{"id":38965,"date":"2019-01-23T00:00:06","date_gmt":"2019-01-22T16:00:06","guid":{"rendered":"https:\/\/case.ntu.edu.tw\/blog\/?p=38965"},"modified":"2022-01-04T08:40:53","modified_gmt":"2022-01-04T00:40:53","slug":"%e7%9c%9f%e5%81%bd%e8%8e%ab%e8%be%a8%e7%9a%84%e4%ba%ba%e5%83%8f%e7%94%a2%e7%94%9f%e5%99%a8","status":"publish","type":"post","link":"https:\/\/case.ntu.edu.tw\/blog\/?p=38965","title":{"rendered":"\u771f\u507d\u83ab\u8fa8\u7684\u4eba\u50cf\u7522\u751f\u5668"},"content":{"rendered":"
\n
\"\u771f\u507d\u83ab\u8fa8\u7684\u4eba\u50cf\u7522\u751f\u5668\"<\/div>\n<\/div>\n
\n
\n

\u7de8\u8b6f\uff0f\u845b\u7ad1\u5fd7<\/strong><\/p>\n

\u4f60\u76f8\u4fe1\u4e0a\u5716\u4e2d\u6240\u6709\u4eba\u7269\u90fd\u662f\u5047\u7684\u55ce\uff1f\u900f\u904e\u5c0d\u6297\u5f0f\u751f\u6210\u7db2\u8def\uff08Generative Adversarial Network\uff0cGAN\uff09\u8207\u65b0\u8208\u6280\u8853\uff0c\u5c31\u80fd\u6191\u7a7a\u7522\u751f\u4e00\u4e9b\u4e0d\u5b58\u5728\u7684\u4eba\u4e8b\u7269\uff0c\u751a\u81f3\u53ef\u6309\u7167\u4f7f\u7528\u8005\u9700\u6c42\uff0c\u8a02\u88fd\u51fa\u6e05\u6670\u4e14\u903c\u771f\u7684\u5f71\u50cf\u3002\u81ea2014\u5e74Ian Goodfellow\u9996\u6b21\u63d0\u51fa\u5f8c\uff0cGAN\u958b\u59cb\u5728\u5404\u5927\u9802\u7d1a\u7814\u8a0e\u6703\u4e0a\u4f54\u6709\u4e00\u5e2d\u4e4b\u5730\u3002\u622a\u81f3\u4eca\u65e5\u70ba\u6b62\uff0c\u5df2\u6709\u8a31\u591a\u9032\u5316\u7248\u672c\u88ab\u63d0\u51fa\u4f86\u3002\u4f46\u5230\u5e95\u70ba\u4f55\u6703\u5982\u6b64\u53d7\u5230\u77da\u76ee\u5462\uff1f<\/p>\n

\u8fd1\u5341\u5e74\u4f86\uff0c\u96a8GPU\u53ca\u5e73\u884c\u904b\u7b97\u7684\u80fd\u529b\u5927\u5e45\u63d0\u5347\uff0c\u795e\u7d93\u7db2\u8def\u5df2\u671d\u5411\u6df1\u5c64\u5316\u8207\u898f\u6a21\u5316\u7684\u65b9\u5411\u767c\u5c55\uff0c\u4f46\u4e5f\u9010\u6f38\u9047\u5230\u4e00\u9805\u56b4\u5cfb\u6311\u6230\uff1a\u5177\u6a19\u8a18\u7684\u8cc7\u6599\u96c6\u53ef\u9047\u4e0d\u53ef\u6c42\u3002\u82e5\u7121\u96c4\u539a\u8cc7\u672c\u8207\u8a08\u7b97\u80fd\u529b\u7684\u52a0\u6301\uff0c\u6240\u8a13\u7df4\u51fa\u7684AI\u6a21\u578b\u8868\u73fe\u6050\u4e0d\u5982\u9810\u671f\u3002\u56e0\u6b64\uff0c\u8207\u5176\u4e00\u5f35\u5f35\u5716\u7247\u8ac4\u8ac4\u5584\u8a98\u6559\u5c0e\u6a5f\u5668\u5982\u4f55\u8fa8\u5225\uff0c\u8a31\u591a\u96fb\u8166\u79d1\u5b78\u5bb6\u66f4\u671f\u671bAI\u53ef\u4ee5\u5b8c\u5168\u812b\u96e2\u4eba\u5de5\u6a19\u8a18\u7684\u8f14\u52a9\u3001\u81ea\u4e3b\u5b78\u7fd2\u3002\u9019\u7a2e\u4e0d\u501a\u8cf4\u4eba\u70ba\u5e72\u6d89\u7684\u5b78\u7fd2\u65b9\u6cd5\u7a31\u4f5c\u300c\u975e\u76e3\u7763\u5f0f\u5b78\u7fd2\u300d\uff08Unsupervised learning\uff09\uff0cGAN\u6b63\u662f\u6b64\u7a2e\u6846\u67b6\u4e0b\u8868\u73fe\u76f8\u7576\u7a81\u51fa\u7684\u6a5f\u5668\u5b78\u7fd2\u6a21\u578b\u3002<\/p>\n

\"\"<\/p>\n

\u5716\u4e00\u3001\u8fd1\u5e74\u4f86\u8207GAN\u76f8\u95dc\u7684\u8ad6\u6587\u767c\u8868\u6578\u91cf\u6025\u907d\u6500\u5347\u3002(\u4f86\u6e90\uff1aThe GAN Zoo)<\/p>\n<\/div>\n

 <\/p>\n

Goodfellow\u7576\u5e74\u5728\u601d\u8003\u5982\u4f55\u4f7f\u7528\u6a5f\u5668\u7522\u751f\u5716\u7247\u6642\uff0c\u4fbf\u63cf\u7e6a\u4e86\u5169\u500b\u76f8\u4e92\u7368\u7acb\u7684\u795e\u7d93\u7db2\u8def\uff1aGenerator\uff08\u751f\u6210\u5668\uff09\u8207Discriminator\uff08\u9451\u5225\u5668\uff09\u3002\u524d\u8005\u5728\u7d66\u5b9a\u7684\u96a8\u6a5f\u5411\u91cf\uff08Latent vector\uff09\u4e0b\uff0c\u900f\u904e\u647a\u7a4d\u5c64\uff08Convolution layer\uff09\u7522\u751f\u5716\u7247\uff1b\u5f8c\u8005\u5247\u5224\u65b7\u8a72\u5716\u7247\u7684\u771f\u507d\u3002\u5728\u5169\u500b\u7db2\u8def\u4f60\u4f86\u6211\u5f80\u3001\u76f8\u4e92\u5207\u78cb\u7684\u904e\u7a0b\u4e2d\uff0cGenerator\u6240\u7522\u751f\u7684\u5716\u7247\u4e5f\u8d8a\u4f86\u8d8a\u903c\u771f\u3002<\/p>\n

\u63d0\u5347\u89e3\u6790\u5ea6<\/strong><\/p>\n

\u8d77\u521d\uff0cGAN\u6240\u751f\u6210\u7684\u5716\u7247\u5927\u5c0f\u4e3b\u8981\u9084\u662f\u572864\u00d764\u4ee5\u4e0b\uff0c\u904e\u5927\u5247\u5bb9\u6613\u5931\u771f\uff0c\u76f4\u5230\u570b\u969b\u77e5\u540dGPU\u88fd\u9020\u5546Nvidia\u7684\u7814\u7a76\u5718\u968a\u63d0\u51fa\u6f38\u9032\u5f0f\u5b78\u7fd2\u65b9\u6cd5\uff08Progressive GAN\uff09[4]\u3002\u4e00\u958b\u59cb\uff0c\u4ed6\u5011\u5c07Generator\u548cDiscriminator\u7684\u5c64\u9650\u5236\u5728\u76f8\u7576\u5c0f\u7684\u7dad\u5ea6\uff0c\u59824\u00d74\uff1b\u540c\u6642\uff0c\u9001\u5230Discriminator\u7684\u771f\u5be6\u5716\u7247\u4e5f\u9700\u540c\u6b65\u964d\u4f4e\u53d6\u6a23\u5927\u5c0f\u81f34\u00d74\u3002\u5f85\u8a13\u7df4\u4e00\u6bb5\u6642\u9593\u5f8c\uff0c\u5728\u5169\u500b\u795e\u7d93\u7db2\u8def\u5404\u758a\u4e00\u5c64\u500d\u589e\u7684\u7dad\u5ea6\uff0c\u59828\u00d78\uff0c\u4e26\u7e7c\u7e8c\u4ee5\u540c\u7dad\u5ea6\u7684\u5716\u7247\u8a13\u7df4\u3002\u9031\u800c\u5fa9\u59cb\u76f4\u81f31024\u00d71024\u6642\uff0c\u5c31\u80fd\u7522\u751f\u9ad8\u756b\u8cea\u3001\u4e0d\u5bb9\u6613\u5931\u771f\u7684\u5716\u7247\u4e86\u3002<\/p>\n

\"\"<\/a><\/p>\n

\u5716\u4e8c\u3001Progressive GAN\u7684\u8a13\u7df4\u67b6\u69cb\u3002(\u4f86\u6e90\uff1aT. Karras et al., 2018.)<\/p>\n<\/div>\n

\u7d30\u5bdfGenerator\u7684\u67b6\u69cb\uff0c\u767c\u73fe\u8d8a\u5e95\u5c64\u7684\u795e\u7d93\u5143\u63a7\u5236\u8457\u4eba\u81c9\u8f03\u7c97\u7cd9\u7684\u7279\u5fb5\uff0c\u5982\u81c9\u578b\u3001\u819a\u8272\u53ca\u9aee\u578b\uff1b\u53cd\u4e4b\uff0c\u8d8a\u4e0a\u5c64\u5247\u8655\u7406\u8d8a\u7d30\u90e8\u7684\u7279\u5fb5\uff0c\u5982\u8868\u60c5\u3001\u773c\u775b\u7684\u9589\u5408\u3001\u9aee\u8272\u5230\u819a\u8cea\u7b49\u7d30\u7bc0\u3002\u6700\u8fd1\uff0c\u540c\u4e00\u5718\u968a\u4e5f\u91cb\u51fa\u5347\u7d1a\u7248\u672c\u2500\u4ee5\u98a8\u683c\u8f49\u63db\u70ba\u57fa\u790e\u7684GAN\uff08Style-based GAN\uff09\u3002<\/p>\n

\u98a8\u683c\u8f49\u63db<\/strong><\/p>\n

\u76f8\u4fe1\u4f60\u4e00\u5b9a\u4e0d\u964c\u751f\u90a3\u4e9b\u80fd\u9810\u6e2c\u4f60\u4e0a\u4e86\u5e74\u7d00\u6216\u8f49\u63db\u6027\u5225\u5f8c\u6703\u9577\u600e\u6a23\u7684\u61c9\u7528\u7a0b\u5f0f\u3002\u7406\u8ad6\u4e0a\uff0c\u85c9\u7531\u8abf\u6574Latent vector\u7684\u67d0\u4e9b\u5143\u7d20\uff0c\u5c31\u80fd\u5de6\u53f3\u5716\u4e2d\u4eba\u7269\u7684\u6027\u5225\u3001\u9aee\u9577\u6216\u819a\u8272\u7b49\uff1b\u7136\u800c\uff0c\u5be6\u969b\u64cd\u4f5c\u53ef\u80fd\u6bd4\u4f60\u60f3\u50cf\u4e2d\u9084\u8981\u4f86\u5f97\u56f0\u96e3\u3002<\/p>\n

\u4f8b\u5982\uff0c\u82e5\u539f\u5148\u7684\u8a13\u7df4\u8cc7\u6599\u96c6\u6709\u6240\u504f\u8aa4\uff08biased\uff09\uff0c\u4f8b\u5982\u4ee5\u4e9e\u6d32\u4eba\u7684\u5716\u50cf\u70ba\u5927\u5b97\u4e86\u8a71\uff0c\u5247\u4efb\u6191\u7d66\u5b9a\u7684Latent vector\u591a\u9ebd\u5730\u4e0d\u540c\uff0cGenerator\u6700\u5f8c\u7684\u8f38\u51fa\u7d50\u679c\u5c31\u53ea\u662f\u7a0d\u6709\u4e0d\u540c\u7684\u9ec3\u819a\u8272\u4eba\u81c9\u3002\u4e3b\u8981\u539f\u56e0\u5728\u65bc\u8a13\u7df4\u8cc7\u6599\u7684\u8b8a\u7570\u5ea6\u904e\u5c0f\uff0c\u5c0e\u81f4Generator\u7684\u8f38\u51fa\u7a7a\u9593\u4e0d\u5920\u5927\uff0c\u5373\u4f7f\u518d\u600e\u9ebc\u8abf\u6574latent vector\u90fd\u7121\u6cd5\u8d85\u51fa\u539f\u5148\u8a13\u7df4\u7684\u7bc4\u7587\u3002<\/p>\n

\u6b64\u5916\uff0c\u7279\u5fb5\u7cfe\u7e8f\uff08Features entanglement\uff09\u73fe\u8c61\u4e5f\u4ee4\u4eba\u96e3\u4ee5\u91d0\u6e05Latent vector\u4e2d\u7684\u54ea\u4e9b\u6210\u4efd\u7a76\u7adf\u63a7\u5236\u8457\u4eba\u81c9\u7684\u54ea\u500b\u7279\u5fb5\u3002\u4f7f\u5f97\u4efb\u4f55\u6578\u503c\u7684\u8abf\u6574\uff0c\u90fd\u662f\u727d\u4e00\u9aee\u800c\u52d5\u5168\u8eab\uff0c\u7a0d\u6709\u4e0d\u614e\uff0c\u4fbf\u662f\u9762\u76ee\u5168\u975e\u3002\u70ba\u6b64\uff0c\u65b0\u63d0\u51fa\u7684\u67b6\u69cb\u4e3b\u8981\u63a1\u7528\u4e86\u4ee5\u4e0b\u4e09\u7a2e\u6838\u5fc3\u6280\u8853\uff1a<\/p>\n

\"\"<\/p>\n

\u5716\u4e09\u3001Style-based GAN\u7684\u67b6\u69cb\u3002(\u4f86\u6e90\uff1aT. Karras et al., 2018.)<\/p>\n<\/div>\n

\u7b2c\u4e00\uff0c\u6709\u5225\u65bc\u50b3\u7d71GAN\u76f4\u63a5\u5c07latent vector\u4f5c\u70ba\u5716\u7247\u5f62\u6210\u7684\u57fa\u5e95\uff0cKarras\u7b49\u4eba\u5247\u5c07\u57fa\u5e95\u56fa\u5b9a\u4f4f\uff08\u5716\u4e094x4x512\u7684\u90e8\u5206\uff09\uff0c\u800c\u5f8c\u501f\u7528\u98a8\u683c\u8f49\u63db\uff08Style transfer\uff09\u7684\u6982\u5ff5\uff0c\u5728\u5404\u500b\u89e3\u6790\u5ea6\u7684\u5c64\u4e0a\uff0c\u9010\u6b65\u70ba\u6240\u8b02\u7684\u300c\u5167\u5bb9\u300d\u52a0\u4e0a\u4f86\u81eaLatent vector\u7684\u300c\u98a8\u683c\u300d\uff0c\u6700\u7d42\u7d44\u5408\u6210\u4e00\u5f35\u756b\u8cea\u76f8\u7576\u9ad8\u7684\u5716\u7247\u3002<\/p>\n

\u7b2c\u4e8c\uff0cLatent vector\u6703\u5148\u901a\u904e\u6578\u500b\u5168\u9023\u7d50\u5c64\uff08Fully connected layer\uff09\u6240\u69cb\u6210\u7684\u7db2\u8def\u3002\u5728\u5718\u968a\u7684\u5047\u8a2d\u4e2d\uff0c\u82e5\u63a1\u50b3\u7d71GAN\u7684\u67b6\u69cb\uff0cLatent vector\u6703\u5728Z\u7a7a\u9593\u4e2d\u6975\u76e1\u6240\u80fd\u5730\u8fd1\u4f3c\u8a13\u7df4\u5716\u7247\u7684\u5206\u4f48\uff0c\u5c07\u76f4\u63a5\u5c0e\u81f4\u7cfe\u7e8f\u73fe\u8c61\uff08Entanglement\uff09\u7684\u767c\u751f\u3002\u70ba\u6b64\uff0c\u4ed6\u5011\u4f7fLatent vector\u900f\u904e\u4e00\u7a2e\u975e\u7dda\u6027\u8f49\u63db\uff08\u5f9eZ\u7a7a\u9593\u5230W\u7a7a\u9593\uff09\u6210w\u5411\u91cf\u5f8c\uff0c\u518d\u5957\u7528\u5230\u300c\u5167\u5bb9\u300d\u8eab\u4e0a\u3002\u900f\u904e\u9019\u500b\u984d\u5916\u7684\u795e\u7d93\u7db2\u8def\uff0c\u64f4\u5c55\u7684Latent vector\u7684\u8f49\u63db\u7a7a\u9593\uff0c\u4f7f\u5f97w\u5411\u91cf\u4e0d\u518d\u53d7\u9650\u65bc\u8a13\u7df4\u8cc7\u6599\u7684\u6a5f\u7387\u5206\u4f48\uff0c\u540c\u6642\u80fd\u964d\u4f4e\u4e0d\u540c\u7279\u5fb5\u4e4b\u9593\u7684\u76f8\u95dc\u6027\uff0c\u9054\u5230\u7279\u5fb5\u89e3\u96e2\uff08Feature disentanglement\uff09\u7684\u76ee\u7684[3]\u3002<\/p>\n

\u7b2c\u4e09\uff0c\u5728\u5404\u500b\u89e3\u6790\u5ea6\u5c64\u4e2d\u52a0\u5165\u300c\u98a8\u683c\u300d\u8207\u300c\u96a8\u6a5f\u6210\u5206\u300d\u3002\u5728\u300c\u98a8\u683c\u300d\u4e0a\uff0c\u4ed6\u5011\u4f7f\u7528\u4eff\u5c04\u8b8a\u63db\uff08Affine transformation\uff0c\u5982\u5716\u4e09a\uff09\uff0c\u5c07w\u5411\u91cf\u8f49\u63db\u70baAdaIN\uff08Adaptive Instance Normalization\uff09\u6240\u9700\u7684\u7e2e\u653e\u8207\u504f\u8aa4\u53c3\u6578\u3002\u53e6\u4e00\u65b9\u9762\uff0c\u70ba\u4e86\u4eff\u771f\u4eba\u81c9\u4e2d\u5982\u982d\u9aee\u3001\u7d30\u7d0b\u7b49\u8f03\u8907\u96dc\u7684\u7d30\u7dfb\u7279\u5fb5\uff0c\u6bcf\u500b\u5c64\u90fd\u6703\u984d\u5916\u52a0\u4e0a\u7d93\u904e\u7e2e\u653e\uff08\u5982\u5716\u4e09b\uff09\u7684\u96dc\u8a0a\uff0c\u4f5c\u70ba\u96a8\u6a5f\u6210\u5206\u7684\u4f86\u6e90\u3002<\/p>\n

\"\"<\/p>\n

\u5716\u56db\u3001\u52a0\u5165\u300c\u98a8\u683c\u300d\u7684\u6d41\u7a0b\u8207AdaIN\u7684\u8f49\u63db\u3002AdaIN\u662f\u98a8\u683c\u8f49\u63db\u4e2d\u5c07\u300c\u5167\u5bb9\u300d\u6b63\u5247\u5316\u70ba\u300c\u98a8\u683c\u300d\u5206\u5e03\u7684\u6280\u5de7\u3002\u5176\u4e2dy\u4f86\u81ea\u65bc\u300c\u98a8\u683c\u300d\u6210\u5206\uff0cx\u4f86\u81ea\u65bc\u300c\u5167\u5bb9\u300d\u6210\u5206\u3002(\u4f86\u6e90\uff1aR. Horev, 2018.)<\/p>\n<\/div>\n

 <\/p>\n

\"\"<\/p>\n

\u5716\u4e94\u3001\u5728\u4e0d\u540c\u5c64\u4e0a\u5408\u6210\u4e0d\u540c\u4eba\u81c9\u7279\u5fb5\u3002(\u4f86\u6e90\uff1a[7])<\/p>\n<\/div>\n

\u8207Progressive GAN\u76f8\u540c\uff0cStyle-based GAN\u8f03\u70ba\u5e95\u5c64\u7684Latent vector\u4ee3\u8868\u4eba\u81c9\u4e0a\u8f03\u7c97\u7565\u7684\u7279\u5fb5\uff0c\u4e2d\u5c64\u63a7\u5236\u8457\u8868\u60c5\u8207\u773c\u775b\uff0c\u6700\u4e0a\u5c64\u5247\u662f\u9aee\u8272\u8207\u81c9\u90e8\u7b49\u7d30\u7bc0\u3002Nvidia\u5718\u968a\u5229\u7528\u5de7\u601d\uff0c\u6539\u8b8aLatent vector\u7684\u8f38\u5165\u65b9\u5f0f\uff0c\u5f9e\u800c\u7531\u4e0d\u540c\u968e\u5c64\u7684\u8f38\u5165\u6539\u8b8a\u5716\u50cf\u98a8\u683c\uff0c\u8457\u5be6\u4ee4\u4eba\u671f\u5f85\u5f80\u5f8c\u7684\u6280\u8853\u7a81\u7834\u3002 <\/p>\n

\u7de8\u8b6f\u4f86\u6e90<\/strong><\/p>\n

Will Knight.\u00a0These incredibly realistic fake faces show how algorithms can now mess with us.<\/a>\u00a0MIT Technology Review. 2018.<\/p>\n

\u53c3\u8003\u8cc7\u6599<\/strong><\/p>\n

    \n
  1. R. Horev, \u201cExplained: A Style-Based Generator Architecture for GANs \u2013 Generating and Tuning Realistic Artificial Faces<\/a>\u201c,\u00a0Towards Data Science,<\/em>\u00a02018.<\/li>\n
  2. W. Knight, \u201cAI software can dream up an entire digital world from a simple sketch<\/a>\u201c,\u00a0MIT Technology Review<\/em>, 2018.<\/li>\n
  3. T. Karras, S. Laine, T. Aila, \u201cA Style-Based Generator Architecture for Generative Adversarial Networks<\/a>\u201c, Arxiv.org, 2018.<\/li>\n
  4. T. Karras, T. Aila, S. Laine, J. Lehtinen, \u201cProgressive Growing of GANs for Improved Quality, Stability, and Variation<\/a>\u201c, ICLR 2018.<\/li>\n
  5. X. Huang, S. Belongie, \u201cArbitrary Style Transfer in Real-time with Adaptive Instance Normalization<\/a>\u201c,\u00a0ICCV 2017.<\/li>\n
  6. A. Hindupur, \u201cThe GAN Zoo\u201d<\/a>, GitHub.<\/li>\n
  7. A Style-Based Generator Architecture for Generative Adversarial Networks<\/a>,\u00a0YouTube, 2018.<\/li>\n<\/ol>\n

    (\u672c\u6587\u7531\u6559\u80b2\u90e8\u88dc\u52a9\uff62AI\u5831\u5831\u2500AI\u79d1\u666e\u63a8\u5ee3\u8a08\u756b\u300d\u57f7\u884c\u5718\u968a\u7de8\u8b6f)<\/p>\n<\/div>\n<\/div>\n","protected":false},"excerpt":{"rendered":"

    \u7de8\u8b6f\uff0f\u845b\u7ad1\u5fd7 \u4f60\u76f8\u4fe1\u4e0a\u5716\u4e2d\u6240\u6709\u4eba\u7269\u90fd\u662f\u5047<\/p>\n","protected":false},"author":21,"featured_media":38966,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[3772,3293],"tags":[6108,6109,5434,6110,3935,3809],"aioseo_notices":[],"jetpack_featured_media_url":"https:\/\/case.ntu.edu.tw\/blog\/wp-content\/uploads\/2022\/01\/facegenerator-0.png","_links":{"self":[{"href":"https:\/\/case.ntu.edu.tw\/blog\/index.php?rest_route=\/wp\/v2\/posts\/38965"}],"collection":[{"href":"https:\/\/case.ntu.edu.tw\/blog\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/case.ntu.edu.tw\/blog\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/case.ntu.edu.tw\/blog\/index.php?rest_route=\/wp\/v2\/users\/21"}],"replies":[{"embeddable":true,"href":"https:\/\/case.ntu.edu.tw\/blog\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=38965"}],"version-history":[{"count":1,"href":"https:\/\/case.ntu.edu.tw\/blog\/index.php?rest_route=\/wp\/v2\/posts\/38965\/revisions"}],"predecessor-version":[{"id":38967,"href":"https:\/\/case.ntu.edu.tw\/blog\/index.php?rest_route=\/wp\/v2\/posts\/38965\/revisions\/38967"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/case.ntu.edu.tw\/blog\/index.php?rest_route=\/wp\/v2\/media\/38966"}],"wp:attachment":[{"href":"https:\/\/case.ntu.edu.tw\/blog\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=38965"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/case.ntu.edu.tw\/blog\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=38965"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/case.ntu.edu.tw\/blog\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=38965"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}