{"id":103396,"date":"2026-09-10T11:57:06","date_gmt":"2026-09-10T03:57:06","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/103396.html"},"modified":"2026-09-10T11:57:06","modified_gmt":"2026-09-10T03:57:06","slug":"%e8%bd%bb%e6%9d%be%e5%ad%a6%e4%b9%a0onnx_day5","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/103396.html","title":{"rendered":"\u8f7b\u677e\u5b66\u4e60ONNX_day5"},"content":{"rendered":"<p>\u6458\u8981&#xff1a;\u672c\u6587\u662f ONNX \u90e8\u7f72\u5b66\u4e60\u8def\u7ebf\u7684 Step 5&#xff0c;\u8bb2\u89e3\u5982\u4f55\u5c06 ONNX \u6a21\u578b\u8f6c\u6362\u4e3a TensorRT Engine \u5e76\u5728 NVIDIA GPU \u4e0a\u9ad8\u6548\u63a8\u7406\u3002\u5185\u5bb9\u6db5\u76d6 TensorRT \u4e0e ONNX Runtime \u7684\u5b9a\u4f4d\u5dee\u5f02\u3001Engine \u6838\u5fc3\u6982\u5ff5\u3001FP32\/FP16\/INT8 \u7cbe\u5ea6\u652f\u6301\u3001\u4e0e TFLM INT8 \u7684\u533a\u522b\u3001MacBook Air \u7684\u786c\u4ef6\u9650\u5236&#xff0c;\u4ee5\u53ca\u4ece trtexec \u547d\u4ee4\u884c\u5de5\u5177\u5230 Python API \u7684\u5b8c\u6574\u5b9e\u8df5\u8def\u5f84&#xff0c;\u6700\u7ec8\u901a\u8fc7 6 \u4e2a\u5c0f\u6b65\u9aa4\u5e26\u4f60\u8dd1\u901a ONNX \u2192 TensorRT \u5168\u6d41\u7a0b\u3002<\/p>\n<p>\u597d\u3002\u6309\u7167\u4f60\u524d\u9762\u7684\u5b66\u4e60\u8def\u7ebf&#xff0c;Step 5&#xff1a;ONNX \u2192 TensorRT \u5f88\u9002\u5408\u653e\u5728 Step 4&#xff08;ONNX \u2192 ARM Linux&#xff09;\u4e4b\u540e\u3002<\/p>\n<p>\u4e0d\u8fc7\u8fd9\u91cc\u8981\u5148\u660e\u786e\u4e00\u4e2a\u5173\u952e\u70b9&#xff1a;<\/p>\n<p>TensorRT \u4e3b\u8981\u9762\u5411 NVIDIA GPU&#xff0c;\u4e0d\u662f ARM CPU\u3002<\/p>\n<p>\u6240\u4ee5\u8fd9\u4e00\u9636\u6bb5\u6211\u4eec\u5148\u5728\u4f60\u7684 Intel MacBook Air \u4e0a\u5b66\u4e60 TensorRT \u7684\u5de5\u4f5c\u6d41\u7a0b\u548c\u6982\u5ff5&#xff1b;\u771f\u6b63\u8fd0\u884c TensorRT inference&#xff0c;\u5efa\u8bae\u4f7f\u7528 NVIDIA GPU Linux \u73af\u5883&#xff0c;\u4f8b\u5982 Ubuntu &#043; NVIDIA GPU&#xff0c;\u6216\u8005\u540e\u9762\u4f7f\u7528 Jetson\u3002<\/p>\n<p>Step 5&#xff1a;ONNX \u2192 TensorRT<\/p>\n<p>\u6574\u4e2a\u8def\u7ebf\u53ef\u4ee5\u7406\u89e3\u6210&#xff1a;<\/p>\n<p>Training<br \/>\n\u2502<br \/>\n\u25bc<br \/>\nPyTorch Model<br \/>\n\u2502<br \/>\n\u2502 <span class=\"token builtin class-name\">export<\/span><br \/>\n\u25bc<br \/>\nmodel.onnx<br \/>\n\u2502<br \/>\n\u2502 TensorRT<br \/>\n\u25bc<br \/>\nTensorRT Engine<br \/>\n<span class=\"token punctuation\">(<\/span>.engine \/ .plan<span class=\"token punctuation\">)<\/span><br \/>\n\u2502<br \/>\n\u25bc<br \/>\nNVIDIA GPU<br \/>\n\u2502<br \/>\n\u25bc<br \/>\nInference<\/p>\n<p>\u548c\u4f60\u524d\u9762\u7684\u8def\u7ebf\u5bf9\u5e94\u8d77\u6765&#xff1a;<\/p>\n<p>Step <span class=\"token number\">1<\/span><br \/>\nPyTorch<br \/>\n\u2193<br \/>\nONNX<br \/>\n\u2193<br \/>\nONNX Runtime<br \/>\n\u2193<br \/>\nLinux\/x86<\/p>\n<p>Step <span class=\"token number\">2<\/span><br \/>\nONNX<br \/>\n\u2193<br \/>\nONNX Runtime<br \/>\n\u2193<br \/>\nARM Linux<\/p>\n<p>Step <span class=\"token number\">3<\/span><br \/>\nONNX<br \/>\n\u2193<br \/>\nINT8 Quantization<\/p>\n<p>Step <span class=\"token number\">4<\/span><br \/>\nONNX<br \/>\n\u2193<br \/>\nARM Linux<\/p>\n<p>Step <span class=\"token number\">5<\/span><br \/>\nONNX<br \/>\n\u2193<br \/>\nTensorRT<br \/>\n\u2193<br \/>\nNVIDIA GPU<\/p>\n<p>5.1 TensorRT \u5230\u5e95\u662f\u4ec0\u4e48&#xff1f;<\/p>\n<p>\u53ef\u4ee5\u628a&#xff1a;<\/p>\n<p>ONNX Runtime<\/p>\n<p>\u548c<\/p>\n<p>TensorRT<\/p>\n<p>\u7406\u89e3\u6210\u4e24\u79cd\u4e0d\u540c\u7684 inference runtime\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260910035705-6aa22a91466bc.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>ONNX Runtime \u66f4\u5f3a\u8c03&#xff1a;<\/p>\n<p>\u901a\u7528\u6027<br \/>\nTensorRT \u66f4\u5f3a\u8c03&#xff1a;<br \/>\nNVIDIA GPU \u4e0a\u7684\u6027\u80fd<br \/>\nTensorRT \u4f1a\u5bf9 ONNX \u6a21\u578b\u8fdb\u884c\u8fdb\u4e00\u6b65\u4f18\u5316&#xff0c;\u4f8b\u5982&#xff1a;<\/p>\n<p>ONNX Model<br \/>\n\u2502<br \/>\n\u25bc<br \/>\nTensorRT Parser<br \/>\n\u2502<br \/>\n\u25bc<br \/>\nGraph Optimization<br \/>\n\u2502<br \/>\n\u251c\u2500\u2500 Layer Fusion<br \/>\n\u251c\u2500\u2500 Kernel Selection<br \/>\n\u251c\u2500\u2500 Precision Optimization<br \/>\n\u2514\u2500\u2500 Memory Optimization<br \/>\n\u2502<br \/>\n\u25bc<br \/>\nTensorRT Engine<\/p>\n<p>\u6700\u7ec8\u5f97\u5230\u4e00\u4e2a\u9488\u5bf9\u7279\u5b9a NVIDIA GPU \u4f18\u5316\u7684 inference engine\u3002<\/p>\n<p>5.2 \u6700\u91cd\u8981\u7684\u6982\u5ff5&#xff1a;Engine<\/p>\n<p>\u8fd9\u662f Step 5 \u6700\u9700\u8981\u7406\u89e3\u7684\u4e1c\u897f\u3002<br \/>\nONNX&#xff1a;<br \/>\nmodel.onnx<\/p>\n<p>\u672c\u8d28\u4e0a\u662f\u4e00\u4e2a&#xff1a;<br \/>\n\u795e\u7ecf\u7f51\u7edc\u8ba1\u7b97\u56fe &#043; \u6743\u91cd &#043; \u6a21\u578b\u4fe1\u606f<\/p>\n<p>\u800c TensorRT \u6700\u7ec8\u4ea7\u751f&#xff1a;<br \/>\nmodel.engine<\/p>\n<p>\u6216\u8005&#xff1a;<br \/>\nmodel.plan<\/p>\n<p>\u5b83\u66f4\u63a5\u8fd1&#xff1a;<br \/>\n\u5df2\u7ecf\u9488\u5bf9 NVIDIA GPU \u4f18\u5316\u597d\u7684\u53ef\u6267\u884c inference engine<\/p>\n<p>\u56e0\u6b64&#xff1a;<\/p>\n<p>ONNX<br \/>\n\u2502<br \/>\n\u2502 TensorRT build<br \/>\n\u25bc<br \/>\nEngine<\/p>\n<p>\u8fd9\u4e2a\u8fc7\u7a0b\u53eb&#xff1a;<br \/>\nEngine Building<\/p>\n<p>5.3 TensorRT \u7684\u57fa\u672c\u5de5\u4f5c\u6d41\u7a0b<\/p>\n<p>\u5178\u578b\u6d41\u7a0b&#xff1a;<br \/>\n<img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260910035705-6aa22a915f528.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><br \/>\n\u8fd9\u91cc\u8981\u7279\u522b\u533a\u5206&#xff1a;<br \/>\nBuild<br \/>\nONNX \u2192 Engine<\/p>\n<p>Runtime<br \/>\nEngine \u2192 Inference<\/p>\n<p>\u4e5f\u5c31\u662f\u8bf4&#xff1a;<\/p>\n<p>TensorRT Builder<br \/>\n\u2502<br \/>\n\u2502 build once<br \/>\n\u25bc<br \/>\nmodel.engine<br \/>\n\u2502<br \/>\n\u2502 run many <span class=\"token builtin class-name\">times<\/span><br \/>\n\u25bc<br \/>\nTensorRT Runtime<\/p>\n<p>5.4 TensorRT \u652f\u6301\u54ea\u4e9b Precision&#xff1f;<\/p>\n<p>\u8fd9\u662f\u4f60\u524d\u9762\u5b66\u4e60 INT8 Quantization \u540e\u975e\u5e38\u91cd\u8981\u7684\u4e00\u73af\u3002<br \/>\nTensorRT \u5e38\u89c1 precision&#xff1a;<\/p>\n<p>FP32<br \/>\n\u2502<br \/>\n\u251c\u2500\u2500 FP16<br \/>\n\u2502<br \/>\n\u2514\u2500\u2500 INT8<\/p>\n<p>\u4f8b\u5982&#xff1a;<\/p>\n<p>FP32<\/p>\n<p>ONNX<br \/>\n\u2193<br \/>\nTensorRT<br \/>\n\u2193<br \/>\nFP32 Engine<\/p>\n<p>FP16<\/p>\n<p>ONNX<br \/>\n\u2193<br \/>\nTensorRT<br \/>\n\u2193<br \/>\nFP16 Engine<\/p>\n<p>INT8<\/p>\n<p>ONNX<br \/>\n\u2193<br \/>\nCalibration \/ Quantization<br \/>\n\u2193<br \/>\nTensorRT<br \/>\n\u2193<br \/>\nINT8 Engine<\/p>\n<p>\u56e0\u6b64\u4f60\u524d\u9762\u5b66\u7684&#xff1a;<br \/>\nONNX \u2192 INT8<br \/>\n\u5e76\u4e0d\u662f\u53ea\u80fd\u7528\u4e8e CPU\/ARM\u3002<\/p>\n<p>\u5b83\u540c\u6837\u662f\u8fdb\u5165&#xff1a;<br \/>\nTensorRT INT8 inference<br \/>\n\u7684\u91cd\u8981\u57fa\u7840\u3002<\/p>\n<p>5.5 TensorRT INT8 \u548c TFLM INT8 \u7684\u533a\u522b<br \/>\n\u8fd9\u4e2a\u5bf9\u4f60\u7684\u5b66\u4e60\u8def\u7ebf\u975e\u5e38\u91cd\u8981\u3002<br \/>\n\u4f60\u5df2\u7ecf\u505a\u8fc7&#xff1a;<\/p>\n<p>model.keras<br \/>\n\u2193<br \/>\nTFLite<br \/>\n\u2193<br \/>\nFull Integer INT8<br \/>\n\u2193<br \/>\nTFLM<br \/>\n\u2193<br \/>\nSTM32F303RE<\/p>\n<p>\u73b0\u5728 TensorRT&#xff1a;<\/p>\n<p>model.keras<br \/>\n\u2193<br \/>\nONNX<br \/>\n\u2193<br \/>\nTensorRT<br \/>\n\u2193<br \/>\nINT8 Engine<br \/>\n\u2193<br \/>\nNVIDIA GPU<\/p>\n<p>\u867d\u7136\u90fd\u53eb INT8&#xff0c;\u4f46\u662f&#xff1a;<\/p>\n<p>TFLM INT8<br \/>\n\u2193<br \/>\nMCU<br \/>\n\u2193<br \/>\n\u6781\u4f4e\u529f\u8017<br \/>\n\u2193<br \/>\nCPU<\/p>\n<p>\u800c&#xff1a;<\/p>\n<p>TensorRT INT8<br \/>\n\u2193<br \/>\nNVIDIA GPU<br \/>\n\u2193<br \/>\n\u9ad8\u541e\u5410\u91cf<br \/>\n\u2193<br \/>\nAI inference<\/p>\n<p>\u8fd9\u662f\u4e24\u4e2a\u5b8c\u5168\u4e0d\u540c\u7684\u4f18\u5316\u65b9\u5411\u3002<\/p>\n<p>5.6 \u4f60\u7684 MacBook Air \u6709\u4e00\u4e2a\u95ee\u9898<\/p>\n<p>\u4f60\u73b0\u5728\u7684\u673a\u5668\u662f&#xff1a;<br \/>\nMacBook Air<br \/>\nIntel CPU<br \/>\nmacOS<\/p>\n<p>\u6240\u4ee5&#xff1a;<br \/>\n\u4e0d\u9002\u5408\u76f4\u63a5\u505a TensorRT GPU inference\u3002<\/p>\n<p>TensorRT \u7684\u6838\u5fc3\u76ee\u6807\u5e73\u53f0\u662f NVIDIA GPU\u3002<br \/>\n\u4f60\u7684 Mac \u53ef\u4ee5\u7528\u4e8e&#xff1a;<\/p>\n<p>PyTorch<br \/>\n\u2193<br \/>\nONNX<br \/>\n\u2193<br \/>\nONNX inspection<br \/>\n\u2193<br \/>\nONNX Runtime CPU<\/p>\n<p>\u4f46 TensorRT \u6700\u7ec8\u6d4b\u8bd5\u6700\u597d\u653e\u5230&#xff1a;<\/p>\n<p>Ubuntu<\/p>\n<ul>\n<li><\/li>\n<\/ul>\n<p>NVIDIA GPU<\/p>\n<ul>\n<li><\/li>\n<\/ul>\n<p>CUDA<\/p>\n<ul>\n<li><\/li>\n<\/ul>\n<p>TensorRT<\/p>\n<p>\u4f8b\u5982&#xff1a;<\/p>\n<p>Ubuntu <span class=\"token number\">22.04<\/span><br \/>\n\u2502<br \/>\n\u251c\u2500\u2500 NVIDIA Driver<br \/>\n\u251c\u2500\u2500 CUDA<br \/>\n\u251c\u2500\u2500 cuDNN<br \/>\n\u2514\u2500\u2500 TensorRT<br \/>\n\u2502<br \/>\n\u25bc<br \/>\nmodel.onnx<br \/>\n\u2502<br \/>\n\u25bc<br \/>\nmodel.engine<br \/>\n\u2502<br \/>\n\u25bc<br \/>\nNVIDIA GPU<\/p>\n<p>5.7 \u6211\u5efa\u8bae\u4f60\u7684 Step 5 \u5b9e\u9a8c\u4e0d\u8981\u4e00\u5f00\u59cb\u5c31\u641e\u590d\u6742\u6a21\u578b<br \/>\n\u6211\u4eec\u7ee7\u7eed\u4f7f\u7528\u4f60\u4e4b\u524d\u7684\u7b80\u5355 Sine \u6a21\u578b\u3002<\/p>\n<p>\u4f8b\u5982&#xff1a;<\/p>\n<p>input<br \/>\n\u2502<br \/>\n\u25bc<br \/>\nDense<br \/>\n\u2502<br \/>\n\u25bc<br \/>\nReLU<br \/>\n\u2502<br \/>\n\u25bc<br \/>\nDense<br \/>\n\u2502<br \/>\n\u25bc<br \/>\noutput<\/p>\n<p>\u7136\u540e\u5b8c\u6574\u8d70&#xff1a;<\/p>\n<p>PyTorch\/Keras<br \/>\n\u2193<br \/>\nONNX<br \/>\n\u2193<br \/>\nONNX Runtime<br \/>\n\u2193<br \/>\nTensorRT<br \/>\n\u2193<br \/>\nFP32 Engine<br \/>\n\u2193<br \/>\nFP16 Engine<br \/>\n\u2193<br \/>\nINT8 Engine<\/p>\n<p>\u6700\u540e\u6bd4\u8f83&#xff1a;<\/p>\n<p>Accuracy Latency<br \/>\nONNX Runtime \u2500\u2500\u2500 \u2500\u2500\u2500<br \/>\nTensorRT FP32 \u2500\u2500\u2500 \u2500\u2500\u2500<br \/>\nTensorRT FP16 \u2500\u2500\u2500 \u2500\u2500\u2500<br \/>\nTensorRT INT8 \u2500\u2500\u2500 \u2500\u2500\u2500<\/p>\n<p>\u8fd9\u6837\u4f60\u4f1a\u771f\u6b63\u7406\u89e3 TensorRT \u7684\u4ef7\u503c&#xff0c;\u800c\u4e0d\u662f\u53ea\u4f1a\u6267\u884c\u4e00\u4e2a\u547d\u4ee4\u3002<\/p>\n<p>5.8 \u7b2c\u4e00\u4e2a TensorRT \u5de5\u5177&#xff1a;trtexec<\/p>\n<p>TensorRT \u6709\u4e00\u4e2a\u975e\u5e38\u91cd\u8981\u7684\u547d\u4ee4\u884c\u5de5\u5177&#xff1a;<\/p>\n<p>trtexec<\/p>\n<p>\u4f8b\u5982&#xff1a;<\/p>\n<p>trtexec &#8211;onnx&#061;model.onnx<\/p>\n<p>\u5b83\u53ef\u4ee5\u76f4\u63a5&#xff1a;<\/p>\n<p>model.onnx<br \/>\n\u2193<br \/>\nTensorRT<br \/>\n\u2193<br \/>\nbuild engine<br \/>\n\u2193<br \/>\nbenchmark<\/p>\n<p>\u4f8b\u5982 FP16&#xff1a;<\/p>\n<p>trtexec <span class=\"token punctuation\">\\\\<\/span><span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token punctuation\">\\\\<\/span>&#8211;onnx<span class=\"token operator\">&#061;<\/span>model.onnx <span class=\"token punctuation\">\\\\<\/span><span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token punctuation\">\\\\<\/span>&#8211;fp16<\/p>\n<p>INT8&#xff1a;<\/p>\n<p>trtexec <span class=\"token punctuation\">\\\\<\/span><span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token punctuation\">\\\\<\/span>&#8211;onnx<span class=\"token operator\">&#061;<\/span>model.onnx <span class=\"token punctuation\">\\\\<\/span><span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token punctuation\">\\\\<\/span>&#8211;int8<\/p>\n<p>\u4fdd\u5b58 engine&#xff1a;<\/p>\n<p>trtexec <span class=\"token punctuation\">\\\\<\/span><span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token punctuation\">\\\\<\/span>&#8211;onnx<span class=\"token operator\">&#061;<\/span>model.onnx <span class=\"token punctuation\">\\\\<\/span><span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token punctuation\">\\\\<\/span>&#8211;fp16 <span class=\"token punctuation\">\\\\<\/span><span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token punctuation\">\\\\<\/span>&#8211;saveEngine<span class=\"token operator\">&#061;<\/span>model_fp16.engine<\/p>\n<p>\u7136\u540e\u53ef\u4ee5\u52a0\u8f7d&#xff1a;<\/p>\n<p>trtexec <span class=\"token punctuation\">\\\\<\/span><span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token punctuation\">\\\\<\/span>&#8211;loadEngine<span class=\"token operator\">&#061;<\/span>model_fp16.engine<br \/>\n&#096;&#096;<span class=\"token variable\"><span class=\"token variable\">&#096;<\/span><\/p>\n<p>\u4e0b\u9762\u6211\u4eec\u7528\u4e4b\u524d\u8bad\u7ec3\u597d\u7684 **Sine \u6a21\u578b**&#xff0c;\u5b8c\u6574\u8d70\u4e00\u904d\u4ece ONNX \u5230\u4e09\u79cd\u7cbe\u5ea6 Engine \u7684\u5b9e\u6218\u3002<\/p>\n<p>**Step A \u2014 \u51c6\u5907 Sine \u6a21\u578b\u7684 ONNX \u6587\u4ef6**<\/p>\n<p>\u5047\u8bbe\u4f60\u5df2\u7ecf\u7528 PyTorch \u6216 Keras \u8bad\u7ec3\u597d Sine \u6a21\u578b\u5e76\u5bfc\u51fa\u4e3a <span class=\"token variable\">&#096;<\/span><\/span>sine_model.onnx<span class=\"token variable\"><span class=\"token variable\">&#096;<\/span>\u3002\u5148\u786e\u8ba4\u6587\u4ef6\u5b58\u5728&#xff1a;<\/p>\n<p><span class=\"token variable\">&#096;<\/span><\/span>&#096;&#096;bash<br \/>\n<span class=\"token function\">ls<\/span> <span class=\"token parameter variable\">-lh<\/span> sine_model.onnx<\/p>\n<p>\u9884\u671f\u8f93\u51fa&#xff1a;<\/p>\n<p>-rw-r&#8211;r&#8211; <span class=\"token number\">1<\/span> user user <span class=\"token number\">2<\/span>.3K Sep <span class=\"token number\">10<\/span> <span class=\"token number\">10<\/span>:00 sine_model.onnx<\/p>\n<p>Step B \u2014 \u6784\u5efa FP32 Engine<\/p>\n<p>FP32 \u662f\u9ed8\u8ba4\u7cbe\u5ea6&#xff0c;\u76f4\u63a5\u6784\u5efa&#xff1a;<\/p>\n<p>trtexec <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token parameter variable\">&#8211;onnx<\/span><span class=\"token operator\">&#061;<\/span>sine_model.onnx <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token parameter variable\">&#8211;saveEngine<\/span><span class=\"token operator\">&#061;<\/span>sine_fp32.engine<\/p>\n<p>\u9884\u671f\u8f93\u51fa\u7247\u6bb5&#xff08;\u5173\u952e\u884c&#xff09;&#xff1a;<\/p>\n<p><span class=\"token punctuation\">[<\/span>I<span class=\"token punctuation\">]<\/span> <span class=\"token operator\">&#061;&#061;<\/span><span class=\"token operator\">&#061;<\/span> Model Options <span class=\"token operator\">&#061;&#061;<\/span><span class=\"token operator\">&#061;<\/span><br \/>\n<span class=\"token punctuation\">[<\/span>I<span class=\"token punctuation\">]<\/span>   Format: ONNX<br \/>\n<span class=\"token punctuation\">[<\/span>I<span class=\"token punctuation\">]<\/span>   Model: sine_model.onnx<br \/>\n<span class=\"token punctuation\">[<\/span>I<span class=\"token punctuation\">]<\/span> <span class=\"token operator\">&#061;&#061;<\/span><span class=\"token operator\">&#061;<\/span> Build Options <span class=\"token operator\">&#061;&#061;<\/span><span class=\"token operator\">&#061;<\/span><br \/>\n<span class=\"token punctuation\">[<\/span>I<span class=\"token punctuation\">]<\/span>   Precision: FP32<br \/>\n<span class=\"token punctuation\">[<\/span>I<span class=\"token punctuation\">]<\/span> <span class=\"token punctuation\">[<\/span>MemUsageChange<span class=\"token punctuation\">]<\/span> Init CUDA: CPU &#043;0, GPU &#043;0, now: CPU <span class=\"token number\">100<\/span>, GPU <span class=\"token number\">300<\/span> <span class=\"token punctuation\">(<\/span>MiB<span class=\"token punctuation\">)<\/span><br \/>\n<span class=\"token punctuation\">[<\/span>I<span class=\"token punctuation\">]<\/span> <span class=\"token punctuation\">[<\/span>MemUsageChange<span class=\"token punctuation\">]<\/span> Init builder: CPU &#043;0, GPU &#043;0, now: CPU <span class=\"token number\">100<\/span>, GPU <span class=\"token number\">400<\/span> <span class=\"token punctuation\">(<\/span>MiB<span class=\"token punctuation\">)<\/span><br \/>\n<span class=\"token punctuation\">[<\/span>I<span class=\"token punctuation\">]<\/span> <span class=\"token punctuation\">[<\/span>MemUsageChange<span class=\"token punctuation\">]<\/span> Init cuBLAS\/cuDNN: CPU &#043;0, GPU &#043;0, now: CPU <span class=\"token number\">200<\/span>, GPU <span class=\"token number\">500<\/span> <span class=\"token punctuation\">(<\/span>MiB<span class=\"token punctuation\">)<\/span><br \/>\n<span class=\"token punctuation\">[<\/span>I<span class=\"token punctuation\">]<\/span> Engine built <span class=\"token keyword\">in<\/span> <span class=\"token number\">0.5<\/span> sec.<br \/>\n<span class=\"token punctuation\">[<\/span>I<span class=\"token punctuation\">]<\/span> <span class=\"token punctuation\">[<\/span>MemUsageChange<span class=\"token punctuation\">]<\/span> Init deserializing engine: CPU &#043;0, GPU &#043;0, now: CPU <span class=\"token number\">100<\/span>, GPU <span class=\"token number\">500<\/span> <span class=\"token punctuation\">(<\/span>MiB<span class=\"token punctuation\">)<\/span><br \/>\n<span class=\"token punctuation\">[<\/span>I<span class=\"token punctuation\">]<\/span> <span class=\"token operator\">&#061;&#061;<\/span><span class=\"token operator\">&#061;<\/span> Performance summary <span class=\"token operator\">&#061;&#061;<\/span><span class=\"token operator\">&#061;<\/span><br \/>\n<span class=\"token punctuation\">[<\/span>I<span class=\"token punctuation\">]<\/span> Throughput: <span class=\"token number\">12034.5<\/span> qps<br \/>\n<span class=\"token punctuation\">[<\/span>I<span class=\"token punctuation\">]<\/span> Latency: min <span class=\"token operator\">&#061;<\/span> <span class=\"token number\">0.082<\/span> ms, max <span class=\"token operator\">&#061;<\/span> <span class=\"token number\">0.115<\/span> ms, mean <span class=\"token operator\">&#061;<\/span> <span class=\"token number\">0.091<\/span> ms<br \/>\n<span class=\"token punctuation\">[<\/span>I<span class=\"token punctuation\">]<\/span> End-to-End Host Latency: min <span class=\"token operator\">&#061;<\/span> <span class=\"token number\">0.095<\/span> ms, max <span class=\"token operator\">&#061;<\/span> <span class=\"token number\">0.130<\/span> ms, mean <span class=\"token operator\">&#061;<\/span> <span class=\"token number\">0.105<\/span> ms<br \/>\n<span class=\"token punctuation\">[<\/span>I<span class=\"token punctuation\">]<\/span> Enqueue Time: min <span class=\"token operator\">&#061;<\/span> <span class=\"token number\">0.070<\/span> ms, max <span class=\"token operator\">&#061;<\/span> <span class=\"token number\">0.100<\/span> ms, mean <span class=\"token operator\">&#061;<\/span> <span class=\"token number\">0.080<\/span> ms<br \/>\n<span class=\"token punctuation\">[<\/span>I<span class=\"token punctuation\">]<\/span> Total Host Walltime: <span class=\"token number\">3.00<\/span> s<\/p>\n<p>Step C \u2014 \u6784\u5efa FP16 Engine<\/p>\n<p>FP16 \u9700\u8981 GPU \u652f\u6301&#xff08;\u51e0\u4e4e\u6240\u6709 NVIDIA GPU \u90fd\u652f\u6301&#xff09;&#xff1a;<\/p>\n<p>trtexec <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token parameter variable\">&#8211;onnx<\/span><span class=\"token operator\">&#061;<\/span>sine_model.onnx <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token parameter variable\">&#8211;fp16<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token parameter variable\">&#8211;saveEngine<\/span><span class=\"token operator\">&#061;<\/span>sine_fp16.engine<\/p>\n<p>\u9884\u671f\u8f93\u51fa\u7247\u6bb5&#xff08;\u5173\u952e\u884c&#xff09;&#xff1a;<\/p>\n<p><span class=\"token punctuation\">[<\/span>I<span class=\"token punctuation\">]<\/span> <span class=\"token operator\">&#061;&#061;<\/span><span class=\"token operator\">&#061;<\/span> Build Options <span class=\"token operator\">&#061;&#061;<\/span><span class=\"token operator\">&#061;<\/span><br \/>\n<span class=\"token punctuation\">[<\/span>I<span class=\"token punctuation\">]<\/span>   Precision: FP16<br \/>\n<span class=\"token punctuation\">[<\/span>I<span class=\"token punctuation\">]<\/span>   FP16: Enabled<br \/>\n<span class=\"token punctuation\">[<\/span>I<span class=\"token punctuation\">]<\/span> Engine built <span class=\"token keyword\">in<\/span> <span class=\"token number\">0.4<\/span> sec.<br \/>\n<span class=\"token punctuation\">[<\/span>I<span class=\"token punctuation\">]<\/span> <span class=\"token operator\">&#061;&#061;<\/span><span class=\"token operator\">&#061;<\/span> Performance summary <span class=\"token operator\">&#061;&#061;<\/span><span class=\"token operator\">&#061;<\/span><br \/>\n<span class=\"token punctuation\">[<\/span>I<span class=\"token punctuation\">]<\/span> Throughput: <span class=\"token number\">21087.3<\/span> qps<br \/>\n<span class=\"token punctuation\">[<\/span>I<span class=\"token punctuation\">]<\/span> Latency: min <span class=\"token operator\">&#061;<\/span> <span class=\"token number\">0.045<\/span> ms, max <span class=\"token operator\">&#061;<\/span> <span class=\"token number\">0.068<\/span> ms, mean <span class=\"token operator\">&#061;<\/span> <span class=\"token number\">0.050<\/span> ms<\/p>\n<p>\u53ef\u4ee5\u770b\u5230 FP16 \u7684\u541e\u5410\u91cf\u660e\u663e\u9ad8\u4e8e FP32\u3002<\/p>\n<p>Step D \u2014 \u6784\u5efa INT8 Engine<\/p>\n<p>INT8 \u9700\u8981 \u6821\u51c6\u6570\u636e&#xff08;Calibration&#xff09;&#xff0c;trtexec \u9ed8\u8ba4\u4f1a\u751f\u6210\u968f\u673a\u8f93\u5165\u4f5c\u4e3a\u6821\u51c6\u6570\u636e&#xff1a;<\/p>\n<p>trtexec <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token parameter variable\">&#8211;onnx<\/span><span class=\"token operator\">&#061;<\/span>sine_model.onnx <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token parameter variable\">&#8211;int8<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token parameter variable\">&#8211;calib<\/span><span class=\"token operator\">&#061;<\/span>calibration_data <span class=\"token punctuation\">\\\\<\/span><br \/>\n<span class=\"token parameter variable\">&#8211;saveEngine<\/span><span class=\"token operator\">&#061;<\/span>sine_int8.engine<\/p>\n<p>\u8bf4\u660e&#xff1a;&#8211;calib \u6307\u5b9a\u6821\u51c6\u7f13\u5b58\u6587\u4ef6\u3002\u5bf9\u4e8e Sine \u8fd9\u79cd\u7b80\u5355\u6a21\u578b&#xff0c;\u968f\u673a\u6821\u51c6\u6570\u636e\u901a\u5e38\u8db3\u591f&#xff1b;\u771f\u5b9e\u9879\u76ee\u4e2d\u5efa\u8bae\u7528\u4ee3\u8868\u6027\u6570\u636e\u96c6\u3002<\/p>\n<p>\u9884\u671f\u8f93\u51fa\u7247\u6bb5&#xff08;\u5173\u952e\u884c&#xff09;&#xff1a;<\/p>\n<p><span class=\"token punctuation\">[<\/span>I<span class=\"token punctuation\">]<\/span> <span class=\"token operator\">&#061;&#061;<\/span><span class=\"token operator\">&#061;<\/span> Build Options <span class=\"token operator\">&#061;&#061;<\/span><span class=\"token operator\">&#061;<\/span><br \/>\n<span class=\"token punctuation\">[<\/span>I<span class=\"token punctuation\">]<\/span>   Precision: INT8<br \/>\n<span class=\"token punctuation\">[<\/span>I<span class=\"token punctuation\">]<\/span>   INT8: Enabled<br \/>\n<span class=\"token punctuation\">[<\/span>I<span class=\"token punctuation\">]<\/span> <span class=\"token punctuation\">[<\/span>TRT<span class=\"token punctuation\">]<\/span> <span class=\"token punctuation\">[<\/span>MemUsageChange<span class=\"token punctuation\">]<\/span> Init cuDNN: CPU &#043;0, GPU &#043;0, now: CPU <span class=\"token number\">200<\/span>, GPU <span class=\"token number\">600<\/span> <span class=\"token punctuation\">(<\/span>MiB<span class=\"token punctuation\">)<\/span><br \/>\n<span class=\"token punctuation\">[<\/span>I<span class=\"token punctuation\">]<\/span> <span class=\"token punctuation\">[<\/span>TRT<span class=\"token punctuation\">]<\/span> Detected <span class=\"token number\">1<\/span> inputs and <span class=\"token number\">1<\/span> outputs <span class=\"token keyword\">in<\/span> the network.<br \/>\n<span class=\"token punctuation\">[<\/span>I<span class=\"token punctuation\">]<\/span> <span class=\"token punctuation\">[<\/span>TRT<span class=\"token punctuation\">]<\/span> Total Host Persistent Memory: <span class=\"token number\">1000<\/span><br \/>\n<span class=\"token punctuation\">[<\/span>I<span class=\"token punctuation\">]<\/span> <span class=\"token punctuation\">[<\/span>TRT<span class=\"token punctuation\">]<\/span> Total Device Persistent Memory: <span class=\"token number\">2000<\/span><br \/>\n<span class=\"token punctuation\">[<\/span>I<span class=\"token punctuation\">]<\/span> Engine built <span class=\"token keyword\">in<\/span> <span class=\"token number\">0.6<\/span> sec.<br \/>\n<span class=\"token punctuation\">[<\/span>I<span class=\"token punctuation\">]<\/span> <span class=\"token operator\">&#061;&#061;<\/span><span class=\"token operator\">&#061;<\/span> Performance summary <span class=\"token operator\">&#061;&#061;<\/span><span class=\"token operator\">&#061;<\/span><br \/>\n<span class=\"token punctuation\">[<\/span>I<span class=\"token punctuation\">]<\/span> Throughput: <span class=\"token number\">35672.8<\/span> qps<br \/>\n<span class=\"token punctuation\">[<\/span>I<span class=\"token punctuation\">]<\/span> Latency: min <span class=\"token operator\">&#061;<\/span> <span class=\"token number\">0.028<\/span> ms, max <span class=\"token operator\">&#061;<\/span> <span class=\"token number\">0.045<\/span> ms, mean <span class=\"token operator\">&#061;<\/span> <span class=\"token number\">0.032<\/span> ms<\/p>\n<p>Step E \u2014 \u5bf9\u6bd4\u4e09\u79cd\u7cbe\u5ea6\u7684\u6027\u80fd<\/p>\n<p>\u5206\u522b\u52a0\u8f7d\u4e09\u4e2a Engine \u505a\u6027\u80fd\u57fa\u51c6\u6d4b\u8bd5&#xff08;benchmark&#xff09;&#xff1a;<\/p>\n<p>trtexec <span class=\"token parameter variable\">&#8211;loadEngine<\/span><span class=\"token operator\">&#061;<\/span>sine_fp32.engine<br \/>\ntrtexec <span class=\"token parameter variable\">&#8211;loadEngine<\/span><span class=\"token operator\">&#061;<\/span>sine_fp16.engine<br \/>\ntrtexec <span class=\"token parameter variable\">&#8211;loadEngine<\/span><span class=\"token operator\">&#061;<\/span>sine_int8.engine<\/p>\n<p>\u6574\u7406\u6210\u5bf9\u6bd4\u8868&#xff1a;<\/p>\n<p>\u7cbe\u5ea6        Throughput <span class=\"token punctuation\">(<\/span>qps<span class=\"token punctuation\">)<\/span>    Latency <span class=\"token punctuation\">(<\/span>ms<span class=\"token punctuation\">)<\/span><br \/>\nFP32        <span class=\"token number\">12034.5<\/span>             <span class=\"token number\">0.091<\/span><br \/>\nFP16        <span class=\"token number\">21087.3<\/span>             <span class=\"token number\">0.050<\/span><br \/>\nINT8        <span class=\"token number\">35672.8<\/span>             <span class=\"token number\">0.032<\/span><\/p>\n<p>\u4e0b\u9762\u628a\u4e09\u79cd\u7cbe\u5ea6\u7684\u5173\u952e\u6307\u6807\u6c47\u603b\u6210\u4e00\u5f20\u66f4\u5b8c\u6574\u7684\u5bf9\u6bd4\u8868&#xff1a;<\/p>\n<table>\n<tr>\u7cbe\u5ea6Throughput (qps)Latency (ms)\u6a21\u578b\u5927\u5c0f\u7cbe\u5ea6\u635f\u5931&#xff08;MSE&#xff09;<\/tr>\n<tbody>\n<tr>\n<td>FP32<\/td>\n<td>12034.5<\/td>\n<td>0.091<\/td>\n<td>2.3 KB<\/td>\n<td>0&#xff08;\u57fa\u51c6&#xff09;<\/td>\n<\/tr>\n<tr>\n<td>FP16<\/td>\n<td>21087.3<\/td>\n<td>0.050<\/td>\n<td>1.2 KB<\/td>\n<td>~1e-7<\/td>\n<\/tr>\n<tr>\n<td>INT8<\/td>\n<td>35672.8<\/td>\n<td>0.032<\/td>\n<td>0.6 KB<\/td>\n<td>~1e-4<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u8bf4\u660e&#xff1a;\u6a21\u578b\u5927\u5c0f\u6309 Sine \u6a21\u578b\u4f30\u7b97&#xff08;FP16 \u7ea6\u4e3a FP32 \u7684\u4e00\u534a&#xff0c;INT8 \u7ea6\u4e3a FP32 \u7684\u56db\u5206\u4e4b\u4e00&#xff09;&#xff1b;MSE \u4e3a\u76f8\u5bf9 FP32 \u8f93\u51fa\u7684\u5747\u65b9\u8bef\u5dee\u91cf\u7ea7&#xff0c;\u5b9e\u9645\u6570\u503c\u4f1a\u56e0\u6821\u51c6\u6570\u636e\u96c6\u4e0d\u540c\u800c\u7565\u6709\u5dee\u5f02\u3002<\/p>\n<p>\u4e09\u79cd\u7cbe\u5ea6\u7684\u9002\u7528\u573a\u666f\u5206\u6790<\/p>\n<ul>\n<li>\n<p>FP32&#xff1a;\u7cbe\u5ea6\u6700\u9ad8\u3001\u517c\u5bb9\u6027\u6700\u597d&#xff0c;\u9002\u5408\u5bf9\u6570\u503c\u7cbe\u5ea6\u8981\u6c42\u82db\u523b\u3001\u4e14\u5bf9\u5ef6\u8fdf\u4e0d\u654f\u611f\u7684\u573a\u666f&#xff0c;\u4f8b\u5982\u6a21\u578b\u6b63\u786e\u6027\u9a8c\u8bc1\u3001\u8c03\u8bd5\u57fa\u51c6\u3002\u5b83\u662f\u6240\u6709\u7cbe\u5ea6\u7684&#034;\u6807\u51c6\u7b54\u6848&#034;&#xff0c;\u7528\u6765\u8861\u91cf\u5176\u4ed6\u7cbe\u5ea6\u7684\u8bef\u5dee\u3002<\/p>\n<\/li>\n<li>\n<p>FP16&#xff1a;\u7cbe\u5ea6\u635f\u5931\u51e0\u4e4e\u53ef\u4ee5\u5ffd\u7565&#xff08;MSE \u7ea6 1e-7&#xff09;&#xff0c;\u4f46\u541e\u5410\u91cf\u63d0\u5347\u7ea6 1.7 \u500d\u3001\u663e\u5b58\u5360\u7528\u51cf\u534a\u3002\u9002\u5408\u7edd\u5927\u591a\u6570\u751f\u4ea7\u73af\u5883&#xff0c;\u5c24\u5176\u662f\u5b9e\u65f6\u63a8\u7406\u3001\u5728\u7ebf\u670d\u52a1\u7b49\u5bf9\u5ef6\u8fdf\u548c\u541e\u5410\u90fd\u6709\u8981\u6c42\u7684\u573a\u666f&#xff0c;\u662f&#034;\u6027\u4ef7\u6bd4&#034;\u6700\u9ad8\u7684\u9009\u62e9\u3002<\/p>\n<\/li>\n<li>\n<p>INT8&#xff1a;\u541e\u5410\u91cf\u63d0\u5347\u7ea6 3 \u500d\u3001\u6a21\u578b\u4f53\u79ef\u538b\u7f29\u5230 1\/4&#xff0c;\u4f46\u7cbe\u5ea6\u635f\u5931\u76f8\u5bf9\u660e\u663e&#xff08;MSE \u7ea6 1e-4&#xff09;\u3002\u9002\u5408\u5bf9\u7cbe\u5ea6\u5bb9\u5fcd\u5ea6\u8f83\u9ad8\u3001\u5bf9\u541e\u5410\u91cf\u548c\u529f\u8017\u6781\u5ea6\u654f\u611f\u7684\u573a\u666f&#xff0c;\u4f8b\u5982\u5927\u89c4\u6a21\u6279\u91cf\u63a8\u7406\u3001\u8fb9\u7f18\u8bbe\u5907\u3001\u89c6\u9891\u6d41\u5904\u7406\u7b49\u3002\u5bf9\u4e8e Sine \u8fd9\u79cd\u7b80\u5355\u6a21\u578b&#xff0c;INT8 \u7684\u8bef\u5dee\u51e0\u4e4e\u4e0d\u5f71\u54cd\u7ed3\u679c&#xff1b;\u4f46\u5bf9\u4e8e\u590d\u6742\u6a21\u578b&#xff08;\u5982\u68c0\u6d4b\u3001\u5206\u5272\u7f51\u7edc&#xff09;&#xff0c;\u9700\u8981\u5148\u7528\u4ee3\u8868\u6027\u6570\u636e\u96c6\u6821\u51c6&#xff0c;\u5e76\u9a8c\u8bc1\u7cbe\u5ea6\u662f\u5426\u5728\u53ef\u63a5\u53d7\u8303\u56f4\u5185\u3002<\/p>\n<\/li>\n<\/ul>\n<p>\u9009\u62e9\u5efa\u8bae&#xff1a;\u5148\u7528 FP32 \u8dd1\u901a\u6d41\u7a0b\u5e76\u4f5c\u4e3a\u7cbe\u5ea6\u57fa\u51c6&#xff0c;\u518d\u5c1d\u8bd5 FP16 \u83b7\u5f97\u6027\u80fd\u63d0\u5347&#xff1b;\u53ea\u6709\u5f53 FP16 \u4ecd\u4e0d\u6ee1\u8db3\u541e\u5410\u8981\u6c42\u3001\u4e14\u7cbe\u5ea6\u635f\u5931\u53ef\u63a5\u53d7\u65f6&#xff0c;\u624d\u8fdb\u4e00\u6b65\u4f7f\u7528 INT8\u3002<\/p>\n<p>\u53ef\u4ee5\u770b\u5230&#xff1a;FP16 \u6bd4 FP32 \u5feb\u7ea6 1.7 \u500d&#xff0c;INT8 \u6bd4 FP32 \u5feb\u7ea6 3 \u500d\u3002\u8fd9\u5c31\u662f TensorRT \u7684\u4ef7\u503c\u3002<\/p>\n<p>\u5e38\u89c1\u62a5\u9519\u6392\u67e5<\/p>\n<p>\u62a5\u9519 1&#xff1a;\u627e\u4e0d\u5230 ONNX \u6587\u4ef6<\/p>\n<p><span class=\"token punctuation\">[<\/span>E<span class=\"token punctuation\">]<\/span> Failed to parse ONNX file: sine_model.onnx<br \/>\n<span class=\"token punctuation\">[<\/span>E<span class=\"token punctuation\">]<\/span> Error: File not found<\/p>\n<p>\u6392\u67e5\u65b9\u6cd5&#xff1a;<\/p>\n<p><span class=\"token comment\"># \u786e\u8ba4\u6587\u4ef6\u5b58\u5728<\/span><br \/>\n<span class=\"token function\">ls<\/span> <span class=\"token parameter variable\">-lh<\/span> sine_model.onnx<\/p>\n<p><span class=\"token comment\"># \u786e\u8ba4\u8def\u5f84\u6b63\u786e&#xff08;\u7528\u7edd\u5bf9\u8def\u5f84\u6700\u7a33\u59a5&#xff09;<\/span><br \/>\ntrtexec <span class=\"token parameter variable\">&#8211;onnx<\/span><span class=\"token operator\">&#061;<\/span><span class=\"token variable\"><span class=\"token variable\">$(<\/span><span class=\"token builtin class-name\">pwd<\/span><span class=\"token variable\">)<\/span><\/span>\/sine_model.onnx<\/p>\n<p>\u62a5\u9519 2&#xff1a;GPU \u4e0d\u652f\u6301 FP16<\/p>\n<p><span class=\"token punctuation\">[<\/span>E<span class=\"token punctuation\">]<\/span> Could not <span class=\"token keyword\">select<\/span> implementation <span class=\"token keyword\">for<\/span> node: <span class=\"token punctuation\">[<\/span>Dense<span class=\"token punctuation\">]<\/span><br \/>\n<span class=\"token punctuation\">[<\/span>E<span class=\"token punctuation\">]<\/span> Error: No implementation found <span class=\"token keyword\">for<\/span> <span class=\"token function\">node<\/span><\/p>\n<p>\u6392\u67e5\u65b9\u6cd5&#xff1a;<\/p>\n<p><span class=\"token comment\"># \u67e5\u770b GPU \u7b97\u529b\u662f\u5426\u652f\u6301 FP16<\/span><br \/>\nnvidia-smi &#8211;query-gpu<span class=\"token operator\">&#061;<\/span>name,compute_cap <span class=\"token parameter variable\">&#8211;format<\/span><span class=\"token operator\">&#061;<\/span>csv<\/p>\n<p><span class=\"token comment\"># \u5982\u679c\u7b97\u529b\u4f4e\u4e8e 5.3&#xff0c;\u53bb\u6389 &#8211;fp16&#xff0c;\u53ea\u7528 FP32<\/span><\/p>\n<p>\u62a5\u9519 3&#xff1a;INT8 \u6821\u51c6\u5931\u8d25<\/p>\n<p><span class=\"token punctuation\">[<\/span>E<span class=\"token punctuation\">]<\/span> <span class=\"token punctuation\">[<\/span>TRT<span class=\"token punctuation\">]<\/span> Calibration failed: no calibration data<\/p>\n<p>\u6392\u67e5\u65b9\u6cd5&#xff1a;<\/p>\n<p><span class=\"token comment\"># \u663e\u5f0f\u6307\u5b9a\u6821\u51c6\u6570\u636e\u6587\u4ef6<\/span><br \/>\ntrtexec <span class=\"token parameter variable\">&#8211;onnx<\/span><span class=\"token operator\">&#061;<\/span>sine_model.onnx <span class=\"token parameter variable\">&#8211;int8<\/span> <span class=\"token parameter variable\">&#8211;calib<\/span><span class=\"token operator\">&#061;<\/span>calibration_data<\/p>\n<p><span class=\"token comment\"># \u6216\u8005\u7528 &#8211;calib \u6307\u5b9a\u4e00\u4e2a\u76ee\u5f55&#xff0c;trtexec \u4f1a\u8bfb\u53d6\u5176\u4e2d\u7684\u56fe\u7247\/\u6570\u636e<\/span><\/p>\n<p>\u62a5\u9519 4&#xff1a;\u663e\u5b58\u4e0d\u8db3&#xff08;OOM&#xff09;<\/p>\n<p><span class=\"token punctuation\">[<\/span>E<span class=\"token punctuation\">]<\/span> <span class=\"token punctuation\">[<\/span>TRT<span class=\"token punctuation\">]<\/span> CUDA Error: out of memory<\/p>\n<p>\u6392\u67e5\u65b9\u6cd5&#xff1a;<\/p>\n<p><span class=\"token comment\"># \u67e5\u770b\u5f53\u524d\u663e\u5b58\u5360\u7528<\/span><br \/>\nnvidia-smi<\/p>\n<p><span class=\"token comment\"># \u5173\u95ed\u5176\u4ed6\u5360\u7528\u663e\u5b58\u7684\u8fdb\u7a0b&#xff0c;\u6216\u51cf\u5c0f batch size<\/span><br \/>\ntrtexec <span class=\"token parameter variable\">&#8211;onnx<\/span><span class=\"token operator\">&#061;<\/span>sine_model.onnx <span class=\"token parameter variable\">&#8211;batch<\/span><span class=\"token operator\">&#061;<\/span><span class=\"token number\">1<\/span><\/p>\n<p>\u62a5\u9519 5&#xff1a;\u7248\u672c\u4e0d\u5339\u914d<\/p>\n<p><span class=\"token punctuation\">[<\/span>E<span class=\"token punctuation\">]<\/span> <span class=\"token punctuation\">[<\/span>TRT<span class=\"token punctuation\">]<\/span> Engine version mismatch: engine built with TensorRT <span class=\"token number\">8.6<\/span>, current version <span class=\"token number\">8.5<\/span><\/p>\n<p>\u6392\u67e5\u65b9\u6cd5&#xff1a;<\/p>\n<p><span class=\"token comment\"># \u67e5\u770b\u5f53\u524d TensorRT \u7248\u672c<\/span><br \/>\ntrtexec <span class=\"token parameter variable\">&#8211;version<\/span><\/p>\n<p><span class=\"token comment\"># \u91cd\u65b0\u6784\u5efa engine&#xff08;engine \u662f\u7ed1\u5b9a GPU \u548c TensorRT \u7248\u672c\u7684&#xff0c;\u6362\u73af\u5883\u5fc5\u987b\u91cd\u65b0 build&#xff09;<\/span><\/p>\n<p>\u8fd9\u51e0\u4e2a\u547d\u4ee4\u975e\u5e38\u503c\u5f97\u638c\u63e1\u3002<\/p>\n<p>5.9 TensorRT Python API<\/p>\n<p>\u4e4b\u540e\u6211\u4eec\u518d\u5b66\u4e60 Python API&#xff1a;<\/p>\n<p><span class=\"token function\">import<\/span> tensorrt as <span class=\"token function\">tr<\/span><\/p>\n<p>\u5927\u81f4\u6d41\u7a0b&#xff1a;<\/p>\n<p>logger <span class=\"token operator\">&#061;<\/span> tr<span class=\"token punctuation\">.<\/span>Logger<span class=\"token punctuation\">(<\/span>tr<span class=\"token punctuation\">.<\/span>Logger<span class=\"token punctuation\">.<\/span>WARNING<span class=\"token punctuation\">)<\/span><\/p>\n<p>builder <span class=\"token operator\">&#061;<\/span> tr<span class=\"token punctuation\">.<\/span>Builder<span class=\"token punctuation\">(<\/span>logger<span class=\"token punctuation\">)<\/span><\/p>\n<p>network <span class=\"token operator\">&#061;<\/span> builder<span class=\"token punctuation\">.<\/span>create_network<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">)<\/span><\/p>\n<p>parser <span class=\"token operator\">&#061;<\/span> tr<span class=\"token punctuation\">.<\/span>OnnxParser<span class=\"token punctuation\">(<\/span>network<span class=\"token punctuation\">,<\/span> logger<span class=\"token punctuation\">)<\/span><\/p>\n<p>parser<span class=\"token punctuation\">.<\/span>parse_from_file<span class=\"token punctuation\">(<\/span><span class=\"token string\">&#034;model.onnx&#034;<\/span><span class=\"token punctuation\">)<\/span><\/p>\n<p>config <span class=\"token operator\">&#061;<\/span> builder<span class=\"token punctuation\">.<\/span>create_builder_config<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">)<\/span><\/p>\n<p>engine <span class=\"token operator\">&#061;<\/span> builder<span class=\"token punctuation\">.<\/span>build_serialized_network<span class=\"token punctuation\">(<\/span><br \/>\nnetwork<span class=\"token punctuation\">,<\/span><br \/>\nconfig<br \/>\n<span class=\"token punctuation\">)<\/span><\/p>\n<p>\u6700\u7ec8&#xff1a;<\/p>\n<p>model<span class=\"token punctuation\">.<\/span>onnx<br \/>\n\u2193<br \/>\nOnnxParser<br \/>\n\u2193<br \/>\nNetwork<br \/>\n\u2193<br \/>\nBuilder<br \/>\n\u2193<br \/>\nEngine<\/p>\n<p>\u4e0d\u8fc7\u7b2c\u4e00\u904d\u4e0d\u5efa\u8bae\u76f4\u63a5\u4ece Python API \u5f00\u59cb\u3002<\/p>\n<p>\u5148\u7528&#xff1a;<\/p>\n<p>trtexec<\/p>\n<p>\u628a\u6574\u4e2a pipeline \u8dd1\u901a&#xff0c;\u518d\u5b66\u4e60 API&#xff0c;\u4f1a\u5bb9\u6613\u5f88\u591a\u3002<\/p>\n<p>5.10 \u6211\u5efa\u8bae\u4f60\u7684 Step 5 \u5206\u6210 6 \u4e2a\u5c0f\u6b65\u9aa4<\/p>\n<p>\u6211\u4eec\u7ee7\u7eed\u91c7\u7528\u4f60\u4e4b\u524d\u7684\u5b66\u4e60\u65b9\u5f0f&#xff0c;\u4e00\u6b65\u4e00\u6b65\u505a\u3002<\/p>\n<p>Step 5.1 \u2014 \u51c6\u5907 NVIDIA TensorRT \u73af\u5883<\/p>\n<p>Ubuntu<br \/>\n\u2193<br \/>\nNVIDIA Driver<br \/>\n\u2193<br \/>\nCUDA<br \/>\n\u2193<br \/>\nTensorRT<br \/>\n\u2193<br \/>\ntrtexec<\/p>\n<p>Step 5.2 \u2014 ONNX \u2192 TensorRT FP32<\/p>\n<p>model<span class=\"token punctuation\">.<\/span>onnx<br \/>\n\u2193<br \/>\ntrtexec<br \/>\n\u2193<br \/>\nmodel_fp32<span class=\"token punctuation\">.<\/span>engine<\/p>\n<p>\u5e76\u6d4b\u8bd5&#xff1a;<\/p>\n<p>latency<br \/>\nthroughput<br \/>\nGPU utilization<\/p>\n<p>Step 5.3 \u2014 ONNX \u2192 TensorRT FP16<\/p>\n<p>model<span class=\"token punctuation\">.<\/span>onnx<br \/>\n\u2193<br \/>\nTensorRT<br \/>\n\u2193<br \/>\nFP16<br \/>\n\u2193<br \/>\nmodel_fp16<span class=\"token punctuation\">.<\/span>engine<\/p>\n<p>\u7136\u540e\u548c FP32 \u6bd4\u8f83\u3002<\/p>\n<p>Step 5.4 \u2014 ONNX \u2192 TensorRT INT8<\/p>\n<p>model<span class=\"token punctuation\">.<\/span>onnx<br \/>\n\u2193<br \/>\nINT8<br \/>\n\u2193<br \/>\nCalibration<br \/>\n\u2193<br \/>\nINT8 Engine<\/p>\n<p>\u8fd9\u662f\u672c Step \u6700\u91cd\u8981\u7684\u90e8\u5206\u3002<\/p>\n<p>Step 5.5 \u2014 TensorRT Python API<\/p>\n<p>\u5b66\u4e60&#xff1a;<\/p>\n<p>Builder<br \/>\nNetwork<br \/>\nParser<br \/>\nConfig<br \/>\nEngine<br \/>\nRuntime<br \/>\nContext<br \/>\nTensor<\/p>\n<p>Step 5.6 \u2014 Benchmark<\/p>\n<p>\u6700\u7ec8\u5f62\u6210&#xff1a; <img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260910035705-6aa22a9171435.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>\u8fd9\u4e00\u6b65\u5b8c\u6210\u540e&#xff0c;\u4f60\u5c31\u771f\u6b63\u628a&#xff1a;<\/p>\n<p>ONNX \u2192 Runtime \u2192 Quantization \u2192 GPU inference<\/p>\n<p>\u8fd9\u4e00\u6574\u6761\u94fe\u4e32\u8d77\u6765\u4e86\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6458\u8981&#xff1a;\u672c\u6587\u662f ONNX \u90e8\u7f72\u5b66\u4e60\u8def\u7ebf\u7684 Step 5&#xff0c;\u8bb2\u89e3\u5982\u4f55\u5c06 ONNX \u6a21\u578b\u8f6c\u6362\u4e3a TensorRT Engine \u5e76\u5728 NVIDIA GPU \u4e0a\u9ad8\u6548\u63a8\u7406\u3002\u5185\u5bb9\u6db5\u76d6 TensorRT \u4e0e ONNX Runtime \u7684\u5b9a\u4f4d\u5dee\u5f02\u3001Engine \u6838\u5fc3\u6982\u5ff5\u3001FP32\/FP16\/INT8 \u7cbe\u5ea6\u652f\u6301\u3001\u4e0e TFLM INT8 \u7684\u533a\u522b\u3001MacBook Air \u7684\u786c\u4ef6\u9650\u5236&#xff0c;\u4ee5\u53ca\u4ece trtexec \u547d\u4ee4\u884c\u5de5\u5177\u5230 Python API \u7684\u5b8c\u6574\u5b9e\u8df5\u8def\u5f84&amp;<\/p>\n","protected":false},"author":2,"featured_media":103393,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[371],"topic":[],"class_list":["post-103396","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-371"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v20.3 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>\u8f7b\u677e\u5b66\u4e60ONNX_day5 - \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\/103396.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u8f7b\u677e\u5b66\u4e60ONNX_day5 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u6458\u8981&#xff1a;\u672c\u6587\u662f ONNX \u90e8\u7f72\u5b66\u4e60\u8def\u7ebf\u7684 Step 5&#xff0c;\u8bb2\u89e3\u5982\u4f55\u5c06 ONNX \u6a21\u578b\u8f6c\u6362\u4e3a TensorRT Engine \u5e76\u5728 NVIDIA GPU \u4e0a\u9ad8\u6548\u63a8\u7406\u3002\u5185\u5bb9\u6db5\u76d6 TensorRT \u4e0e ONNX Runtime \u7684\u5b9a\u4f4d\u5dee\u5f02\u3001Engine \u6838\u5fc3\u6982\u5ff5\u3001FP32\/FP16\/INT8 \u7cbe\u5ea6\u652f\u6301\u3001\u4e0e TFLM INT8 \u7684\u533a\u522b\u3001MacBook Air \u7684\u786c\u4ef6\u9650\u5236&#xff0c;\u4ee5\u53ca\u4ece trtexec \u547d\u4ee4\u884c\u5de5\u5177\u5230 Python API \u7684\u5b8c\u6574\u5b9e\u8df5\u8def\u5f84&amp;\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/103396.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-09-10T03:57:06+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260910035705-6aa22a91466bc.png\" \/>\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=\"6 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/103396.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/103396.html\",\"name\":\"\u8f7b\u677e\u5b66\u4e60ONNX_day5 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-09-10T03:57:06+00:00\",\"dateModified\":\"2026-09-10T03:57:06+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/103396.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/103396.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/103396.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u8f7b\u677e\u5b66\u4e60ONNX_day5\"}]},{\"@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":"\u8f7b\u677e\u5b66\u4e60ONNX_day5 - \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\/103396.html","og_locale":"zh_CN","og_type":"article","og_title":"\u8f7b\u677e\u5b66\u4e60ONNX_day5 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u6458\u8981&#xff1a;\u672c\u6587\u662f ONNX \u90e8\u7f72\u5b66\u4e60\u8def\u7ebf\u7684 Step 5&#xff0c;\u8bb2\u89e3\u5982\u4f55\u5c06 ONNX \u6a21\u578b\u8f6c\u6362\u4e3a TensorRT Engine \u5e76\u5728 NVIDIA GPU \u4e0a\u9ad8\u6548\u63a8\u7406\u3002\u5185\u5bb9\u6db5\u76d6 TensorRT \u4e0e ONNX Runtime \u7684\u5b9a\u4f4d\u5dee\u5f02\u3001Engine \u6838\u5fc3\u6982\u5ff5\u3001FP32\/FP16\/INT8 \u7cbe\u5ea6\u652f\u6301\u3001\u4e0e TFLM INT8 \u7684\u533a\u522b\u3001MacBook Air \u7684\u786c\u4ef6\u9650\u5236&#xff0c;\u4ee5\u53ca\u4ece trtexec \u547d\u4ee4\u884c\u5de5\u5177\u5230 Python API \u7684\u5b8c\u6574\u5b9e\u8df5\u8def\u5f84&amp;","og_url":"https:\/\/www.wsisp.com\/helps\/103396.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-09-10T03:57:06+00:00","og_image":[{"url":"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260910035705-6aa22a91466bc.png"}],"author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"6 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/103396.html","url":"https:\/\/www.wsisp.com\/helps\/103396.html","name":"\u8f7b\u677e\u5b66\u4e60ONNX_day5 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-09-10T03:57:06+00:00","dateModified":"2026-09-10T03:57:06+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/103396.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/103396.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/103396.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\u8f7b\u677e\u5b66\u4e60ONNX_day5"}]},{"@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\/103396","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=103396"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/103396\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media\/103393"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=103396"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=103396"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=103396"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=103396"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}