{"id":752,"date":"2024-08-21T08:00:37","date_gmt":"2024-08-21T00:00:37","guid":{"rendered":"https:\/\/www.zhoubin.asia\/?p=752"},"modified":"2024-08-20T14:10:38","modified_gmt":"2024-08-20T06:10:38","slug":"hailo-dataflow-compiler","status":"publish","type":"post","link":"https:\/\/www.zhoubin.asia\/index.php\/2024\/08\/21\/hailo-dataflow-compiler\/","title":{"rendered":"\u6df1\u5165\u4e86\u89e3 Hailo Dataflow Compiler \u91cf\u5316\u6a21\u578b\u811a\u672c"},"content":{"rendered":"

Hailo Dataflow Compiler \u91cf\u5316\u6a21\u578b\u811a\u672c\u8be6\u89e3<\/h1>\n

\u4e3b\u8981\u4ecb\u7ecd\u4e0b Hailo Dataflow compiler \u4e2d\u6700\u91cd\u8981\u91cf\u5316\u8fd9\u90e8\u5206\u7684 Model scripts\u3002\u8fd9\u91cc\u6211\u53ef\u80fd\u53ea\u9488\u5bf9\u6211\u81ea\u5df1\u6bd4\u8f83\u4e86\u89e3\u4e00\u70b9\u7684\u90e8\u5206\u4f1a\u8fdb\u884c\u8be6\u7ec6\u89e3\u91ca\uff0c\u6211\u81ea\u5df1\u4e0d\u4e86\u89e3\u7684\u90e8\u5206\u4ee5\u53ca\u4e0d\u91cd\u8981\u7684\u90e8\u5206\u53ef\u80fd\u4e0d\u505a\u4efb\u4f55\u89e3\u91ca\uff0c\u8be6\u7ec6\u53ef\u4ee5\u67e5\u770b Hailo \u7684\u5b98\u65b9\u6587\u6863\u3002<\/p>\n

\u53ef\u4ee5\u7b80\u5355\u7406\u89e3\uff0cHailo \u652f\u6301\u81ea\u52a8\u91cf\u5316\u4e00\u822c\u7684\u6a21\u578b\uff0c\u4f46\u662f\u4e5f\u53ef\u4ee5\u624b\u52a8\u6dfb\u52a0\u5f88\u591a\u4e0d\u540c\u547d\u4ee4\u53bb\u5b9e\u73b0\u81ea\u52a8\u91cf\u5316\u4e0b\u4e0d\u652f\u6301\u7684\u529f\u80fd\u6216\u8005\u89e3\u51b3\u91cf\u5316\u62a5\u9519\u3002\u6211\u4eec\u4e00\u822c\u5c06\u8fd9\u4e2a model scripts \u4fdd\u5b58\u4e3a alls<\/code> \u547d\u540d\u7684\u6587\u4ef6\uff0c\u4f8b\u5982\u6a21\u578b\u5e93\u91cc\u9762\u7684alls<\/a>\u76ee\u5f55\u3002\u6240\u6709\u7684 alls<\/code> \u6587\u4ef6\u90fd\u662f\u9488\u5bf9\u4e0d\u540c\u6a21\u578b\u7684 Model scripts\u3002<\/p>\n

\u5e38\u89c1\u7684 Optimization \u547d\u4ee4<\/h2>\n

\u5176\u4e2d\uff0c\u6700\u5e38\u89c1\u7684\u662f\u5f52\u4e00\u5316\uff08\u5c06\u5f52\u4e00\u5316\u653e\u5728 Hailo \u5904\u7406\uff09\uff1a<\/p>\n

normalization0 = normalization([0.0, 0.0, 0.0], [255.0, 255.0, 255.0])<\/code><\/pre>\n
\u6b63\u5e38\u6765\u8bf4\uff0c\u5927\u81f4\u5206\u4e3a 6 \u7c7b optimization \u547d\u4ee4\uff1a<\/p>\n
1. model_modification_commands<\/h3>\n
\u6a21\u578b\u53d8\u5316\uff0c\u8fd9\u4e2a\u4e00\u822c\u6307\u7684\u662f\uff0c\u9488\u5bf9\u672c\u8eab\u6a21\u578b\u6ca1\u6709\u7684\uff0c\u4f46\u662f\u6211\u4eec\u5e0c\u671b\u5728 Hailo \u5185\u8fd0\u884c\u7684\u4e00\u4e9b\u529f\u80fd\uff1a<\/p>\n
    \n
  • input_conversion<\/code> – \u8f93\u5165\u683c\u5f0f\u7684\u8f6c\u6362\uff0c\u4e3b\u8981\u662f\u989c\u8272\u7a7a\u95f4\u8f6c\u6362\uff0c\u4f8b\u5982 nv21torgb<\/code>\uff0c\u53ef\u4ee5\u5b9e\u73b0\u8f93\u5165 YUV \u7684\u6570\u636e\u5230\u4e00\u4e2a RGB \u7684\u6a21\u578b\u4e2d\uff0c\u989c\u8272\u8f6c\u6362\u5728 Hailo \u8fd0\u884c\u3002<\/li>\n
  • transpose<\/code> – \u6a21\u578b\u8f6c\u7f6e\uff0c\u56e0\u4e3a Hailo \u5bf9\u4e8e width > height<\/code> \u7684\u6a21\u578b\u8fd0\u884c\u6548\u7387\u8981\u9ad8\u4e8e height > width<\/code> \u7684\u6a21\u578b\uff0c\u6240\u4ee5\u5b9e\u5728\u6709\u9700\u8981\u7684\u65f6\u5019\u53ef\u4ee5\u8003\u8651\u3002<\/li>\n
  • normalization<\/code> – \u5f52\u4e00\u5316\uff0c\u6700\u5e38\u7528\u7684\u529f\u80fd\uff0cHailo \u672c\u8eab\u662f int \u7684 NPU\uff0c\u5f3a\u70c8\u5efa\u8bae\u5f52\u4e00\u5316\u5728 Hailo \u8fd0\u884c\u3002<\/li>\n
  • nms_postprocess<\/code> – NMS \u540e\u5904\u7406\u8fd0\u884c\u5728 Hailo\uff0c\u4e3b\u8981\u7528\u4e8e YOLO \u6a21\u578b\u7684\u540e\u5904\u7406 NMS \u90e8\u5206\uff0c\u901a\u8fc7\u8fd9\u90e8\u5206\u5c06\u540e\u5904\u7406\u8fd0\u884c\u5728 Hailort API \u5185\u90e8\uff0c\u76f4\u63a5\u5f97\u5230 NMS \u540e\u7684 BBox \u7ed3\u679c\u3002<\/li>\n
  • change_output_activation<\/code> – \u6dfb\u52a0 activation\uff0c\u4e3b\u8981\u7528\u4e8e\u4e00\u4e9b YOLO \u6a21\u578b\u7684 regression \u8282\u70b9\uff0c\u901a\u8fc7\u6dfb\u52a0\u8fd9\u4e2a\u907f\u514d\u5728\u540e\u5904\u7406\u4e2d\u6dfb\u52a0 sigmoid\u3002<\/li>\n
  • logits_layer<\/code><\/li>\n
  • set_seed<\/code><\/li>\n
  • resize<\/code><\/li>\n<\/ul>\n
    2. model_optimization_flavor<\/h3>\n
    \u901a\u8fc7\u8bbe\u7f6e\u538b\u7f29\u7ea7\u522b\u548c\u4f18\u5316\u7ea7\u522b\u6765\u914d\u7f6e\u6a21\u578b\u91cf\u5316\u4ee5\u53ca optimization \u5de5\u4f5c\uff1a<\/p>\n
    model_optimization_flavor(optimization_level=4)<\/code><\/pre>\n
    model_optimization_flavor(optimization_level=2, compression_level=1)<\/code><\/pre>\n
    model_optimization_flavor(optimization_level=2, batch_size=4)<\/code><\/pre>\n
    \u8bbe\u7f6e optimization_level<\/code> \u4ee5\u53ca compression_level<\/code>\uff08\u91cf\u5316 4bit \u7387\uff09\uff0c\u4e5f\u53ef\u4ee5\u901a\u8fc7 batch_size<\/code> \u907f\u514d GPU \u7206\u5185\u5b58\u3002<\/p>\n
    3. model_optimization_config<\/h3>\n
    \u4e0e model_optimization_flavor<\/code> \u7c7b\u4f3c\uff0c\u4f46\u662f\u5f15\u5165\u4e86\u66f4\u591a\u7684\u63a7\u5236\u91cf\u5316\u7684\u65f6\u5019\u4f7f\u7528\u7684\u7cfb\u5217\u53c2\u6570\uff0c\u5176\u4e2d\u5305\u542b\uff1a<\/p>\n
      \n
    • compression_params<\/code> – \u7c7b\u4f3c\u4e0a\u9762\u7684 compression_level<\/code>\uff0c\u4f46\u662f\u66f4\u52a0\u7cbe\u7ec6\uff0c\u800c\u4e14\u652f\u6301 16bits \u7684\u91cf\u5316\uff0c\u6bd4\u5982\u8bbe\u7f6e\u6240\u6709\u652f\u6301\u7684\u5c42\u4e3a 16bits\uff1a\n
      model_optimization_config(compression_params, auto_16bit_weights_ratio=1)<\/code><\/pre>\n<\/li>\n
    • negative_exponent<\/code><\/li>\n
    • globals<\/code><\/li>\n
    • calibration<\/code> – \u4e3b\u8981\u662f\u901a\u8fc7\u9ed8\u8ba4\u662f\u53ea\u4f7f\u7528 64 \u5f20\u56fe\u7247\u8fdb\u884c\u91cf\u5316\u6570\u636e\u5206\u6790\u7684\uff0c\u5982\u679c\u7cbe\u5ea6\u51fa\u73b0\u5f02\u5e38\uff0c\u53ef\u4ee5\u5c1d\u8bd5\u589e\u52a0\u8fd9\u4e2a\u6570\u5b57\uff0c\u4f46\u662f\u589e\u52a0\u8fd9\u4e2a\u53ef\u80fd\u5bfc\u81f4\u91cf\u5316\u6548\u679c\u4e0d\u597d\uff0c\u4e00\u822c\u4f1a\u914d\u5408 clipping \u4f7f\u7528\uff1a\n
      model_optimization_config(calibration, batch_size=4, calibset_size=128)<\/code><\/pre>\n<\/li>\n
    • checker_cfg<\/code><\/li>\n<\/ul>\n
      4. quantization_param<\/h3>\n
      \u4e3b\u8981\u662f\u7528\u4e8e\u624b\u52a8\u8c03\u6574\u6bcf\u5c42\u6216\u8005\u6307\u5b9a\u5c42\u7684\u4e00\u4e9b\u91cf\u5316\u53c2\u6570\uff0c\u57fa\u672c\u683c\u5f0f\u4e3a\uff1a<\/p>\n
      quantization_param(<layer>, <parameter>=<value>)<\/code><\/pre>\n
      \u57fa\u672c\u53c2\u6570\u6709\uff1a<\/p>\n
        \n
      • bias_mode<\/code> – \u8bbe\u7f6e\u5c42\u7684 bias \u884c\u4e3a\uff0c\u4f8b\u5982\uff1a\u4f7f\u7528 16bits \u800c\u975e\u9ed8\u8ba4\u7684 8bits\u3002<\/li>\n
      • precision_mode<\/code> – \u5e38\u7528\uff0c\u7cbe\u5ea6\u6a21\u5f0f\u8bbe\u7f6e\u7528\u4e8e\u5c42\u7684\u6743\u91cd\u548c\u6fc0\u6d3b\u8868\u793a\u7684\u4f4d\u6570\uff1a\n
        quantization_param([conv1, maxpool2], precision_mode=a16_w16)<\/code><\/pre>\n<\/li>\n
      • quantization_groups<\/code><\/li>\n
      • force_range_out<\/code> – \u6b64\u547d\u4ee4\u5728\u91cf\u5316\u8fc7\u7a0b\u4e2d\u5f3a\u5236\u5c06\u6307\u5b9a\u8303\u56f4\u5e94\u7528\u4e8e\u7ed9\u5b9a\u5c42\u7684\u8f93\u51fa\u3002\u4e00\u822c\u662f\u901a\u8fc7 SNR \u5f97\u5230\u5f53\u524d\u5c42\u7684\u4e3b\u8981\u6570\u636e\u8f93\u51fa\uff0c\u5229\u7528\u8fd9\u4e2a\u6765\u5f3a\u5236\u9650\u5236\uff0c\u6d88\u9664 noise\uff0c\u67e5\u770b\u7cbe\u5ea6\u662f\u5426\u6539\u5584\u3002<\/li>\n
      • max_elementwise_feed_repeat<\/code><\/li>\n
      • max_bias_feed_repeat<\/code><\/li>\n
      • null_channels_cutoff_factor<\/code><\/li>\n
      • output_encoding_vector<\/code><\/li>\n
      • gpu_policy<\/code><\/li>\n<\/ul>\n
        5. pre_quantization_optimization<\/h3>\n
        \u6b64\u547d\u4ee4\u7684\u6240\u6709\u529f\u80fd\u5728\u91cf\u5316\u8fc7\u7a0b\u4e4b\u524d\u4f18\u5316\u6a21\u578b\u3002\u5176\u4e2d\u4e00\u4e9b\u547d\u4ee4\u4f1a\u4fee\u6539\u6a21\u578b\u7ed3\u6784\uff0c\u5e76\u5728\u5176\u4f59\u547d\u4ee4\u4e4b\u524d\u6267\u884c\uff1a<\/p>\n
          \n
        • dead_channels_removal<\/code><\/li>\n
        • zero_static_channels<\/code><\/li>\n
        • zero_static_channels per-layer<\/code><\/li>\n
        • se_optimization<\/code><\/li>\n
        • equalization<\/code> – \u6b64\u5b50\u547d\u4ee4\u5141\u8bb8\u5728\u91cf\u5316\u524d\u914d\u7f6e\u5168\u5c40\u5747\u8861\u884c\u4e3a\uff0c\u5e76\u66ff\u4ee3\u4e86 quantize()<\/code> API \u4e2d\u7684\u65e7\u5747\u8861\u53c2\u6570\u3002\u8fd9\u4e2a\u5176\u5b9e\u5c31\u662f model_optimization_flavor(optimization_level=0)<\/code> \u6240\u505a\u7684\u4efb\u52a1\uff0c\u5f53\u7136\uff0c\u7279\u6b8a\u6a21\u578b\u53ef\u4ee5\u901a\u8fc7\u8fd9\u4e2a\u547d\u4ee4\u5173\u95ed\u6b64\u529f\u80fd\u3002<\/li>\n
        • equalization per-layer<\/code><\/li>\n
        • dead_layers_removal<\/code><\/li>\n
        • weights_clipping<\/code> – \u6b64\u547d\u4ee4\u5141\u8bb8\u5728\u8fd0\u884c\u91cf\u5316 API \u65f6\u66f4\u6539\u9009\u5b9a\u5c42\u7684\u6b64\u884c\u4e3a\u5e76\u5e94\u7528\u6743\u91cd\u526a\u88c1\u3002\u5bf9\u4e8e\u6743\u91cd\u503c\u5f02\u5e38\u7684\u60c5\u51b5\uff0c\u6b64\u547d\u4ee4\u53ef\u80fd\u6709\u52a9\u4e8e\u51cf\u5c11\u91cf\u5316\u76f8\u5173\u7684\u9000\u5316\u3002\u5b83\u4ec5\u9002\u7528\u4e8e\u6709\u6743\u91cd\u7684\u5c42\u3002\u975e\u5e38\u91cd\u8981\u7684\u901a\u8fc7 SNR \u6765\u8fdb\u884c\u64cd\u4f5c\u5b9e\u73b0\u7cbe\u5ea6\u63d0\u5347\u7684\u65b9\u5f0f\u3002<\/li>\n
        • activation_clipping<\/code> – \u7c7b\u4f3c weights_clipping<\/code>\uff0c\u53ea\u662f\u7528\u4e8e\u6fc0\u6d3b\u5c42\u3002<\/li>\n
        • ew_add_fusing<\/code> – \u5f53\u542f\u7528\u6b64\u529f\u80fd\u65f6\uff0cew_add<\/code> \u5c42\u5c06\u878d\u5408\u5230\u5377\u79ef\u5c42\uff08conv<\/code>\uff09\u548c\u52a0\u6cd5\u5c42\uff08add<\/code>\uff09\u4e2d\u3002\u5177\u6709\u4e0d\u517c\u5bb9\u7cbe\u5ea6\u6a21\u5f0f\u7684\u5c42\u5c06\u4e0d\u4f1a\u88ab\u878d\u5408\u3002\u7c7b\u6bd4\u7406\u89e3 BN \u878d\u5408\u5165 Conv\u3002<\/li>\n
        • layer_decomposition<\/code><\/li>\n
        • smart_softmax_stats<\/code><\/li>\n
        • defuse<\/code> – \u6b64\u547d\u4ee4\u5141\u8bb8\u6839\u636e\u53bb\u878d\u5408\u7c7b\u578b\u5bf9\u5c42\u8fdb\u884c\u53bb\u878d\u5408\uff1a\n
            \n
          • INPUT FEATURES<\/code> – \u5bf9\u4e8e\u9009\u5b9a\u7684\u5168\u8fde\u63a5\u5c42\u6216\u5377\u79ef\u5c42\uff0c\u5c06\u8f93\u5165\u7279\u5f81\u53bb\u878d\u5408\u4e3a\u6307\u5b9a\u6570\u91cf\u7684\u5206\u5272\u3002\u4e5f\u53ef\u4ee5\u7528\u6765\u7981\u7528\u67d0\u5c42\u7684\u53bb\u878d\u5408\u3002<\/li>\n
          • MULTI-HEAD ATTENTION (MHA)<\/code> – \u5141\u8bb8\u5c06\u7531\u5176\u7b2c\u4e00\u4e2a\u77e9\u9635\u4e58\u6cd5\u8868\u793a\u7684\u591a\u5934\u6ce8\u610f\u529b\u5757\u53bb\u878d\u5408\u4e3a\u6307\u5b9a\u6570\u91cf\u7684\u5206\u5272\u3002\u89e3\u51b3\u4e00\u4e9b\u5c42\u6bd4\u8f83\u5927\uff0c\u4e0d\u80fd\u91cf\u5316\u7684\u95ee\u9898\u3002<\/li>\n<\/ul>\n<\/li>\n
          • resolution_reduction<\/code> – \u51cf\u5c11\u6240\u6709\u8f93\u5165\u5c42\u7684\u6a21\u578b\u5206\u8fa8\u7387\uff0c\u4ee5\u4fbf\u66f4\u6709\u6548\u5730\u4f18\u5316\u6a21\u578b\u3002\u8fd9\u5bf9\u7cbe\u5ea6\u4f1a\u6709\u5f71\u54cd\u3002<\/li>\n
          • resolution_reduction per-layer<\/code><\/li>\n
          • global_avgpool_reduction<\/code> – \u6b64\u547d\u4ee4\u5141\u8bb8\u901a\u8fc7\u4f7f\u7528\u989d\u5916\u7684 avgpool \u6765\u51cf\u5c11 avgpool \u7684\u7a7a\u95f4\u7ef4\u5ea6\u3002\u6dfb\u52a0\u7684 avgpool \u7684\u6838\u5927\u5c0f\u5c06\u4e3a [1, h \/\/ division_factors[0], w \/\/ division_factors[1], 1]\u3002\u53ef\u4ee5\u89e3\u51b3\u67d0\u4e9b\u65f6\u5019\u91cf\u5316\u7684\u65f6\u5019\u9488\u5bf9 avgpool \u7684\u91cf\u5316\u62a5\u9519\u3002<\/li>\n
          • add_shortcut_layer<\/code><\/li>\n
          • matmul_correction<\/code><\/li>\n
          • matmul_equalization<\/code><\/li>\n
          • split_ew_mult_by_bit_significance<\/code><\/li>\n<\/ul>\n
            6. post_quantization_optimization<\/h3>\n
            \u6b64\u547d\u4ee4\u7684\u6240\u6709\u529f\u80fd\u5728\u91cf\u5316\u8fc7\u7a0b\u4e4b\u540e\u5bf9\u91cf\u5316\u7684\u6a21\u578b\u8fdb\u884c\u4f18\u5316\uff0c\u4ee5\u4fbf\u63d0\u5347\u6a21\u578b\u7684\u7cbe\u5ea6\uff1a<\/p>\n
              \n
            • bias_correction<\/code> – \u6b64\u5b50\u547d\u4ee4\u5141\u8bb8\u5728\u91cf\u5316\u540e\u8fc7\u7a0b\u671f\u95f4\u914d\u7f6e\u5168\u5c40\u504f\u7f6e\u6821\u6b63\u884c\u4e3a\uff0c\u6b64\u547d\u4ee4\u66ff\u4ee3\u4e86 quantize()<\/code> API \u4e2d\u7684\u65e7 ibc<\/code> \u53c2\u6570\u3002<\/li>\n
            • bias_correction per-layer<\/code><\/li>\n
            • train_encoding<\/code><\/li>\n
            • finetune<\/code> – \u6b64\u5b50\u547d\u4ee4\u542f\u7528\u57fa\u4e8e\u77e5\u8bc6\u84b8\u998f\u7684\u91cf\u5316\u6765\u505a\u6a21\u578b\u5fae\u8c03\u3002\u7b80\u5355\u6765\u8bf4\u5c31\u662f\u5229\u7528\u6d6e\u70b9\u6a21\u578b\u6765\u77e5\u8bc6\u84b8\u998f\u91cf\u5316\u6a21\u578b\uff0c\u4ee5\u4fbf\u63d0\u9ad8\u91cf\u5316\u6a21\u578b\u7684\u7cbe\u5ea6\u3002optimization_level=2<\/code> \u5c31\u662f\u8fd9\u4e2a\u7684\u7b80\u6613\u7248\u672c\u3002<\/li>\n
            • adaround<\/code> – \u540c finetune<\/code>\uff0c\u53ea\u662f\u5bf9\u6bcf\u4e00\u5c42\u90fd\u8fdb\u884c\u77e5\u8bc6\u84b8\u998f\uff08optimization_level=4<\/code>\uff09\uff0c\u975e\u5e38\u8017\u65f6\u3002<\/li>\n
            • adaround per-layer<\/code><\/li>\n
            • mix_precision_search<\/code><\/li>\n<\/ul>\n
               <\/p>\n","protected":false},"excerpt":{"rendered":"
              Hailo Dataflow Compiler \u91cf\u5316\u6a21\u578b\u811a\u672c\u8be6\u89e3 \u4e3b\u8981\u4ecb\u7ecd\u4e0b Hailo Dataflow c […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_bbp_topic_count":0,"_bbp_reply_count":0,"_bbp_total_topic_count":0,"_bbp_total_reply_count":0,"_bbp_voice_count":0,"_bbp_anonymous_reply_count":0,"_bbp_topic_count_hidden":0,"_bbp_reply_count_hidden":0,"_bbp_forum_subforum_count":0,"footnotes":""},"categories":[16,1],"tags":[],"class_list":["post-752","post","type-post","status-publish","format-standard","hentry","category-dfc","category-tech"],"_links":{"self":[{"href":"https:\/\/www.zhoubin.asia\/index.php\/wp-json\/wp\/v2\/posts\/752","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.zhoubin.asia\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.zhoubin.asia\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.zhoubin.asia\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.zhoubin.asia\/index.php\/wp-json\/wp\/v2\/comments?post=752"}],"version-history":[{"count":1,"href":"https:\/\/www.zhoubin.asia\/index.php\/wp-json\/wp\/v2\/posts\/752\/revisions"}],"predecessor-version":[{"id":753,"href":"https:\/\/www.zhoubin.asia\/index.php\/wp-json\/wp\/v2\/posts\/752\/revisions\/753"}],"wp:attachment":[{"href":"https:\/\/www.zhoubin.asia\/index.php\/wp-json\/wp\/v2\/media?parent=752"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.zhoubin.asia\/index.php\/wp-json\/wp\/v2\/categories?post=752"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.zhoubin.asia\/index.php\/wp-json\/wp\/v2\/tags?post=752"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}