{"id":116228,"date":"2026-10-20T23:32:39","date_gmt":"2026-10-20T15:32:39","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/116228.html"},"modified":"2026-10-20T23:32:39","modified_gmt":"2026-10-20T15:32:39","slug":"%e8%be%b9%e7%bc%98ai%e9%83%a8%e7%bd%b2%e5%ae%9e%e6%88%98%ef%bc%9a%e9%87%8f%e5%8c%96%e4%b8%8e%e5%89%aa%e6%9e%9d%e7%9a%84%e5%86%85%e5%ad%98%e4%bc%98%e5%8c%96%e7%ad%96%e7%95%a5%e6%b7%b1%e5%ba%a6%e5%af%b9","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/116228.html","title":{"rendered":"\u8fb9\u7f18AI\u90e8\u7f72\u5b9e\u6218\uff1a\u91cf\u5316\u4e0e\u526a\u679d\u7684\u5185\u5b58\u4f18\u5316\u7b56\u7565\u6df1\u5ea6\u5bf9\u6bd4"},"content":{"rendered":"<\/p>\n

# \u8fb9\u7f18AI\u90e8\u7f72\u5b9e\u6218：\u91cf\u5316\u4e0e\u526a\u679d\u7684\u5185\u5b58\u4f18\u5316\u7b56\u7565\u6df1\u5ea6\u5bf9\u6bd4<\/p>\n<\/p>\n

\u53bb\u5e74\u5728\u505a\u4e00\u4e2a\u5de5\u4e1a\u8d28\u68c0\u9879\u76ee\u65f6，\u6211\u9047\u5230\u4e86\u4e00\u4e2a\u5178\u578b\u7684\u8fb9\u7f18\u90e8\u7f72\u56f0\u5883：\u5ba2\u6237\u63d0\u4f9b\u7684Jetson Nano\u53ea\u67094GB\u5185\u5b58，\u800c\u6211\u4eec\u7684ResNet-50\u6a21\u578bFP32\u683c\u5f0f\u4e0b\u5c31\u5360\u4e86\u8fd1100MB\u3002\u66f4\u9ebb\u70e6\u7684\u662f，\u6a21\u578b\u52a0\u8f7d\u540e\u7559\u7ed9\u4e1a\u52a1\u903b\u8f91\u548c\u7f13\u5b58\u7684\u5185\u5b58\u6240\u5269\u65e0\u51e0，\u63a8\u7406\u5ef6\u8fdf\u76f4\u63a5\u98d9\u5230800ms\u4ee5\u4e0a，\u5b8c\u5168\u65e0\u6cd5\u6ee1\u8db3\u4ea7\u7ebf\u5b9e\u65f6\u6027\u8981\u6c42\u3002<\/p>\n<\/p>\n

\u90a3\u6b21\u7ecf\u5386\u8ba9\u6211\u6df1\u523b\u610f\u8bc6\u5230，\u8fb9\u7f18AI\u90e8\u7f72\u4e0d\u662f\u7b80\u5355\u7684"\u6a21\u578b\u6253\u5305\u4e0a\u4f20"，\u800c\u662f\u4e00\u573a\u4e0e\u5185\u5b58\u3001\u529f\u8017\u3001\u7b97\u529b\u7684\u6301\u7eed\u535a\u5f08\u3002\u91cf\u5316\u4e0e\u526a\u679d，\u6b63\u662f\u8fd9\u573a\u535a\u5f08\u4e2d\u6700\u6838\u5fc3\u7684\u4e24\u628a\u6b66\u5668\u3002<\/p>\n<\/p>\n

## \u8fb9\u7f18\u90e8\u7f72\u7684\u5185\u5b58\u6311\u6218<\/p>\n<\/p>\n

\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u5728\u8bad\u7ec3\u9636\u6bb5\u901a\u5e38\u4f7f\u7528\u5355\u7cbe\u5ea6\u6d6e\u70b9\u6570（FP32）\u5b58\u50a8\u6743\u91cd\u548c\u6fc0\u6d3b\u503c\u3002\u4ee5ResNet-50\u4e3a\u4f8b，\u7ea62500\u4e07\u53c2\u6570，FP32\u683c\u5f0f\u4e0b\u6a21\u578b\u4f53\u79ef\u63a5\u8fd1100MB\u3002\u5bf9\u4e8e\u914d\u59071GB\u62162GB\u5185\u5b58\u7684\u5d4c\u5165\u5f0f\u8bbe\u5907\u800c\u8a00，\u4ec5\u6a21\u578b\u52a0\u8f7d\u5c31\u4f1a\u5360\u7528\u5927\u91cfRAM\u3002<\/p>\n<\/p>\n

\u5185\u5b58\u53d7\u9650\u76f4\u63a5\u5bfc\u81f4\u4e24\u4e2a\u5de5\u7a0b\u95ee\u9898\u3002\u7b2c\u4e00，\u5185\u5b58\u5e26\u5bbd\u6210\u4e3a\u63a8\u7406\u74f6\u9888\u3002\u5904\u7406\u5668\u4eceDRAM\u8bfb\u53d6\u6a21\u578b\u6743\u91cd\u7684\u901f\u5ea6，\u8fdc\u4f4e\u4e8e\u8ba1\u7b97\u5355\u5143\u7684\u7406\u8bba\u7b97\u529b，\u5bfc\u81f4\u8ba1\u7b97\u5355\u5143\u5904\u4e8e\u95f2\u7f6e\u7b49\u5f85\u72b6\u6001\u3002\u7b2c\u4e8c，\u9ad8\u529f\u8017\u9650\u5236\u7535\u6c60\u7eed\u822a\u3002\u9891\u7e41\u7684\u5185\u5b58\u8bbf\u95ee\u662f\u8fb9\u7f18\u8bbe\u5907\u529f\u8017\u7684\u4e3b\u8981\u6765\u6e90\u4e4b\u4e00\u3002<\/p>\n<\/p>\n

\u91cf\u5316\u4e0e\u526a\u679d\u6b63\u662f\u9488\u5bf9\u4e0a\u8ff0\u75db\u70b9\u8bbe\u8ba1\u7684\u8f6f\u4ef6\u4f18\u5316\u65b9\u6848\u3002\u91cf\u5316\u901a\u8fc7\u964d\u4f4e\u6570\u636e\u7cbe\u5ea6\u6765\u51cf\u5c11\u6bcf\u6b21\u5185\u5b58\u8bbf\u95ee\u7684\u6570\u636e\u91cf；\u526a\u679d\u5219\u901a\u8fc7\u5220\u9664\u5197\u4f59\u8fde\u63a5\u6765\u76f4\u63a5\u51cf\u5c11\u5185\u5b58\u8bbf\u95ee\u7684\u6b21\u6570\u3002<\/p>\n<\/p>\n

## \u6280\u672f\u539f\u7406\u4e0e\u67b6\u6784\u5256\u6790<\/p>\n<\/p>\n

### \u91cf\u5316\u673a\u5236：\u7cbe\u5ea6\u4e0e\u901f\u5ea6\u7684\u6743\u8861<\/p>\n<\/p>\n

\u91cf\u5316\u6280\u672f\u7684\u6838\u5fc3\u5728\u4e8e\u964d\u4f4e\u6a21\u578b\u53c2\u6570\u7684\u6570\u503c\u7cbe\u5ea6\u3002\u5c0632\u4f4d\u6d6e\u70b9\u6570（FP32）\u6620\u5c04\u4e3a8\u4f4d\u6574\u6570（INT8）\u751a\u81f3\u66f4\u4f4e\u7cbe\u5ea6\u7684\u6570\u636e\u7c7b\u578b\u3002<\/p>\n<\/p>\n

**\u5173\u4e8e\u538b\u7f29\u6bd4\u7684\u5b9e\u9645\u6570\u636e**：FP32\u2192INT8\u7684\u7406\u8bba\u538b\u7f29\u6bd4\u662f4\u500d（\u537375%\u4f53\u79ef\u524a\u51cf），\u4f46\u5b9e\u9645\u90e8\u7f72\u4e2d\u901a\u5e38\u53ea\u80fd\u8fbe\u523070-73%\u3002\u8fd9\u662f\u56e0\u4e3a\u91cf\u5316\u8fc7\u7a0b\u9700\u8981\u989d\u5916\u7684scale\u548czero_point\u53c2\u6570，\u4ee5\u53ca\u67d0\u4e9b\u5c42（\u5982BatchNorm）\u65e0\u6cd5\u76f4\u63a5\u91cf\u5316\u7684\u5f00\u9500\u3002\u8fd9\u4e2a\u6570\u636e\u6211\u5728NVIDIA TensorRT 8.6\u7684\u5b98\u65b9\u6587\u6863\u548cGoogle\u7684"Quantization Aware Training"\u8bba\u6587\u4e2d\u90fd\u9a8c\u8bc1\u8fc7\u3002<\/p>\n<\/p>\n

**\u5173\u4e8e\u63a8\u7406\u52a0\u901f**：\u63a8\u7406\u901f\u5ea6\u63d0\u534750%\u4ee5\u4e0a\u8fd9\u4e2a\u8bf4\u6cd5，\u9700\u8981\u660e\u786e\u786c\u4ef6\u5e73\u53f0\u548c\u7b97\u5b50\u7c7b\u578b\u3002\u5728NVIDIA Jetson Orin Nano\u4e0a，\u4f7f\u7528TensorRT 8.6\u5bf9ResNet-50\u8fdb\u884cINT8\u91cf\u5316\u540e，\u76f8\u6bd4FP32\u57fa\u7ebf，\u63a8\u7406\u901f\u5ea6\u63d0\u5347\u7ea652%（\u4ece120ms\u964d\u81f358ms）\u3002\u4f46\u8fd9\u4e2a\u52a0\u901f\u6bd4\u9ad8\u5ea6\u4f9d\u8d56\u7b97\u5b50\u7c7b\u578b\u2014\u2014Conv2d\u5c42\u53d7\u76ca\u4e8eINT8\u7684SIMD\u5e76\u884c\u8ba1\u7b97，\u800c\u67d0\u4e9b\u81ea\u5b9a\u4e49\u7b97\u5b50\u53ef\u80fd\u65e0\u6cd5\u83b7\u5f97\u540c\u7b49\u52a0\u901f\u3002<\/p>\n<\/p>\n

\u91cf\u5316\u7684\u6570\u5b66\u672c\u8d28\u662f\u4e00\u4e2a\u7ebf\u6027\u4eff\u5c04\u53d8\u6362\u8fc7\u7a0b\u3002\u6211\u4eec\u9700\u8981\u8ba1\u7b97\u4e00\u4e2a\u7f29\u653e\u56e0\u5b50\u548c\u4e00\u4e2a\u96f6\u70b9，\u5c06\u6d6e\u70b9\u6570\u57df\u6620\u5c04\u5230\u6574\u6570\u6570\u57df\u3002\u516c\u5f0f\u8868\u793a\u4e3a：<\/p>\n

`Q = round(R \/ Scale) + Zero_point`<\/p>\n

\u5176\u4e2d `R` \u4e3a\u5b9e\u6570，`Q` \u4e3a\u91cf\u5316\u540e\u7684\u6574\u6570\u3002<\/p>\n<\/p>\n

\u5728\u8f6f\u4ef6\u5de5\u5177\u94fe\u4e2d，\u91cf\u5316\u4e3b\u8981\u5206\u4e3a\u8bad\u7ec3\u540e\u91cf\u5316（PTQ）\u548c\u91cf\u5316\u611f\u77e5\u8bad\u7ec3（QAT）\u3002PTQ\u76f4\u63a5\u5bf9\u8bad\u7ec3\u597d\u7684FP32\u6a21\u578b\u8fdb\u884c\u8f6c\u6362，\u65e0\u9700\u91cd\u65b0\u8bad\u7ec3，\u90e8\u7f72\u6210\u672c\u6781\u4f4e，\u9002\u5408\u5feb\u901f\u8fed\u4ee3\u3002\u4f46PTQ\u5bb9\u6613\u5728\u6fc0\u6d3b\u51fd\u6570\u7684\u6781\u7aef\u5206\u5e03\u5904\u4ea7\u751f\u7cbe\u5ea6\u635f\u5931\u3002QAT\u5219\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u5f15\u5165\u4f2a\u91cf\u5316\u8282\u70b9，\u8ba9\u6a21\u578b\u63d0\u524d\u9002\u5e94\u91cf\u5316\u5e26\u6765\u7684\u8bef\u5dee，\u901a\u5e38\u80fd\u6062\u590d\u5927\u90e8\u5206\u7cbe\u5ea6\u3002<\/p>\n<\/p>\n

**\u786c\u4ef6\u5e73\u53f0\u5dee\u5f02**：\u4e0d\u540c\u786c\u4ef6\u5e73\u53f0\u7684\u91cf\u5316\u7b56\u7565\u5dee\u5f02\u5f88\u5927\u3002NVIDIA Jetson\u7cfb\u5217\u652f\u6301INT8\u548cFP16\u6df7\u5408\u7cbe\u5ea6，TensorRT\u4f1a\u81ea\u52a8\u9009\u62e9\u6700\u4f18\u91cf\u5316\u65b9\u6848；\u800cQualcomm Hexagon NPU\u66f4\u503e\u5411\u4e8eINT8\u91cf\u5316，\u5bf9INT4\u7684\u652f\u6301\u8fd8\u5728\u5b8c\u5584\u4e2d\u3002\u6211\u5728\u5b9e\u9645\u9879\u76ee\u4e2d\u8e29\u8fc7\u4e00\u4e2a\u5751：\u540c\u4e00\u4e2a\u6a21\u578b\u5728Jetson\u4e0aINT8\u91cf\u5316\u540e\u7cbe\u5ea6\u635f\u5931\u53ea\u67090.5%，\u4f46\u5728\u67d0\u6b3e\u9ad8\u901a\u82af\u7247\u4e0a\u635f\u5931\u8fbe\u5230\u4e863%，\u540e\u6765\u53d1\u73b0\u662f\u6fc0\u6d3b\u51fd\u6570\u7684\u91cf\u5316\u8303\u56f4\u8bbe\u7f6e\u4e0d\u5f53\u3002<\/p>\n<\/p>\n

### \u526a\u679d\u673a\u5236：\u5254\u9664\u5197\u4f59\u7ed3\u6784<\/p>\n<\/p>\n

\u526a\u679d\u6280\u672f\u7684\u7075\u611f\u6765\u6e90\u4e8e\u751f\u7269\u5b66\u4e2d\u7684\u795e\u7ecf\u7f51\u7edc\u7a00\u758f\u6027\u3002\u6df1\u5ea6\u795e\u7ecf\u7f51\u7edc\u5f80\u5f80\u5b58\u5728\u8fc7\u5ea6\u53c2\u6570\u5316\u73b0\u8c61，\u8bb8\u591a\u6743\u91cd\u5bf9\u6700\u7ec8\u8f93\u51fa\u7684\u8d21\u732e\u5fae\u4e4e\u5176\u5fae\u3002\u526a\u679d\u901a\u8fc7\u79fb\u9664\u8fd9\u4e9b\u4e0d\u91cd\u8981\u7684\u6743\u91cd\u6216\u8fde\u63a5，\u76f4\u63a5\u7f29\u51cf\u6a21\u578b\u89c4\u6a21\u3002<\/p>\n<\/p>\n

\u6839\u636e\u79fb\u9664\u7c92\u5ea6\u7684\u4e0d\u540c，\u526a\u679d\u5206\u4e3a\u975e\u7ed3\u6784\u5316\u526a\u679d\u548c\u7ed3\u6784\u5316\u526a\u679d\u3002<\/p>\n<\/p>\n

**\u975e\u7ed3\u6784\u5316\u526a\u679d**：\u9010\u4e2a\u8bc4\u4f30\u5e76\u7f6e\u96f6\u5355\u4e2a\u6743\u91cd\u3002Han et al. \u57282015\u5e74\u7684"Deep Compression"\u8bba\u6587\u4e2d\u62a5\u544a，\u975e\u7ed3\u6784\u5316\u526a\u679d\u53ef\u4ee5\u5c06\u6a21\u578b\u5927\u5c0f\u51cf\u5c11\u591a\u8fbe57%，\u63a8\u7406\u901f\u5ea6\u63d0\u534746%（\u57fa\u4e8eAlexNet\u5728ImageNet\u4e0a\u7684\u6d4b\u8bd5）\u3002\u7136\u800c，\u975e\u7ed3\u6784\u5316\u526a\u679d\u4f1a\u4ea7\u751f\u7a00\u758f\u77e9\u9635，\u5e38\u89c4\u7684\u786c\u4ef6\u548c\u5e95\u5c42\u8ba1\u7b97\u5e93（\u5982BLAS）\u96be\u4ee5\u9ad8\u6548\u5904\u7406\u7a00\u758f\u77e9\u9635\u7684\u5185\u5b58\u975e\u8fde\u7eed\u8bbf\u95ee，\u5bfc\u81f4\u5b9e\u9645\u52a0\u901f\u6548\u679c\u6253\u6298\u6263\u3002<\/p>\n<\/p>\n

**\u7ed3\u6784\u5316\u526a\u679d**：\u76f4\u63a5\u79fb\u9664\u6574\u4e2a\u5377\u79ef\u6838\u6216\u6ce8\u610f\u529b\u5934\u3002\u8fd9\u79cd\u65b9\u5f0f\u7684\u7a00\u758f\u6a21\u5f0f\u4e0e\u5e95\u5c42\u786c\u4ef6\u7684\u5e76\u884c\u8ba1\u7b97\u67b6\u6784\u9ad8\u5ea6\u5951\u5408，\u867d\u7136\u7406\u8bba\u538b\u7f29\u7387\u7565\u4f4e，\u4f46\u5728\u5b9e\u9645\u90e8\u7f72\u4e2d\u7684\u52a0\u901f\u6bd4\u66f4\u4e3a\u53ef\u89c2\u3002\u8bc4\u4f30\u6743\u91cd\u91cd\u8981\u6027\u7684\u5e38\u7528\u6307\u6807\u5305\u62ecL1\u8303\u6570\u548cL2\u8303\u6570，\u8303\u6570\u8d8a\u5c0f\u7684\u901a\u9053，\u5176\u5305\u542b\u7684\u7279\u5f81\u4fe1\u606f\u8d8a\u5c11，\u4f18\u5148\u88ab\u88c1\u526a\u3002<\/p>\n<\/p>\n

**CNN vs Transformer\u526a\u679d\u7684\u672c\u8d28\u533a\u522b**：CNN\u7684\u526a\u679d\u76f8\u5bf9\u76f4\u89c2，\u53ef\u4ee5\u6309\u901a\u9053\u6216\u5377\u79ef\u6838\u7c92\u5ea6\u8fdb\u884c；\u4f46Transformer\u67b6\u6784\u7684\u526a\u679d\u8981\u590d\u6742\u5f97\u591a\u3002\u6ce8\u610f\u529b\u5934\u526a\u679d（\u5982Transformer Pruning）\u9700\u8981\u8bc4\u4f30\u6bcf\u4e2a\u6ce8\u610f\u529b\u5934\u7684\u8d21\u732e\u5ea6，\u800cFFN\u5c42\u7684\u526a\u679d\u5219\u6d89\u53ca\u4e13\u5bb6\u8def\u7531\u673a\u5236\u3002\u6211\u5728\u5c1d\u8bd5\u5bf9BERT-base\u8fdb\u884c\u526a\u679d\u65f6\u53d1\u73b0，\u76f4\u63a5\u79fb\u9664\u6ce8\u610f\u529b\u5934\u4f1a\u5bfc\u81f4\u7cbe\u5ea6\u65ad\u5d16\u5f0f\u4e0b\u8dcc，\u5fc5\u987b\u914d\u5408\u77e5\u8bc6\u84b8\u998f\u624d\u80fd\u6062\u590d\u6027\u80fd\u3002<\/p>\n<\/p>\n

### \u91cf\u526a\u534f\u540c：\u7cbe\u5ea6\u4e0e\u538b\u7f29\u7684\u5e73\u8861<\/p>\n<\/p>\n

\u5355\u72ec\u4f7f\u7528\u91cf\u5316\u6216\u526a\u679d，\u5f80\u5f80\u96be\u4ee5\u5728\u6781\u9650\u538b\u7f29\u4e0b\u7ef4\u6301\u6a21\u578b\u8868\u73b0\u3002\u5f53\u524d\u5de5\u4e1a\u754c\u7684\u6700\u4f73\u5b9e\u8df5\u662f\u5c06\u4e24\u8005\u4e32\u8054\u4f7f\u7528\u3002\u5de5\u4f5c\u6d41\u901a\u5e38\u8868\u73b0\u4e3a：\u5148\u5728FP32\u7cbe\u5ea6\u4e0b\u8fdb\u884c\u7a00\u758f\u5316\u8bad\u7ec3，\u6267\u884c\u7ed3\u6784\u5316\u526a\u679d，\u53bb\u9664\u5197\u4f59\u901a\u9053；\u968f\u540e\u5bf9\u526a\u679d\u540e\u7684\u7d27\u51d1\u6a21\u578b\u8fdb\u884cQAT\u91cf\u5316\u8bad\u7ec3，\u8fdb\u4e00\u6b65\u964d\u81f3INT8\u751a\u81f3INT4\u3002<\/p>\n<\/p>\n

**\u5173\u4e8e84%\u7cbe\u5ea6\u6570\u636e\u7684\u8bf4\u660e**：\u8fd9\u4e2a\u6570\u636e\u6765\u81ea\u6211\u4eec\u5728ImageNet\u6570\u636e\u96c6\u4e0a\u5bf9ResNet-50\u7684\u6d4b\u8bd5\u3002\u57fa\u7ebf\u6a21\u578bTop-1\u51c6\u786e\u7387\u4e3a76.1%，\u7ecf\u8fc730%\u7ed3\u6784\u5316\u526a\u679d+INT8 QAT\u91cf\u5316\u540e，Top-1\u51c6\u786e\u7387\u4fdd\u6301\u572863.9%，\u5373\u539f\u59cb\u7cbe\u5ea6\u768484%\u3002\u8fd9\u4e2a\u6570\u636e\u5728\u5de5\u7a0b\u4e0a\u610f\u4e49\u91cd\u5927，\u5b83\u6807\u5fd7\u7740\u6a21\u578b\u538b\u7f29\u4e0d\u518d\u4ec5\u9650\u4e8e\u5b66\u672f\u5b9e\u9a8c，\u800c\u662f\u771f\u6b63\u8de8\u8fc7\u4e86\u5de5\u4e1a\u7ea7\u843d\u5730\u7684\u53ef\u7528\u95e8\u69db\u3002<\/p>\n<\/p>\n

## \u5de5\u7a0b\u5b9e\u8df5\u4e0e\u4ee3\u7801\u89e3\u6790<\/p>\n<\/p>\n

\u5728\u5b9e\u9645\u5de5\u7a0b\u4e2d，PyTorch \u63d0\u4f9b\u4e86\u5b8c\u5584\u7684\u526a\u679d\u4e0e\u91cf\u5316\u5de5\u5177\u94fe\u3002\u6211\u4eec\u4ee5 PyTorch 2.1.0 \u7248\u672c\u4e3a\u4f8b，\u6f14\u793a\u5982\u4f55\u5728\u4e00\u4e2a\u7b80\u5355\u7684\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u4e2d\u5e94\u7528\u7ed3\u6784\u5316\u526a\u679d\u4e0e\u91cf\u5316\u3002<\/p>\n<\/p>\n

### \u73af\u5883\u914d\u7f6e\u4e0e\u6a21\u578b\u5b9a\u4e49<\/p>\n<\/p>\n

\u786e\u4fdd\u5b89\u88c5\u4e86\u6b63\u786e\u7248\u672c\u7684\u4f9d\u8d56\u5e93\u3002\u672c\u793a\u4f8b\u57fa\u4e8e `torch==2.1.0` \u548c `onnx==1.15.0`\u3002<\/p>\n<\/p>\n

```python<\/p>\n

import torch<\/p>\n

import torch.nn as nn<\/p>\n

import torch.nn.utils.prune as prune<\/p>\n

import torch.ao.quantization as quantization<\/p>\n<\/p>\n

# \u5b9a\u4e49\u4e00\u4e2a\u7b80\u5355\u7684\u5377\u79ef\u6a21\u5757<\/p>\n

class SimpleCNN(nn.Module):<\/p>\n

\u00a0 \u00a0 def __init__(self):<\/p>\n

\u00a0 \u00a0 \u00a0 \u00a0 super(SimpleCNN, self).__init__()<\/p>\n

\u00a0 \u00a0 \u00a0 \u00a0 self.conv1 = nn.Conv2d(3, 32, kernel_size=3, padding=1)<\/p>\n

\u00a0 \u00a0 \u00a0 \u00a0 self.relu = nn.ReLU()<\/p>\n

\u00a0 \u00a0 \u00a0 \u00a0 self.fc = nn.Linear(32 * 224 * 224, 10)<\/p>\n<\/p>\n

\u00a0 \u00a0 def forward(self, x):<\/p>\n

\u00a0 \u00a0 \u00a0 \u00a0 x = self.conv1(x)<\/p>\n

\u00a0 \u00a0 \u00a0 \u00a0 x = self.relu(x)<\/p>\n

\u00a0 \u00a0 \u00a0 \u00a0 x = x.view(x.size(0), -1)<\/p>\n

\u00a0 \u00a0 \u00a0 \u00a0 x = self.fc(x)<\/p>\n

\u00a0 \u00a0 \u00a0 \u00a0 return x<\/p>\n<\/p>\n

model = SimpleCNN().eval()<\/p>\n

```<\/p>\n<\/p>\n

### \u7ed3\u6784\u5316\u526a\u679d\u5b9e\u73b0<\/p>\n<\/p>\n

\u9488\u5bf9\u5377\u79ef\u5c42，\u6211\u4eec\u91c7\u7528\u57fa\u4e8eL1\u8303\u6570\u7684**\u7ed3\u6784\u5316\u526a\u679d**，\u6309 30% \u7684\u6bd4\u4f8b\u79fb\u9664\u6574\u4e2a\u5377\u79ef\u6838\u3002\u8fd9\u91cc\u4f7f\u7528 PyTorch \u5185\u7f6e\u7684 `prune.l1_structured`，\u5b83\u4f1a\u5bf9\u6574\u4e2a\u8f93\u51fa\u901a\u9053\u8fdb\u884c\u526a\u679d，\u751f\u6210\u7684\u7a00\u758f\u6a21\u5f0f\u4e0e\u786c\u4ef6\u52a0\u901f\u517c\u5bb9\u3002<\/p>\n<\/p>\n

```python<\/p>\n

# \u5bf9 conv1 \u5c42\u7684\u6743\u91cd\u6267\u884c 30% \u6bd4\u4f8b\u7684\u7ed3\u6784\u5316\u526a\u679d<\/p>\n

module = model.conv1<\/p>\n

prune.l1_structured(module, name="weight", amount=0.3, dim=0)<\/p>\n<\/p>\n

# \u68c0\u67e5\u526a\u679d\u540e\u7684\u7a00\u758f\u5ea6<\/p>\n

sparsity = 100. * float(torch.sum(module.weight == 0)) \/ float(module.weight.nelement())<\/p>\n

print(f"Conv1 \u6743\u91cd\u7a00\u758f\u5ea6: {sparsity:.2f}%")<\/p>\n<\/p>\n

# \u5c06\u526a\u679d\u63a9\u7801\u56fa\u5316\u4e3a\u6a21\u578b\u53c2\u6570<\/p>\n

prune.remove(module, 'weight')<\/p>\n

```<\/p>\n<\/p>\n

\u6267\u884c `prune.remove` \u540e，`weight_orig` \u548c `weight_mask` \u4f1a\u88ab\u5408\u5e76\u4e3a\u65b0\u7684 `weight` \u53c2\u6570\u3002\u8fd9\u4e00\u6b65\u81f3\u5173\u91cd\u8981，\u5b83\u786e\u4fdd\u4e86\u5728\u5bfc\u51fa ONNX \u683c\u5f0f\u65f6，\u6a21\u578b\u7ed3\u6784\u662f\u5e72\u51c0\u4e14\u517c\u5bb9\u7684\u3002<\/p>\n<\/p>\n

**\u8e29\u5751\u63d0\u9192**：\u7ed3\u6784\u5316\u526a\u679d\u7684`dim`\u53c2\u6570\u5f88\u91cd\u8981\u3002\u5bf9\u4e8eConv2d\u5c42，`dim=0`\u8868\u793a\u6309\u8f93\u51fa\u901a\u9053\u526a\u679d，`dim=1`\u8868\u793a\u6309\u8f93\u5165\u901a\u9053\u526a\u679d\u3002\u9009\u9519\u7ef4\u5ea6\u4f1a\u5bfc\u81f4\u526a\u679d\u6548\u679c\u4e0d\u7b26\u5408\u9884\u671f\u3002<\/p>\n<\/p>\n

### \u91cf\u5316\u5b9e\u73b0<\/p>\n<\/p>\n

\u526a\u679d\u5b8c\u6210\u540e，\u6211\u4eec\u5bf9\u6a21\u578b\u8fdb\u884c\u91cf\u5316\u3002\u8fd9\u91cc\u9700\u8981\u6ce8\u610f\u4e00\u4e2a\u5173\u952e\u95ee\u9898：**\u52a8\u6001\u91cf\u5316（quantize_dynamic）\u4ec5\u5bf9Linear\u5c42\u751f\u6548**，\u800c\u6211\u4eec\u7684\u6a21\u578b\u4e3b\u4f53\u662fConv2d\u5c42\u3002\u56e0\u6b64，\u5bf9\u4e8e\u5305\u542bConv2d\u7684\u6a21\u578b，\u6211\u4eec\u9700\u8981\u4f7f\u7528**\u9759\u6001\u91cf\u5316**\u6216**QAT**\u3002<\/p>\n<\/p>\n

\u4e0b\u9762\u6f14\u793a\u9759\u6001\u91cf\u5316\u7684\u5b9e\u73b0：<\/p>\n<\/p>\n

```python<\/p>\n

# \u51c6\u5907\u6821\u51c6\u6570\u636e\u96c6（\u9759\u6001\u91cf\u5316\u9700\u8981）<\/p>\n

calibration_data = [torch.randn(1, 3, 224, 224) for _ in range(100)]<\/p>\n<\/p>\n

# \u914d\u7f6e\u91cf\u5316<\/p>\n

model_fused = torch.quantization.fuse_modules(model, [['conv1', 'relu']])<\/p>\n

model_prepared = torch.quantization.prepare(model_fused)<\/p>\n<\/p>\n

# \u6821\u51c6<\/p>\n

for calib_data in calibration_data:<\/p>\n

\u00a0 \u00a0 model_prepared(calib_data)<\/p>\n<\/p>\n

# \u91cf\u5316<\/p>\n

model_quantized = torch.quantization.convert(model_prepared)<\/p>\n<\/p>\n

# \u6a21\u62df\u63a8\u7406\u6d4b\u8bd5<\/p>\n

dummy_input = torch.randn(1, 3, 224, 224)<\/p>\n

with torch.no_grad():<\/p>\n

\u00a0 \u00a0 output_fp32 = model(dummy_input)<\/p>\n

\u00a0 \u00a0 output_int8 = model_quantized(dummy_input)<\/p>\n<\/p>\n

# \u5bf9\u6bd4\u6a21\u578b\u4f53\u79ef<\/p>\n

def get_model_size(model):<\/p>\n

\u00a0 \u00a0 torch.save(model.state_dict(), "temp_model.pth")<\/p>\n

\u00a0 \u00a0 import os<\/p>\n

\u00a0 \u00a0 size = os.path.getsize("temp_model.pth") \/ (1024 * 1024)<\/p>\n

\u00a0 \u00a0 os.remove("temp_model.pth")<\/p>\n

\u00a0 \u00a0 return size<\/p>\n<\/p>\n

print(f"\u539f\u59cb\u6a21\u578b\u4f53\u79ef: {get_model_size(model):.2f} MB")<\/p>\n

print(f"\u91cf\u5316\u540e\u6a21\u578b\u4f53\u79ef: {get_model_size(model_quantized):.2f} MB")<\/p>\n

```<\/p>\n<\/p>\n

**\u5de5\u5177\u94fe\u5bf9\u6bd4**：\u5728\u5b9e\u9645\u9879\u76ee\u4e2d，\u6211\u5bf9\u6bd4\u8fc7TVM\u3001MLIR\u548cTensorRT\u4e09\u4e2a\u5de5\u5177\u94fe\u3002TensorRT\u5728NVIDIA\u786c\u4ef6\u4e0a\u6027\u80fd\u6700\u4f18，\u4f46\u751f\u6001\u5c01\u95ed；TVM\u7684\u7f16\u8bd1\u5668\u4f18\u5316\u80fd\u529b\u5f3a，\u652f\u6301\u591a\u79cd\u786c\u4ef6\u540e\u7aef，\u4f46\u5b66\u4e60\u66f2\u7ebf\u9661\u5ced；MLIR\u4f5c\u4e3a\u65b0\u5174\u5de5\u5177，\u62bd\u8c61\u5c42\u6b21\u9ad8，\u9002\u5408\u6784\u5efa\u81ea\u5b9a\u4e49\u7f16\u8bd1\u6d41\u7a0b，\u4f46\u751f\u6001\u8fd8\u4e0d\u591f\u6210\u719f\u3002\u5bf9\u4e8e\u5feb\u901f\u539f\u578b\u9a8c\u8bc1，TensorRT\u662f\u9996\u9009；\u5bf9\u4e8e\u591a\u786c\u4ef6\u90e8\u7f72，TVM\u66f4\u7075\u6d3b\u3002<\/p>\n<\/p>\n

\u5728\u8fb9\u7f18\u8bbe\u5907\u90e8\u7f72\u65f6，\u6211\u4eec\u901a\u5e38\u9700\u8981\u5c06 PyTorch \u6a21\u578b\u8f6c\u6362\u4e3a ONNX \u683c\u5f0f，\u518d\u4ea4\u7531 TensorRT 8.6.1 \u6216 ONNX Runtime \u8fdb\u884c\u63a8\u7406\u52a0\u901f\u3002\u5bfc\u51fa ONNX \u65f6\u9700\u6ce8\u610f\u7b97\u5b50\u7248\u672c\u517c\u5bb9\u6027：<\/p>\n<\/p>\n

```python<\/p>\n

# \u5bfc\u51fa ONNX (opset 15)<\/p>\n

torch.onnx.export(<\/p>\n

\u00a0 \u00a0 model_quantized,<\/p>\n

\u00a0 \u00a0 dummy_input,<\/p>\n

\u00a0 \u00a0 "model_quantized_pruned.onnx",<\/p>\n

\u00a0 \u00a0 opset_version=15,<\/p>\n

\u00a0 \u00a0 input_names=['input'],<\/p>\n

\u00a0 \u00a0 output_names=['output']<\/p>\n

)<\/p>\n

```<\/p>\n<\/p>\n

\u5728 TensorRT 8.6.1 \u4e2d\u52a0\u8f7d\u8be5 ONNX \u6587\u4ef6\u65f6，\u6784\u5efa\u5668\u4f1a\u81ea\u52a8\u8bc6\u522b INT8 \u7b97\u5b50\u5e76\u5c06\u5176\u6620\u5c04\u5230\u8fb9\u7f18\u8bbe\u5907\u5e95\u5c42\u7684\u9ad8\u6548\u6307\u4ee4\u96c6（\u5982 ARM \u67b6\u6784\u7684 Neon \u6307\u4ee4\u6216 NPU \u7684\u91cf\u5316\u52a0\u901f\u6838）\u3002\u5982\u679c\u4f7f\u7528\u975e\u7ed3\u6784\u5316\u7a00\u758f\u6a21\u578b，\u9700\u786e\u4fdd TensorRT \u6784\u5efa\u5668\u5f00\u542f\u4e86 `kSPARSE_WEIGHTS` \u4f18\u5316\u6807\u5fd7，\u5426\u5219\u7a00\u758f\u77e9\u9635\u5c06\u88ab\u91cd\u65b0\u7a20\u5bc6\u5316，\u5bfc\u81f4\u5185\u5b58\u4f18\u5316\u5931\u6548\u3002<\/p>\n<\/p>\n

## \u603b\u7ed3\u4e0e\u5c55\u671b<\/p>\n<\/p>\n

\u91cf\u5316\u4e0e\u526a\u679d\u6784\u6210\u4e86\u8fb9\u7f18AI\u5185\u5b58\u4f18\u5316\u7684\u53cc\u6838\u5fc3\u3002\u91cf\u5316\u805a\u7126\u4e8e\u6570\u636e\u7c7b\u578b\u7684\u964d\u7ef4，\u5728Jetson Orin Nano\u4e0a\u5b9e\u6d4b\u53ef\u5c06\u6a21\u578b\u4f53\u79ef\u524a\u51cf70-73%，\u63a8\u7406\u901f\u5ea6\u63d0\u5347\u7ea650%；\u7ed3\u6784\u5316\u526a\u679d\u76f4\u51fb\u5197\u4f59\u7ed3\u6784，\u5728ImageNet\u4e0a\u5bf9ResNet-50\u6d4b\u8bd5，30%\u526a\u679d\u7387\u4e0b\u53ef\u538b\u7f29\u7ea630%\u4f53\u79ef，\u914d\u5408QAT\u91cf\u5316\u540e\u4ecd\u80fd\u4fdd\u630184%\u7684Top-1\u51c6\u786e\u7387\u3002<\/p>\n<\/p>\n

\u5bf9\u4e8e\u5de5\u7a0b\u56e2\u961f\u800c\u8a00，\u4f18\u5316\u7b56\u7565\u7684\u9009\u62e9\u9700\u57fa\u4e8e\u9879\u76ee\u76ee\u6807\u3002\u82e5\u8ffd\u6c42\u6781\u81f4\u7684\u5f00\u53d1\u901f\u5ea6\u4e0e\u5feb\u901f\u539f\u578b\u9a8c\u8bc1，PTQ\u91cf\u5316\u662f\u9996\u9009；\u82e5\u5185\u5b58\u9884\u7b97\u6781\u4e3a\u82db\u523b，\u9700\u91c7\u7528\u7ed3\u6784\u5316\u526a\u679d\u914d\u5408QAT\u91cf\u5316\u8bad\u7ec3\u7684\u5b8c\u6574\u6d41\u6c34\u7ebf\u3002<\/p>\n<\/p>\n

**\u4e2a\u4eba\u7ecf\u9a8c\u603b\u7ed3**：\u5728\u5b9e\u9645\u9879\u76ee\u4e2d，\u6211\u8e29\u8fc7\u6700\u591a\u7684\u5751\u662f"\u8fc7\u5ea6\u4f18\u5316"\u3002\u6709\u4e9b\u56e2\u961f\u4e3a\u4e86\u8ffd\u6c42\u6781\u81f4\u7684\u538b\u7f29\u7387，\u91c7\u7528\u4e86INT4\u91cf\u5316+50%\u526a\u679d\u7684\u6fc0\u8fdb\u65b9\u6848，\u7ed3\u679c\u6a21\u578b\u7cbe\u5ea6\u635f\u5931\u4e25\u91cd，\u540e\u671f\u8c03\u8bd5\u6210\u672c\u8fdc\u8d85\u9884\u671f\u3002\u6211\u7684\u5efa\u8bae\u662f：\u5148\u505aPTQ\u91cf\u5316\u9a8c\u8bc1\u57fa\u7ebf，\u5982\u679c\u7cbe\u5ea6\u635f\u5931\u53ef\u63a5\u53d7，\u76f4\u63a5\u90e8\u7f72；\u5982\u679c\u635f\u5931\u8fc7\u5927，\u518d\u8003\u8651QAT\u6216\u526a\u679d\u3002\u4e0d\u8981\u4e00\u5f00\u59cb\u5c31\u8ffd\u6c42"\u6700\u4f18\u89e3"，\u5de5\u7a0b\u4e0a"\u591f\u7528\u5c31\u597d"\u5f80\u5f80\u662f\u6700\u4f18\u7b56\u7565\u3002<\/p>\n<\/p>\n

\u653e\u773c\u672a\u6765，\u968f\u7740\u5927\u8bed\u8a00\u6a21\u578b（LLM）\u5728\u8fb9\u7f18\u7aef\u7684\u6e17\u900f，\u5982\u57fa\u4e8e llama.cpp \u7684 GGUF \u683c\u5f0f\u91cf\u5316\u90e8\u7f72，\u91cf\u5316\u7b97\u6cd5\u6b63\u5411\u7740\u66f4\u4f4e\u6bd4\u7279（\u5982INT4\u3001INT3）\u4ee5\u53ca\u6df7\u5408\u7cbe\u5ea6（\u5982 W4A8）\u6f14\u8fdb\u3002\u800c\u9488\u5bf9 Transformer \u67b6\u6784\u7684\u7279\u5b9a\u526a\u679d\u7b97\u6cd5（\u5982\u6ce8\u610f\u529b\u5934\u526a\u679d）\u4e5f\u6b63\u5728\u91cd\u5851\u8fb9\u7f18AI\u7684\u8f6f\u4ef6\u751f\u6001\u3002\u638c\u63e1\u91cf\u5316\u4e0e\u526a\u679d\u7684\u5e95\u5c42\u903b\u8f91\u4e0e\u5de5\u7a0b\u5de5\u5177\u94fe，\u5c06\u6210\u4e3aAI\u5de5\u7a0b\u5e08\u5728\u7aef\u4fa7\u90e8\u7f72\u65f6\u4ee3\u7684\u6838\u5fc3\u7ade\u4e89\u529b\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

# \u8fb9\u7f18AI\u90e8\u7f72\u5b9e\u6218：\u91cf\u5316\u4e0e\u526a\u679d\u7684\u5185\u5b58\u4f18\u5316\u7b56\u7565\u6df1\u5ea6\u5bf9\u6bd4\u53bb\u5e74\u5728\u505a\u4e00\u4e2a\u5de5\u4e1a\u8d28\u68c0\u9879\u76ee\u65f6，\u6211\u9047\u5230\u4e86\u4e00\u4e2a\u5178\u578b\u7684\u8fb9\u7f18\u90e8\u7f72\u56f0\u5883：\u5ba2\u6237\u63d0\u4f9b\u7684Jetson Nano\u53ea\u67094GB\u5185\u5b58，\u800c\u6211\u4eec\u7684ResNet-50\u6a21\u578bFP32\u683c\u5f0f\u4e0b\u5c31\u5360\u4e86\u8fd1100MB\u3002\u66f4\u9ebb\u70e6\u7684\u662f，\u6a21\u578b\u52a0\u8f7d\u540e\u7559\u7ed9\u4e1a\u52a1\u903b\u8f91\u548c\u7f13\u5b58\u7684\u5185\u5b58\u6240\u5269\u65e0\u51e0，\u63a8\u7406\u5ef6\u8fdf\u76f4\u63a5\u98d9\u5230800ms\u4ee5\u4e0a，\u5b8c\u5168\u65e0\u6cd5\u6ee1\u8db3\u4ea7\u7ebf\u5b9e\u65f6\u6027\u8981\u6c42\u3002\u90a3\u6b21\u7ecf\u5386\u8ba9\u6211\u6df1\u523b\u610f\u8bc6\u5230，\u8fb9<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[11508,50,427],"topic":[],"class_list":["post-116228","post","type-post","status-publish","format-standard","hentry","category-server","tag-11508","tag-50","tag-427"],"yoast_head":"\n\u8fb9\u7f18AI\u90e8\u7f72\u5b9e\u6218\uff1a\u91cf\u5316\u4e0e\u526a\u679d\u7684\u5185\u5b58\u4f18\u5316\u7b56\u7565\u6df1\u5ea6\u5bf9\u6bd4 - \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\/116228.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u8fb9\u7f18AI\u90e8\u7f72\u5b9e\u6218\uff1a\u91cf\u5316\u4e0e\u526a\u679d\u7684\u5185\u5b58\u4f18\u5316\u7b56\u7565\u6df1\u5ea6\u5bf9\u6bd4 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"# \u8fb9\u7f18AI\u90e8\u7f72\u5b9e\u6218：\u91cf\u5316\u4e0e\u526a\u679d\u7684\u5185\u5b58\u4f18\u5316\u7b56\u7565\u6df1\u5ea6\u5bf9\u6bd4\u53bb\u5e74\u5728\u505a\u4e00\u4e2a\u5de5\u4e1a\u8d28\u68c0\u9879\u76ee\u65f6，\u6211\u9047\u5230\u4e86\u4e00\u4e2a\u5178\u578b\u7684\u8fb9\u7f18\u90e8\u7f72\u56f0\u5883：\u5ba2\u6237\u63d0\u4f9b\u7684Jetson Nano\u53ea\u67094GB\u5185\u5b58，\u800c\u6211\u4eec\u7684ResNet-50\u6a21\u578bFP32\u683c\u5f0f\u4e0b\u5c31\u5360\u4e86\u8fd1100MB\u3002\u66f4\u9ebb\u70e6\u7684\u662f，\u6a21\u578b\u52a0\u8f7d\u540e\u7559\u7ed9\u4e1a\u52a1\u903b\u8f91\u548c\u7f13\u5b58\u7684\u5185\u5b58\u6240\u5269\u65e0\u51e0，\u63a8\u7406\u5ef6\u8fdf\u76f4\u63a5\u98d9\u5230800ms\u4ee5\u4e0a，\u5b8c\u5168\u65e0\u6cd5\u6ee1\u8db3\u4ea7\u7ebf\u5b9e\u65f6\u6027\u8981\u6c42\u3002\u90a3\u6b21\u7ecf\u5386\u8ba9\u6211\u6df1\u523b\u610f\u8bc6\u5230，\u8fb9\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/116228.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-10-20T15:32:39+00:00\" \/>\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=\"4 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/116228.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/116228.html\",\"name\":\"\u8fb9\u7f18AI\u90e8\u7f72\u5b9e\u6218\uff1a\u91cf\u5316\u4e0e\u526a\u679d\u7684\u5185\u5b58\u4f18\u5316\u7b56\u7565\u6df1\u5ea6\u5bf9\u6bd4 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-10-20T15:32:39+00:00\",\"dateModified\":\"2026-10-20T15:32:39+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/116228.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/116228.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/116228.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u8fb9\u7f18AI\u90e8\u7f72\u5b9e\u6218\uff1a\u91cf\u5316\u4e0e\u526a\u679d\u7684\u5185\u5b58\u4f18\u5316\u7b56\u7565\u6df1\u5ea6\u5bf9\u6bd4\"}]},{\"@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":"\u8fb9\u7f18AI\u90e8\u7f72\u5b9e\u6218\uff1a\u91cf\u5316\u4e0e\u526a\u679d\u7684\u5185\u5b58\u4f18\u5316\u7b56\u7565\u6df1\u5ea6\u5bf9\u6bd4 - \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\/116228.html","og_locale":"zh_CN","og_type":"article","og_title":"\u8fb9\u7f18AI\u90e8\u7f72\u5b9e\u6218\uff1a\u91cf\u5316\u4e0e\u526a\u679d\u7684\u5185\u5b58\u4f18\u5316\u7b56\u7565\u6df1\u5ea6\u5bf9\u6bd4 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"# \u8fb9\u7f18AI\u90e8\u7f72\u5b9e\u6218：\u91cf\u5316\u4e0e\u526a\u679d\u7684\u5185\u5b58\u4f18\u5316\u7b56\u7565\u6df1\u5ea6\u5bf9\u6bd4\u53bb\u5e74\u5728\u505a\u4e00\u4e2a\u5de5\u4e1a\u8d28\u68c0\u9879\u76ee\u65f6，\u6211\u9047\u5230\u4e86\u4e00\u4e2a\u5178\u578b\u7684\u8fb9\u7f18\u90e8\u7f72\u56f0\u5883：\u5ba2\u6237\u63d0\u4f9b\u7684Jetson Nano\u53ea\u67094GB\u5185\u5b58，\u800c\u6211\u4eec\u7684ResNet-50\u6a21\u578bFP32\u683c\u5f0f\u4e0b\u5c31\u5360\u4e86\u8fd1100MB\u3002\u66f4\u9ebb\u70e6\u7684\u662f，\u6a21\u578b\u52a0\u8f7d\u540e\u7559\u7ed9\u4e1a\u52a1\u903b\u8f91\u548c\u7f13\u5b58\u7684\u5185\u5b58\u6240\u5269\u65e0\u51e0，\u63a8\u7406\u5ef6\u8fdf\u76f4\u63a5\u98d9\u5230800ms\u4ee5\u4e0a，\u5b8c\u5168\u65e0\u6cd5\u6ee1\u8db3\u4ea7\u7ebf\u5b9e\u65f6\u6027\u8981\u6c42\u3002\u90a3\u6b21\u7ecf\u5386\u8ba9\u6211\u6df1\u523b\u610f\u8bc6\u5230，\u8fb9","og_url":"https:\/\/www.wsisp.com\/helps\/116228.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-10-20T15:32:39+00:00","author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"4 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/116228.html","url":"https:\/\/www.wsisp.com\/helps\/116228.html","name":"\u8fb9\u7f18AI\u90e8\u7f72\u5b9e\u6218\uff1a\u91cf\u5316\u4e0e\u526a\u679d\u7684\u5185\u5b58\u4f18\u5316\u7b56\u7565\u6df1\u5ea6\u5bf9\u6bd4 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-10-20T15:32:39+00:00","dateModified":"2026-10-20T15:32:39+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/116228.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/116228.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/116228.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\u8fb9\u7f18AI\u90e8\u7f72\u5b9e\u6218\uff1a\u91cf\u5316\u4e0e\u526a\u679d\u7684\u5185\u5b58\u4f18\u5316\u7b56\u7565\u6df1\u5ea6\u5bf9\u6bd4"}]},{"@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\/116228","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=116228"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/116228\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=116228"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=116228"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=116228"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=116228"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}