002<\/p>\n\n\n\n

    CLIP\u5c5e\u4e8e(b)\u7c7b\u578b\u3002<\/strong>\u6bcf\u4e2a\u6a21\u6001\u5355\u72ec\u4f7f\u7528\u91cd\u7684transformer encoder\uff0c\u4f7f\u7528\u6c60\u5316\u540e\u7684\u56fe\u50cf\u7279\u5f81\u70b9\u79ef\u8ba1\u7b97\u7279\u5f81\u76f8\u4f3c\u6027\u3002<\/p>\n\n\n\n

    ViLBERT\u3001UNTER\u548cPixel-BERT\u5c5e\u4e8e(c)\u7c7b\u578b\u3002<\/strong>\u8fd9\u4e9b\u65b9\u6cd5\u4f7f\u7528\u6df1\u5c42transformer\u8fdb\u884c\u4ea4\u4e92\u4f5c\u7528\uff0c\u4f46\u662f\u7531\u4e8eVE\u4ecd\u7136\u4f7f\u7528\u91cd\u7684\u5377\u79ef\u7f51\u7edc\u8fdb\u884c\u7279\u5f81\u62bd\u53d6\uff0c\u5bfc\u81f4\u8ba1\u7b97\u91cf\u4f9d\u7136\u5f88\u5927\u3002<\/p>\n\n\n\n

    \u4f5c\u8005\u63d0\u51fa\u7684ViLT\u5c5e\u4e8e(d)\u7c7b\u578b\u3002<\/strong>ViLT\u662f\u9996\u4e2a\u5c06VE\u8bbe\u8ba1\u7684\u5982TE\u4e00\u6837\u8f7b\u91cf\u7684\u65b9\u6cd5\uff0c\u8be5\u65b9\u6cd5\u7684\u4e3b\u8981\u8ba1\u7b97\u91cf\u90fd\u96c6\u4e2d\u5728\u6a21\u6001\u4ea4\u4e92\u4e0a\u3002<\/strong><\/em><\/p>\n\n\n\n

    Modality Interaction Schema<\/strong><\/h3>\n\n\n\n

    \u6a21\u6001\u4ea4\u4e92\u90e8\u5206\u53ef\u4ee5\u5206\u6210\u4e24\u79cd\u65b9\u5f0f\uff1a\u4e00\u79cd\u662fsingle-stream<\/strong>(\u5982BERT\u548cUNITER)\uff0c\u53e6\u4e00\u79cd\u662fdual-stream<\/strong>(\u5982ViLBERT\u548cLXMERT)\u3002\u5176\u4e2dsingle-stream\u662f\u5bf9\u56fe\u50cf\u548c\u6587\u672cconcate\u7136\u540e\u8fdb\u884c\u4ea4\u4e92\u64cd\u4f5c\uff0c\u800cdual-stream\u662f\u4e0d\u5bf9\u56fe\u50cf\u548c\u6587\u672cconcate\u7136\u540e\u8fdb\u884c\u4ea4\u4e92\u64cd\u4f5c\u3002ViLT\u5ef6\u7528single-stream\u7684\u4ea4\u4e92\u65b9\u5f0f\uff0c\u56e0\u4e3adual-stream\u4f1a\u5f15\u5165\u989d\u5916\u7684\u8ba1\u7b97\u91cf\u3002<\/p>\n\n\n\n

    \u73b0\u6709\u7684VLP\u6a21\u578b\u7684text embedding\u57fa\u672c\u4e0a\u90fd\u4f7f\u7528\u7c7bBERT\u7ed3\u6784(\u56fe1)\uff0c\u4f46\u662fvisual embedding\u5b58\u5728\u7740\u5dee\u5f02\u3002\u5728\u5927\u591a\u6570\u60c5\u51b5\u4e0b\uff0cvisual embedding\u662f\u73b0\u6709VLP\u6a21\u578b\u7684\u74f6\u9888\u3002visual embedding\u7684\u65b9\u6cd5\u603b\u5171\u6709\u4e09\u5927\u7c7b\uff0c\u5176\u4e2dregion feature<\/strong>\u65b9\u6cd5\u901a\u5e38\u91c7\u7528Faster R-CNN\u4e8c\u9636\u6bb5\u68c0\u6d4b\u5668\u63d0\u53d6region\u7684\u7279\u5f81\uff0cgrid feature<\/strong>\u65b9\u6cd5\u76f4\u63a5\u4f7f\u7528CNN\u63d0\u53d6grid\u7684\u7279\u5f81\uff0cpatch projection<\/strong>\u65b9\u6cd5\u5c06\u8f93\u5165\u56fe\u7247\u5207\u7247\u6295\u5f71\u63d0\u53d6\u7279\u5f81\u3002ViLT\u662f\u9996\u4e2a\u4f7f\u7528patch projection\u6765\u505avisual embedding\u7684\u65b9\u6cd5\u3002<\/p>\n\n\n\n

    \u7f51\u7edc\u7ed3\u6784ViLT<\/strong><\/h3>\n\n\n\n
    \"\"<\/figure>\n\n\n\n

    \u4f5c\u8005\u63d0\u51fa\u7684ViLT\u53ef\u4ee5\u8ba4\u4e3a\u662f\u76ee\u524d\u6700\u7b80\u5355\u7684\u591a\u6a21\u6001Transformer\u65b9\u6cd5\u3002ViLT\u4f7f\u7528\u9884\u8bad\u7ec3\u7684ViT\u6765\u521d\u59cb\u5316\u4ea4\u4e92\u7684transformer\uff0c\u8fd9\u6837\u5c31\u53ef\u4ee5\u76f4\u63a5\u5229\u7528\u4ea4\u4e92\u5c42\u6765\u5904\u7406\u89c6\u89c9\u7279\u5f81\uff0c\u4e0d\u9700\u8981\u989d\u5916\u589e\u52a0\u4e00\u4e2a\u89c6\u89c9encoder\u3002<\/p>\n\n\n\n

    \u6587\u672c\u7279\u5f81\u8f93\u5165\u90e8\u5206\uff0c\u5c06\u6587\u672c\u770b\u6210\u4e00\u4e2a\u8bcd\u5e8f\u5217\uff0c\u901a\u8fc7word embedding matrix\u8f6c\u5316\u6210word embedding\uff0c\u7136\u540e\u548cposition embedding\u8fdb\u884c\u76f8\u52a0\uff0c\u6700\u540e\u548cmodal-type embedding\u8fdb\u884cconcate\u3002<\/p>\n\n\n\n

    \u56fe\u50cf\u7279\u5f81\u8f93\u5165\u90e8\u5206\uff0c\u5c06\u56fe\u50cf\u5207\u5757\u770b\u6210\u4e00\u4e2a\u56fe\u50cf\u5757\u5e8f\u5217\uff0c\u901a\u8fc7linear projection\u8f6c\u5316\u6210visual embedding\uff0c\u7136\u540e\u548cpostion embedding\u8fdb\u884c\u76f8\u52a0\uff0c\u6700\u540e\u548cmodal-type embedding\u8fdb\u884cconcate\u3002<\/p>\n\n\n\n

    \u5176\u4e2dword embedding\u548cvisual embedding\u901a\u8fc7\u53ef\u5b66\u4e60\u7684modal-type embedding\u6807\u5fd7\u4f4d\u6765\u533a\u5206\uff0c\u5176\u4e2d0\u6807\u5fd7\u4f4d\u8868\u793aword embedding\u90e8\u5206\uff0c1\u6807\u5fd7\u4f4d\u8868\u793avisual embedding\u90e8\u5206\u3002<\/p>\n\n\n\n

    wrod embedding\u548cvisual embedding\u5206\u522b\u90fd\u5d4c\u5165\u4e86\u4e00\u4e2a\u989d\u5916\u7684\u53ef\u5b66\u4e60[class] embedding\uff0c\u65b9\u4fbf\u548c\u4e0b\u6e38\u4efb\u52a1\u5bf9\u63a5\u3002<\/p>\n\n\n\n

    Pretraining Objectives<\/strong><\/p>\n\n\n\n

    ViLT\u9884\u8bad\u7ec3\u7684\u4f18\u5316\u76ee\u6807\u6709\u4e24\u4e2a\uff1a\u4e00\u4e2a\u662fimage text matching(ITM)<\/strong>\uff0c\u53e6\u4e00\u4e2a\u662fmasked language modeling(MLM)<\/strong>\u3002<\/p>\n\n\n\n

    ImageText Matching<\/strong>\uff1a\u968f\u673a\u4ee50.5\u7684\u6982\u7387\u5c06\u6587\u672c\u5bf9\u5e94\u7684\u56fe\u7247\u66ff\u6362\u6210\u4e0d\u540c\u7684\u56fe\u7247\uff0c\u7136\u540e\u5bf9\u6587\u672c\u6807\u5fd7\u4f4d\u5bf9\u5e94\u8f93\u51fa\u4f7f\u7528\u4e00\u4e2a\u7ebf\u6027\u7684ITM head\u5c06\u8f93\u51fafeature\u6620\u5c04\u6210\u4e00\u4e2a\u4e8c\u503clogits\uff0c\u7528\u6765\u5224\u65ad\u56fe\u50cf\u6587\u672c\u662f\u5426\u5339\u914d\u3002\u53e6\u5916ViLT\u8fd8\u8bbe\u8ba1\u4e86\u4e00\u4e2aword patch alignment (WPA)\u6765\u8ba1\u7b97teextual subset\u548cvisual subset\u7684\u5bf9\u9f50\u5206\u6570\u3002<\/p>\n\n\n\n

    Masked Language Modeling<\/strong>\uff1aMLM\u7684\u76ee\u6807\u662f\u901a\u8fc7\u6587\u672c\u7684\u4e0a\u4e0b\u6587\u4fe1\u606f\u53bb\u9884\u6d4bmasked\u7684\u6587\u672ctokens\u3002\u968f\u673a\u4ee50.15\u7684\u6982\u7387mask\u6389tokens\uff0c\u7136\u540e\u6587\u672c\u8f93\u51fa\u63a5\u4e24\u5c42MLP\u4e0e\u8f66mask\u6389\u7684tokens\u3002<\/p>\n\n\n\n

    Whole Word Masking<\/strong>\uff1a\u53e6\u5916ViLT\u8fd8\u4f7f\u7528\u4e86whole word masking\u6280\u5de7\u3002whole word masking\u662f\u5c06\u8fde\u7eed\u7684\u5b50\u8bcdtokens\u8fdb\u884cmask\u7684\u6280\u5de7\uff0c\u907f\u514d\u4e86\u53ea\u901a\u8fc7\u5355\u8bcd\u4e0a\u4e0b\u6587\u8fdb\u884c\u9884\u6d4b\u3002\u6bd4\u5982\u5c06\u201cgiraffe\u201d\u8bcdtokenized\u62103\u4e2a\u90e8\u5206[“gi”, “##raf”, “##fe”]\uff0c\u53ef\u4ee5mask\u6210[“gi”, “[MASK]”, “##fe”]\uff0c\u6a21\u578b\u4f1a\u901a\u8fc7mask\u7684\u4e0a\u4e0b\u6587\u4fe1\u606f[\u201cgi\u201d\uff0c\u201c##fe\u201d]\u6765\u9884\u6d4bmask\u7684\u201c##raf\u201d\uff0c\u5c31\u4f1a\u5bfc\u81f4\u4e0d\u5229\u7528\u56fe\u50cf\u4fe1\u606f\u3002<\/p>\n\n\n\n

    Experiment<\/h3>\n\n\n\n

    \u672c\u6587\u63d0\u51fa\u7684\u65b9\u6cd5\u5728\u6548\u7387\u4e0a\u5927\u5927\u63d0\u5347\u4e14\u8868\u73b0\u51fa\u76f8\u4f3c\u7684\u6027\u80fd\uff0c\u76f8\u6bd4\u4e8eregion feature\u7684\u65b9\u6cd5\u901f\u5ea6\u5feb\u4e8660\u500d\uff0c\u76f8\u6bd4\u4e8egrid feature\u7684\u65b9\u6cd5\u5feb\u4e864\u500d\uff0c\u800c\u4e14\u4e0b\u6e38\u4efb\u52a1\u8868\u73b0\u51fa\u76f8\u4f3c\u751a\u81f3\u66f4\u597d\u7684\u6027\u80fd\u3002<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    \u5982\u56fe\u6240\u793a\uff0cViLT\u76f8\u6bd4\u4e8eregion feature\u7684\u65b9\u6cd5\u901f\u5ea6\u5feb\u4e8660\u500d\uff0c\u76f8\u6bd4\u4e8egrid feature\u7684\u65b9\u6cd5\u5feb\u4e864\u500d\uff0c\u800c\u4e14\u4e0b\u6e38\u4efb\u52a1\u8868\u73b0\u51fa\u76f8\u4f3c\u751a\u81f3\u66f4\u597d\u7684\u6027\u80fd\u3002<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n
    \"\"<\/figure>\n\n\n\n
    \"\"<\/figure>\n\n\n\n
    \"\"<\/figure>\n\n\n\n
    \"\"<\/figure>\n\n\n\n
    \"\"<\/figure>\n\n\n\n

    \u7f3a\u70b9\uff1a<\/h2>\n\n\n\n

    1\u3001\u6027\u80fd\u4e0d\u591f\u9ad8\uff0c\u5728\u4e00\u4e9b\u6570\u636e\u96c6\u4e0a\u5f97\u8868\u73b0\u6bd4\u4e0d\u8fc7C\u7c7b\u65b9\u6cd5\uff0c\u6709\u53ef\u80fd\u56e0\u4e3a\u5bf9\u4e8e\u73b0\u6709\u7684\u4efb\u52a1\u6765\u8bf4\uff0c\u56e0\u4e3a\u6570\u636e\u96c6\u7684bias\uff0c\u6216\u8005\u8fd9\u4e2a\u4efb\u52a1\u9700\u8981\u66f4\u591a\u7684\u89c6\u89c9\u4fe1\u606f\uff0c\u56e0\u6b64\u9700\u8981\u66f4\u591a\u5f97\u89c6\u89c9\u90e8\u5206\uff0c\u6700\u540e\u7684\u6548\u679c\u624d\u80fd\u597d\u3002<\/p>\n\n\n\n

    2\u3001\u867d\u7136\u63a8\u7406\u65f6\u95f4\u5feb\uff0c\u4f46\u662f\u8bad\u7ec3\u901f\u5ea6\u5f88\u6162\u3002\u53ea\u662f\u7ed3\u6784\u4e0a\u7b80\u5316\u4e86\u591a\u6a21\u6001\u5b66\u4e60\uff0c\u4f46\u4e00\u822c\u4eba\u8fd8\u662f\u73a9\u4e0d\u8d77\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

    paper: https:\/\/arxiv.org\/abs\/2102.03334 ICML 2021 code: … \u7ee7\u7eed\u9605\u8bfb\u591a\u6a21\u6001\u9884\u8bad\u7ec3 | ViLT<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[21,4,9],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/10899"}],"collection":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/comments?post=10899"}],"version-history":[{"count":39,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/10899\/revisions"}],"predecessor-version":[{"id":10994,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/10899\/revisions\/10994"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=10899"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=10899"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=10899"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}