{"id":553,"date":"2026-07-08T03:12:48","date_gmt":"2026-07-07T19:12:48","guid":{"rendered":"https:\/\/tech-renaissance.cn\/blog\/?p=553"},"modified":"2026-07-08T21:19:21","modified_gmt":"2026-07-08T13:19:21","slug":"%e6%8d%9f%e5%a4%b1%e5%87%bd%e6%95%b0%e3%80%81%e5%ad%a6%e4%b9%a0%e7%8e%87%e8%b0%83%e5%ba%a6%e4%b8%8e%e5%8f%82%e6%95%b0%e5%88%9d%e5%a7%8b%e5%8c%96%ef%bc%9a%e8%ae%ad%e7%bb%83%e7%ae%97%e6%b3%95%e9%85%8d","status":"publish","type":"post","link":"https:\/\/tech-renaissance.cn\/blog\/index.php\/2026\/07\/08\/553\/","title":{"rendered":"(22) \u635f\u5931\u51fd\u6570\u3001\u5b66\u4e60\u7387\u8c03\u5ea6\u4e0e\u53c2\u6570\u521d\u59cb\u5316\uff1a\u8bad\u7ec3\u7b97\u6cd5\u914d\u7f6e\u5c42"},"content":{"rendered":"\n<h2 class=\"wp-block-heading\">\u2014\u2014\u201c\u4e00\u4e2a\u4eba\u7528AI\u5982\u4f55\u5199\u51fa\u6bd4PyTorch\u66f4\u5feb\u7684\u81ea\u7814\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u201d\u7cfb\u5217\u6587\u7ae0\u4e4b\u4e8c\u5341\u4e8c<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u524d\u9762\u7684\u6587\u7ae0\u82b1\u4e86\u5f88\u5927\u7bc7\u5e45\u8bb2 Tech-Renaissance \u7684\u201c\u9aa8\u67b6\u201d\uff1a\u5f20\u91cf\u3001\u8ba1\u7b97\u56fe\u3001\u7f16\u8bd1\u5668\u3001\u663e\u5b58\u89c4\u5212\u3001\u591a\u6d41\u6267\u884c\u3001CUDA Graph \u6355\u83b7\u3001\u878d\u5408\u7b97\u5b50\u3002\u9aa8\u67b6\u518d\u5f3a\u58ee\uff0c\u5982\u679c\u6ca1\u6709\u8bad\u7ec3\u7b97\u6cd5\u7684\u6b63\u786e\u914d\u7f6e\uff0c\u6a21\u578b\u4e5f\u8dd1\u4e0d\u8d77\u6765\u2014\u2014\u5c31\u597d\u6bd4\u4e00\u8f86\u8d5b\u8f66\uff0c\u53d1\u52a8\u673a\u3001\u53d8\u901f\u7bb1\u3001\u60ac\u6302\u90fd\u8c03\u597d\u4e86\uff0c\u4f46\u5982\u679c\u6ca1\u6709\u5408\u9002\u7684\u71c3\u6cb9\u6807\u53f7\u3001\u70b9\u706b\u65f6\u673a\u548c\u8f6e\u80ce\u6c14\u538b\uff0c\u5b83\u4f9d\u7136\u5230\u4e0d\u4e86\u7ec8\u70b9\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bad\u7ec3\u7b97\u6cd5\u7684\u914d\u7f6e\u5c42\uff0c\u901a\u5e38\u88ab\u79f0\u4e3a\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u91cc\u201c\u6700\u4e0d\u8d77\u773c\u3001\u5374\u6700\u5f71\u54cd\u7ed3\u679c\u201d\u7684\u90e8\u5206\u3002\u5b83\u5305\u62ec\u635f\u5931\u51fd\u6570\u3001\u5b66\u4e60\u7387\u8c03\u5ea6\u3001\u53c2\u6570\u521d\u59cb\u5316\uff0c\u518d\u914d\u5408\u4f18\u5316\u5668\uff08\u4e0a\u4e00\u7bc7\u5df2\u7ecf\u4e13\u95e8\u8bb2\u8fc7\uff09\u548c\u968f\u673a\u6570\u751f\u6210\uff08\u4e0b\u4e00\u7bc7\u4f1a\u7ec6\u8bb2\uff09\uff0c\u5171\u540c\u51b3\u5b9a\u4e86\u4e00\u6b21\u8bad\u7ec3\u80fd\u5426\u6536\u655b\u3001\u6536\u655b\u5230\u591a\u597d\u7684\u4f4d\u7f6e\u3001\u4ee5\u53ca\u6536\u655b\u5f97\u591a\u5feb\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u7bc7\u6587\u7ae0\u5c31\u6765\u804a\u804a Tech-Renaissance \u91cc\u8fd9\u4e00\u5c42\u662f\u600e\u4e48\u8bbe\u8ba1\u7684\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e00\u3001\u8bad\u7ec3\u7b97\u6cd5\u914d\u7f6e\u5c42\u5230\u5e95\u5728\u914d\u4ec0\u4e48<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">1.1 \u635f\u5931\u51fd\u6570\uff1a\u628a\u9884\u6d4b\u53d8\u6210\u53ef\u4f18\u5316\u7684\u76ee\u6807<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u6700\u540e\u8f93\u51fa\u7684\u4e00\u822c\u662f logits \u6216\u6982\u7387\u5206\u5e03\uff0c\u800c\u635f\u5931\u51fd\u6570\u8d1f\u8d23\u8861\u91cf\u8fd9\u4e2a\u5206\u5e03\u4e0e\u771f\u5b9e\u6807\u7b7e\u4e4b\u95f4\u7684\u5dee\u8ddd\u3002\u5bf9\u4e8e\u5206\u7c7b\u4efb\u52a1\uff0c\u6700\u5e38\u7528\u7684\u662f<strong>\u4ea4\u53c9\u71b5\u635f\u5931\uff08Cross-Entropy Loss\uff09<\/strong>\u3002\u5b83\u7684\u6838\u5fc3\u601d\u60f3\u5f88\u6734\u7d20\uff1a\u6a21\u578b\u5bf9\u6b63\u786e\u7c7b\u522b\u7684\u9884\u6d4b\u6982\u7387\u8d8a\u9ad8\uff0c\u635f\u5931\u8d8a\u4f4e\uff1b\u5982\u679c\u6a21\u578b\u5bf9\u9519\u8bef\u7b54\u6848\u5f88\u81ea\u4fe1\uff0c\u635f\u5931\u4f1a\u8fc5\u901f\u653e\u5927\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5177\u4f53\u5730\uff0c\u5bf9\u4e8e\u5355\u4e2a\u6837\u672c\uff0c\u82e5\u771f\u5b9e\u6807\u7b7e\u4e3a one-hot \u5411\u91cf <code>y<\/code>\uff0c\u6a21\u578b\u8f93\u51fa\u6982\u7387\u4e3a <code>p<\/code>\uff0c\u5219\u4ea4\u53c9\u71b5\u53ef\u4ee5\u5199\u6210\u4e0b\u9762\u7684\u4f2a\u4ee3\u7801\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"python\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\"># y[i] : \u7b2c i \u7c7b\u7684 one-hot \u6807\u7b7e\uff08\u771f\u5b9e\u7c7b\u522b\u4e3a 1\uff0c\u5176\u4f59\u4e3a 0\uff09\n# p[i] : \u6a21\u578b\u9884\u6d4b\u7684\u7b2c i \u7c7b\u6982\u7387\uff0c\u7531 softmax \u8f93\u51fa\nloss = -sum(y[i] * log(p[i]) for i in range(num_classes))\n# \u7531\u4e8e one-hot \u4e2d\u53ea\u6709\u6b63\u786e\u7c7b\u522b y[target] == 1\uff0c\u4e0a\u5f0f\u7b49\u4ef7\u4e8e\nloss = -log(p[target])<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a\u5f0f\u5b50\u9f13\u52b1\u6a21\u578b\u628a\u6b63\u786e\u7c7b\u522b\u7684\u6982\u7387\u63a8\u5411 1\uff1b\u540c\u65f6\u7531\u4e8e <code>log<\/code> \u5728\u63a5\u8fd1 0 \u65f6\u4e0b\u964d\u5f88\u5feb\uff0c\u4e5f\u8ba9\u201c\u81ea\u4fe1\u4f46\u9519\u8bef\u201d\u7684\u6837\u672c\u4ed8\u51fa\u6781\u9ad8\u4ee3\u4ef7\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4ea4\u53c9\u71b5\u901a\u5e38\u548c softmax \u4e00\u8d77\u4f7f\u7528\u3002\u4e3a\u4e86\u8ba9\u6570\u503c\u66f4\u7a33\u5b9a\uff0c\u73b0\u4ee3\u6846\u67b6\u4e0d\u4f1a\u5148\u7b97 softmax \u518d\u53d6 log\uff0c\u800c\u662f\u76f4\u63a5\u505a<strong>fused log-softmax + NLL<\/strong>\uff0c\u5229\u7528 log-sum-exp \u6280\u5de7\u907f\u514d\u6307\u6570\u7206\u70b8\u6216\u4e0b\u6ea2\u3002PyTorch \u7684 <code>nn.CrossEntropyLoss<\/code> \u548c TensorFlow \u7684 <code>SparseCategoricalCrossentropy<\/code> \u90fd\u9075\u5faa\u8fd9\u4e00\u60ef\u4f8b\uff0c\u8981\u6c42\u7528\u6237\u4f20\u5165\u539f\u59cb logits \u800c\u4e0d\u662f\u5df2\u7ecf\u5f52\u4e00\u5316\u7684\u6982\u7387\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u5206\u7c7b\u4efb\u52a1\u4e2d\uff0c\u4e00\u4e2a\u5e38\u7528\u7684\u6280\u5de7\u662f<strong>\u6807\u7b7e\u5e73\u6ed1\uff08Label Smoothing\uff09<\/strong>\u3002\u4f20\u7edf one-hot \u6807\u7b7e\u4f1a\u628a\u5168\u90e8\u6982\u7387\u653e\u5728\u6b63\u786e\u7c7b\u522b\u4e0a\uff0c\u6a21\u578b\u4e3a\u4e86\u62df\u5408\u8fd9\u79cd\u201c\u786c\u201d\u76ee\u6807\uff0c\u5bb9\u6613\u5bf9\u8bad\u7ec3\u6837\u672c\u8fc7\u5ea6\u81ea\u4fe1\uff0c\u6cdb\u5316\u80fd\u529b\u4e0b\u964d\u3002\u6807\u7b7e\u5e73\u6ed1\u628a\u76ee\u6807\u5206\u5e03\u6539\u6210\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"python\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\"># eps : \u5e73\u6ed1\u7cfb\u6570\uff0c\u5e38\u89c1\u53d6\u503c\u4e3a 0.1\n# num_classes : \u7c7b\u522b\u6570\n# one_hot : \u771f\u5b9e\u7c7b\u522b\u7684 one-hot \u5206\u5e03\nsmoothed = (1 - eps) * one_hot + eps \/ num_classes\n# \u7b49\u4ef7\u4e8e\uff1a\u628a\u6b63\u786e\u7c7b\u522b\u7684\u6982\u7387\u4ece 1 \u964d\u5230 1 - eps + eps\/num_classes\uff0c\n#         \u5176\u4f59\u7c7b\u522b\u5404\u5206 eps\/num_classes<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u5176\u4e2d <code>\u03b5<\/code> \u901a\u5e38\u53d6 0.1\uff0c\u76f8\u5f53\u4e8e\u544a\u8bc9\u6a21\u578b\u201c\u6b63\u786e\u7b54\u6848\u7684\u7f6e\u4fe1\u5ea6\u4e0d\u8981\u62c9\u6ee1\u201d\u3002\u5728 ImageNet \u7b49\u5927\u89c4\u6a21\u89c6\u89c9\u4efb\u52a1\u91cc\uff0c0.1 \u7684\u6807\u7b7e\u5e73\u6ed1\u51e0\u4e4e\u662f\u6807\u51c6\u914d\u65b9\uff0c\u80fd\u591f\u8ba9 top-1\/top-5 \u8f7b\u5fae\u63d0\u5347\uff0c\u540c\u65f6\u8ba9\u6a21\u578b\u8f93\u51fa\u7684\u6982\u7387\u5206\u5e03\u66f4\u52a0\u6821\u51c6\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4ea4\u53c9\u71b5\u8fd8\u6709\u4e00\u4e2a\u91cd\u8981\u4f18\u52bf\uff1a\u5f53\u5b83\u548c softmax \u4e00\u8d77\u6c42\u5bfc\u65f6\uff0c\u68af\u5ea6\u5f62\u5f0f\u975e\u5e38\u7b80\u6d01\uff0c\u7b49\u4e8e <code>softmax(logits) - target<\/code>\u3002\u8fd9\u610f\u5473\u7740\u5373\u4f7f\u6a21\u578b\u5728\u65e9\u671f\u8f93\u51fa\u6781\u5dee\u7684 logits\uff0c\u68af\u5ea6\u4e5f\u4e0d\u4f1a\u50cf\u5747\u65b9\u8bef\u5dee\u90a3\u6837\u56e0\u4e3a\u94fe\u5f0f\u6c42\u5bfc\u800c\u8fc5\u901f\u8870\u51cf\uff0c\u6df1\u5c42\u7f51\u7edc\u56e0\u6b64\u66f4\u5bb9\u6613\u83b7\u5f97\u6709\u6548\u7684\u66f4\u65b0\u4fe1\u53f7\u3002\u8fd9\u4e5f\u662f\u5206\u7c7b\u4efb\u52a1\u51e0\u4e4e\u4e0d\u7528 MSE \u4f5c\u4e3a\u635f\u5931\u51fd\u6570\u7684\u539f\u56e0\u4e4b\u4e00\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">1.2 \u5b66\u4e60\u7387\u8c03\u5ea6\uff1a\u63a7\u5236\u6bcf\u4e00\u6b65\u7684\u6b65\u957f<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b66\u4e60\u7387\u5927\u6982\u662f\u6df1\u5ea6\u5b66\u4e60\u8bad\u7ec3\u91cc\u6700\u8ba9\u4eba\u53c8\u7231\u53c8\u6068\u7684\u8d85\u53c2\u6570\u3002\u592a\u5927\uff0c\u635f\u5931\u9707\u8361\u751a\u81f3\u53d1\u6563\uff1b\u592a\u5c0f\uff0c\u6536\u655b\u6162\u8fd8\u5bb9\u6613\u6389\u8fdb\u978d\u70b9\u3002\u66f4\u590d\u6742\u7684\u662f\uff0c\u8bad\u7ec3\u4e0d\u540c\u9636\u6bb5\u9700\u8981\u4e0d\u540c\u7684\u5b66\u4e60\u7387\uff1a\u521a\u5f00\u59cb\u65f6\u68af\u5ea6\u65b9\u5411\u4e0d\u7a33\u5b9a\uff0c\u9700\u8981<strong>Warmup<\/strong>\u2014\u2014\u4ece\u5c0f\u5b66\u4e60\u7387\u7ebf\u6027\u722c\u5347\u5230\u76ee\u6807\u5b66\u4e60\u7387\uff1b\u4e2d\u540e\u671f\u5219\u9700\u8981\u9010\u6b65\u8870\u51cf\uff0c\u8ba9\u6a21\u578b\u5728\u4f4e\u8c37\u9644\u8fd1\u7cbe\u7ec6\u641c\u7d22\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Warmup \u5728\u5927 batch \u8bad\u7ec3\u4e2d\u5c24\u4e3a\u91cd\u8981\u3002\u5f53 batch size \u5f88\u5927\u65f6\uff0c\u68af\u5ea6\u4f30\u8ba1\u66f4\u7cbe\u786e\u4f46\u65b9\u5dee\u4e5f\u66f4\u5927\uff0c\u4e00\u5f00\u59cb\u5c31\u7528\u9ad8\u5b66\u4e60\u7387\u5bb9\u6613\u5bfc\u81f4\u53c2\u6570\u66f4\u65b0\u8fc7\u731b\u3002\u5148\u7528\u51e0\u4e2a epoch \u628a\u5b66\u4e60\u7387\u4ece\u5c0f\u6162\u6162 ramp \u4e0a\u6765\uff0c\u76f8\u5f53\u4e8e\u7ed9\u4f18\u5316\u5668\u4e00\u4e2a\u201c\u6696\u673a\u201d\u8fc7\u7a0b\u3002\u4e4b\u540e\u518d\u63a5\u5404\u79cd\u8870\u51cf\u7b56\u7565\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>StepLR<\/strong>\uff1a\u6bcf\u9694\u56fa\u5b9a epoch \u628a\u5b66\u4e60\u7387\u4e58\u4e00\u4e2a\u56e0\u5b50 <code>gamma<\/code>\uff0c\u7b80\u5355\u76f4\u63a5\u4f46\u8870\u51cf\u4e0d\u8fde\u7eed\uff1b<\/li>\n\n\n\n<li><strong>MultiStepLR<\/strong>\uff1a\u5728\u6307\u5b9a\u91cc\u7a0b\u7891\u5904\u6253\u6298\uff0c\u9002\u5408\u6709\u660e\u786e\u9636\u6bb5\u5212\u5206\u7684\u8bad\u7ec3\uff1b<\/li>\n\n\n\n<li><strong>ExponentialLR<\/strong>\uff1a\u6309 <code>lr = base_lr * gamma^epoch<\/code> \u8fde\u7eed\u8870\u51cf\uff1b<\/li>\n\n\n\n<li><strong>PolynomialLR<\/strong>\uff1a\u6309 <code>(1 - progress)^power<\/code> \u8870\u51cf\uff0cpower \u63a7\u5236\u66f2\u7ebf\u5f62\u72b6\uff1b<\/li>\n\n\n\n<li><strong>CosineAnnealingLR<\/strong>\uff1a\u6309\u4f59\u5f26\u66f2\u7ebf\u4ece <code>base_lr<\/code> \u964d\u5230 <code>eta_min<\/code>\uff0c\u5e73\u6ed1\u81ea\u7136\uff1b<\/li>\n\n\n\n<li><strong>CosineAnnealingWarmRestarts<\/strong>\uff1a\u4f59\u5f26\u9000\u706b\u5468\u671f\u91cd\u542f\uff0c\u5e38\u7528\u4e8e Transformer \u8bad\u7ec3\uff1b<\/li>\n\n\n\n<li><strong>WSD\uff08Warmup-Stable-Decay\uff09<\/strong>\uff1a\u524d\u6bb5\u4fdd\u6301 <code>base_lr<\/code> \u4e0d\u53d8\uff0c\u540e\u6bb5\u7ebf\u6027\u8870\u51cf\u5230 <code>end_lr<\/code>\uff0c\u8fd1\u5e74\u6765\u5728\u5927\u6a21\u578b\u8bad\u7ec3\u91cc\u5f88\u6d41\u884c\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">PyTorch \u91cc\u5e38\u7528 <code>LambdaLR<\/code> \u628a\u8fd9\u4e9b\u7b56\u7565\u50cf\u79ef\u6728\u4e00\u6837\u62fc\u63a5\uff0c\u518d\u7528 <code>SequentialLR<\/code> \u6309\u9636\u6bb5\u5207\u6362\u3002\u4f46\u4e0d\u7ba1\u516c\u5f0f\u600e\u4e48\u53d8\uff0c\u672c\u8d28\u4e0a\u90fd\u662f\u7ed9\u5b9a\u201c\u5f53\u524d\u6b65\u6570\u201d\uff0c\u8fd4\u56de\u4e00\u4e2a\u201c\u5b66\u4e60\u7387\u201d\u3002\u5b66\u4e60\u7387\u8c03\u5ea6\u5668\u672c\u8eab\u4e0d\u4fee\u6539\u6a21\u578b\u53c2\u6570\uff0c\u5b83\u53ea\u662f\u544a\u8bc9\u4f18\u5316\u5668\uff1a\u8fd9\u4e00\u6b65\u8be5\u7528\u591a\u5927\u7684\u6b65\u957f\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4ee5\u7ebf\u6027 Warmup \u4e3a\u4f8b\uff0c\u5176\u5b66\u4e60\u7387\u8ba1\u7b97\u516c\u5f0f\u53ef\u4ee5\u5199\u6210\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"python\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\"># t          : \u5f53\u524d\u6b65\u6570\uff08\u4ece 0 \u5f00\u59cb\u8ba1\u6570\uff09\n# W          : Warmup \u603b\u6b65\u6570\n# start_lr   : Warmup \u8d77\u59cb\u5b66\u4e60\u7387\n# base_lr    : Warmup \u7ed3\u675f\u540e\u7684\u76ee\u6807\u5b66\u4e60\u7387\n# \u6ce8\u610f\uff1a\u5f53 t == W \u65f6\uff0clr \u6070\u597d\u7b49\u4e8e base_lr\nif t &lt;= W:\n    lr = start_lr + (base_lr - start_lr) * t \/ W\nelse:\n    lr = decay_schedule(t - W)  # \u8fdb\u5165\u540e\u7eed\u8870\u51cf\u9636\u6bb5<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a\u7b80\u5355\u516c\u5f0f\u80cc\u540e\u662f\u5927\u6279\u91cf\u8bad\u7ec3\u7684\u7ecf\u9a8c\uff1a\u5728 Warmup \u7ed3\u675f\u65f6\uff0c\u53c2\u6570\u5df2\u7ecf\u79fb\u52a8\u5230\u4e00\u4e2a\u76f8\u5bf9\u7a33\u5b9a\u7684\u533a\u57df\uff0c\u6b64\u65f6\u518d\u7528 <code>base_lr<\/code> \u5c31\u4e0d\u5bb9\u6613\u9020\u6210\u9707\u8361\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">1.3 \u53c2\u6570\u521d\u59cb\u5316\uff1a\u7ed9\u8bad\u7ec3\u4e00\u4e2a\u5408\u7406\u7684\u8d77\u70b9<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u795e\u7ecf\u7f51\u7edc\u4e0d\u662f\u4ece\u4efb\u610f\u72b6\u6001\u90fd\u80fd\u8bad\u7ec3\u6210\u529f\u7684\u3002\u5982\u679c\u6743\u91cd\u5168\u8bbe\u4e3a 0\uff0c\u6240\u6709\u795e\u7ecf\u5143\u4f1a\u5b66\u5230\u540c\u6837\u7684\u4e1c\u897f\uff0c\u5bf9\u79f0\u6027\u65e0\u6cd5\u6253\u7834\uff1b\u5982\u679c\u6743\u91cd\u592a\u5927\u6216\u592a\u5c0f\uff0c\u4fe1\u53f7\u4f1a\u5728\u6df1\u5c42\u7f51\u7edc\u91cc\u7206\u70b8\u6216\u6d88\u5931\u3002\u53c2\u6570\u521d\u59cb\u5316\u7684\u76ee\u6807\uff0c\u5c31\u662f\u8ba9\u524d\u5411\u4f20\u64ad\u548c\u53cd\u5411\u4f20\u64ad\u4e2d\u7684\u4fe1\u53f7\u65b9\u5dee\u4fdd\u6301\u76f8\u5bf9\u7a33\u5b9a\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6700\u7ecf\u5178\u7684\u4e24\u79cd\u601d\u8def\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Xavier \/ Glorot \u521d\u59cb\u5316<\/strong>\uff1a\u5047\u8bbe\u6fc0\u6d3b\u51fd\u6570\u8fd1\u4f3c\u7ebf\u6027\uff0c\u901a\u8fc7 <code>fan_in<\/code> \u548c <code>fan_out<\/code> \u7684\u5e73\u5747\u503c\u6765\u63a7\u5236\u6743\u91cd\u65b9\u5dee\uff0c\u9002\u7528\u4e8e tanh\u3001sigmoid \u7b49\u5bf9\u79f0\u6fc0\u6d3b\u3002<\/li>\n\n\n\n<li><strong>Kaiming \/ He \u521d\u59cb\u5316<\/strong>\uff1a\u9488\u5bf9 ReLU \u7c7b\u6fc0\u6d3b\u51fd\u6570\uff0c\u628a\u65b9\u5dee\u4e58\u4ee5 2 \u6765\u8865\u507f ReLU \u7684\u201c\u622a\u65ad\u4e00\u534a\u201d\u6548\u5e94\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u7528\u4f2a\u4ee3\u7801\u8868\u793a\u5b83\u4eec\u7684\u91c7\u6837\u53c2\u6570\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"python\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\"># fan_in  : \u6743\u91cd\u5f20\u91cf\u8f93\u5165\u4fa7\u7684\u8fde\u63a5\u6570\n# fan_out : \u6743\u91cd\u5f20\u91cf\u8f93\u51fa\u4fa7\u7684\u8fde\u63a5\u6570\n# gain    : \u7531\u6fc0\u6d3b\u51fd\u6570\u51b3\u5b9a\u7684\u989d\u5916\u7f29\u653e\u56e0\u5b50\uff1bReLU \u9ed8\u8ba4 gain = sqrt(2)\n\n# Xavier \/ Glorot \u6b63\u6001\u5206\u5e03\nxavier_std = gain * sqrt(2.0 \/ (fan_in + fan_out))\n\n# Kaiming \/ He \u6b63\u6001\u5206\u5e03\nkaiming_std = gain \/ sqrt(fan_in)              # ReLU \u9ed8\u8ba4 gain = sqrt(2)\n\n# Kaiming \/ He \u5747\u5300\u5206\u5e03\uff1a\u533a\u95f4 [-bound, bound]\nkaiming_bound = gain * sqrt(3.0 \/ fan_in)<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u9009\u62e9\u54ea\u79cd\u521d\u59cb\u5316\uff0c\u5173\u952e\u770b\u6fc0\u6d3b\u51fd\u6570\u3002ReLU \u53ca\u5176\u53d8\u4f53\u51e0\u4e4e\u603b\u662f\u7528 Kaiming\uff1btanh\u3001sigmoid\u3001swish \u7b49\u5bf9\u79f0\u6fc0\u6d3b\u66f4\u9002\u5408 Xavier\u3002\u5982\u679c\u9009\u9519\uff0c\u6df1\u5c42\u7f51\u7edc\u524d\u51e0\u8f6e\u5c31\u4f1a\u51fa\u73b0\u6fc0\u6d3b\u503c\u8fc5\u901f\u584c\u9677\u6216\u7206\u70b8\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9664\u6b64\u4e4b\u5916\uff0c\u89c6\u89c9\u9886\u57df\u8fd8\u5e38\u7528<strong>\u622a\u65ad\u6b63\u6001\u5206\u5e03\uff08Truncated Normal\uff09<\/strong>\uff0c\u628a\u91c7\u6837\u503c\u9650\u5236\u5728 \u00b12\u03c3 \u4ee5\u5185\uff0c\u907f\u514d\u6781\u7aef\u503c\u7834\u574f\u65e9\u671f\u8bad\u7ec3\uff1b\u6279\u5f52\u4e00\u5316\u7684 gamma \u901a\u5e38\u521d\u59cb\u5316\u4e3a 1\u3001bias \u4e3a 0\uff1bResNet \u91cc\u8fd8\u4f1a\u628a\u90e8\u5206 BN \u7684 gamma \u521d\u59cb\u5316\u4e3a 0\uff0c\u5f62\u6210\u6240\u8c13\u7684 Zero Gamma \u521d\u59cb\u5316\uff0c\u8ba9\u6df1\u5c42\u7f51\u7edc\u5728\u8bad\u7ec3\u521d\u671f\u8fd1\u4f3c\u6052\u7b49\u6620\u5c04\uff0c\u7f13\u89e3\u68af\u5ea6\u4f20\u64ad\u56f0\u96be\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">PyTorch \u7684 <code>nn.init<\/code> \u6a21\u5757\u548c TensorFlow \u7684 <code>tf.keras.initializers<\/code> \u90fd\u5b9e\u73b0\u4e86\u4e0a\u8ff0\u65b9\u6848\u3002\u5b83\u4eec\u770b\u8d77\u6765\u53ea\u662f\u51e0\u884c\u968f\u673a\u6570\u751f\u6210\uff0c\u4f46\u80cc\u540e\u76f4\u63a5\u5f71\u54cd\u7684\u662f\u8bad\u7ec3\u80fd\u5426\u542f\u52a8\u3001\u6536\u655b\u901f\u5ea6\u548c\u6700\u7ec8\u7cbe\u5ea6\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6279\u5f52\u4e00\u5316\u5c42\u7684\u521d\u59cb\u5316\u4e5f\u6709\u8bb2\u7a76\u3002<code>\u03b3<\/code>\uff08weight\uff09\u521d\u59cb\u5316\u4e3a 1\u3001<code>\u03b2<\/code>\uff08bias\uff09\u521d\u59cb\u5316\u4e3a 0\uff0c\u662f\u4e3a\u4e86\u8ba9 BN \u5728\u8bad\u7ec3\u521d\u671f\u7b49\u4ef7\u4e8e\u6052\u7b49\u53d8\u6362\uff0c\u4e0d\u7834\u574f\u524d\u4e00\u5c42\u5df2\u7ecf\u521d\u59cb\u5316\u597d\u7684\u4fe1\u53f7\u5206\u5e03\u3002ResNet \u8fdb\u4e00\u6b65\u63d0\u51fa Zero Gamma\uff1a\u5bf9\u7279\u5b9a\u6b8b\u5dee\u5206\u652f\u672b\u5c3e\u7684 BN\uff0c\u628a <code>\u03b3<\/code> \u521d\u59cb\u5316\u4e3a 0\uff0c\u8fd9\u6837\u6574\u4e2a\u5206\u652f\u5728\u8bad\u7ec3\u521a\u5f00\u59cb\u65f6\u8f93\u51fa\u51e0\u4e4e\u4e3a 0\uff0c\u7f51\u7edc\u5148\u5b66\u4e60\u6d45\u5c42\u6052\u7b49\u6620\u5c04\uff0c\u518d\u9010\u6b65\u6fc0\u6d3b\u6df1\u5c42\u8def\u5f84\uff0c\u4ece\u800c\u7f13\u89e3\u6781\u6df1\u7f51\u7edc\u7684\u4f18\u5316\u56f0\u96be\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e8c\u3001Tech-Renaissance \u7684\u8bbe\u8ba1\u54f2\u5b66\uff1a\u65e0\u72b6\u6001\u3001\u7eaf\u51fd\u6570\u3001\u914d\u7f6e\u8fdb\u56fe<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 Tech-Renaissance \u91cc\uff0c\u635f\u5931\u51fd\u6570\u3001\u5b66\u4e60\u7387\u8c03\u5ea6\u5668\u3001\u53c2\u6570\u521d\u59cb\u5316\u5668\u90fd\u88ab\u8bbe\u8ba1\u6210<strong>\u65e0\u72b6\u6001\u7684\u7eaf\u914d\u7f6e\u5bf9\u8c61<\/strong>\u3002\u5b83\u4eec\u4e0d\u7ef4\u62a4 <code>current_step<\/code>\u3001<code>current_lr<\/code> \u4e4b\u7c7b\u7684\u53ef\u53d8\u72b6\u6001\uff0c\u4e5f\u4e0d\u53c2\u4e0e\u8fd0\u884c\u65f6\u7684\u9010\u6b65\u66f4\u65b0\u3002\u8fd9\u79cd\u8bbe\u8ba1\u4e0d\u662f\u5076\u7136\uff0c\u800c\u662f\u670d\u52a1\u4e8e\u6574\u4e2a\u6846\u67b6\u7684\u4e00\u4e2a\u6838\u5fc3\u7ea6\u675f\uff1a<strong>\u591a\u5361\u5206\u5e03\u5f0f\u8bad\u7ec3\u5fc5\u987b\u4fdd\u6301\u4e00\u81f4\uff0c\u4e14\u4e0d\u80fd\u4f9d\u8d56\u8fd0\u884c\u65f6\u540c\u6b65<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 PyTorch \u91cc\uff0c\u8c03\u5ea6\u5668\u901a\u5e38\u548c\u4f18\u5316\u5668\u7ed1\u5b9a\uff0c<code>scheduler.step()<\/code> \u6bcf\u8f6e\u4fee\u6539\u4f18\u5316\u5668\u5185\u90e8\u7684\u5b66\u4e60\u7387\u3002\u5982\u679c\u591a\u5361\u8bad\u7ec3\u65f6\u67d0\u4e2a rank \u7684 step \u6570\u56e0\u4e3a\u6570\u636e\u5206\u7247\u4e0d\u5747\u8861\u800c\u4e0d\u4e00\u81f4\uff0c\u5b66\u4e60\u7387\u5c31\u4f1a\u5206\u53c9\u3002Tech-Renaissance \u7684\u505a\u6cd5\u662f\u53cd\u8fc7\u6765\uff1a<strong>\u7ed9\u5b9a epoch \u548c batch\uff0c\u76f4\u63a5\u7b97\u51fa\u5b66\u4e60\u7387<\/strong>\u3002\u6240\u6709 rank \u53ea\u8981\u770b\u5230\u76f8\u540c\u7684 <code>(epoch, batch)<\/code>\uff0c\u5c31\u4f1a\u5f97\u5230\u76f8\u540c\u7684\u7ed3\u679c\uff0c\u65e0\u9700\u4efb\u4f55\u540c\u6b65\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u79cd\u65e0\u72b6\u6001\u8bbe\u8ba1\u4e5f\u7b80\u5316\u4e86 CUDA Graph \u7684\u6355\u83b7\u3002\u5982\u679c\u5b66\u4e60\u7387\u662f\u4e00\u4e2a\u5bf9\u8c61\u5185\u90e8\u7684 mutable \u72b6\u6001\uff0c\u6bcf\u6b21\u53d8\u5316\u90fd\u53ef\u80fd\u89e6\u53d1\u56fe\u7684\u91cd\u6784\uff1b\u800c\u628a\u5b83\u53d8\u6210 <code>(epoch, batch) \u2192 lr<\/code> \u7684\u7eaf\u51fd\u6570\uff0c\u5b66\u4e60\u7387\u5c31\u53ef\u4ee5\u88ab\u5f53\u4f5c\u4e00\u4e2a\u666e\u901a\u6807\u91cf\u8f93\u5165\uff0c\u8fd0\u884c\u65f6\u53ea\u9700\u8981\u66f4\u65b0\u8fd9\u4e2a\u6807\u91cf\u5373\u53ef\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e09\u4e2a\u914d\u7f6e\u5bf9\u8c61\u6700\u7ec8\u90fd\u4f1a\u8fdb\u5165\u7f16\u8bd1\u7ba1\u7ebf\uff1a\u635f\u5931\u53c2\u6570\u5199\u5165 <code>MemoryPlan<\/code> \u7684\u6807\u91cf\u5f20\u91cf\uff0c\u5b66\u4e60\u7387\u5728\u6bcf\u4e2a batch \u5f00\u5934\u901a\u8fc7\u4e00\u6b21\u8f7b\u91cf H2D \u62f7\u8d1d\u6ce8\u5165\u663e\u5b58\uff0c\u521d\u59cb\u5316\u7b56\u7565\u5219\u4ee5 <code>InitConfig<\/code> \u7684\u5f62\u5f0f\u7ed1\u5b9a\u5230\u6bcf\u4e00\u4e2a DTensor\u3002\u63a5\u4e0b\u6765\u6211\u4eec\u9010\u4e00\u5c55\u5f00\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e09\u3001\u635f\u5931\u51fd\u6570\uff1a\u4e00\u671f\u805a\u7126\u4ea4\u53c9\u71b5<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Tech-Renaissance \u5f53\u524d\u7248\u672c\u5728\u635f\u5931\u51fd\u6570\u4e0a\u975e\u5e38\u805a\u7126\uff0c\u53ea\u5b9e\u73b0\u4e86 <code>CrossEntropyLoss<\/code>\u3002\u8fd9\u4e0d\u662f\u80fd\u529b\u4e0a\u9650\uff0c\u800c\u662f\u8bad\u7ec3\u573a\u666f\u7684\u9009\u62e9\uff1a\u672c\u6846\u67b6\u76ee\u524d\u4e3b\u8981\u9762\u5411\u56fe\u50cf\u5206\u7c7b\u548c\u7c7b\u4f3c\u76d1\u7763\u5b66\u4e60\u4efb\u52a1\uff0c\u4ea4\u53c9\u71b5\u52a0\u6807\u7b7e\u5e73\u6ed1\u5df2\u7ecf\u8986\u76d6\u4e86\u4ece MNIST \u5230 ImageNet \u7684\u4e3b\u6d41\u9700\u6c42\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4ee3\u7801\u5c42\u9762\uff0c<code>CrossEntropyLoss<\/code> \u662f\u4e00\u4e2a\u7eaf\u914d\u7f6e\u7c7b\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ include\/renaissance\/algo\/loss.h\nclass CrossEntropyLoss {\npublic:\n    CrossEntropyLoss() = default;\n\n    CrossEntropyLoss&amp; label_smoothing(float value) {\n        TR_CHECK(value >= 0.0f &amp;&amp; value &lt;= 0.20001f, ValueError,\n                 \"label_smoothing must be in [0, 0.2], got \" &lt;&lt; value);\n        label_smoothing_ = value;\n        GlobalRegistry::instance().set_label_smoothing(value);\n        return *this;\n    }\n\n    [[nodiscard]] float label_smoothing() const noexcept {\n        return label_smoothing_;\n    }\n\nprivate:\n    float label_smoothing_ = 0.0f;\n};<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6ce8\u610f\u8fd9\u91cc\u7684\u4e00\u4e2a\u5173\u952e\u7ec6\u8282\uff1a<code>CrossEntropyLoss::label_smoothing()<\/code> \u4e00\u88ab\u8c03\u7528\uff0c\u5c31\u628a\u503c\u5199\u8fdb\u4e86 <code>GlobalRegistry<\/code>\uff0c\u56e0\u6b64\u5b83\u662f\u5168\u5c40\u552f\u4e00\u7684\u6743\u5a01\u503c\uff1b\u800c <code>DeepLearningTask::loss()<\/code> \u672c\u8eab\u53ea\u505a\u4e00\u4e2a\u9636\u6bb5\u68c0\u67e5\uff0c\u786e\u8ba4\u7528\u6237\u8fd8\u5728 PLANNING \u9636\u6bb5\u3002\u5bf9\u6846\u67b6\u6765\u8bf4\uff0c\u771f\u6b63\u88ab\u7f16\u8bd1\u5668\u8bfb\u53d6\u7684\u662f\u6ce8\u518c\u8868\u91cc\u7684\u503c\uff0c\u800c\u4e0d\u662f\u914d\u7f6e\u5bf9\u8c61\u5185\u90e8\u7684\u526f\u672c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7f16\u8bd1\u9636\u6bb5\uff0c<code>Compiler::compile()<\/code> \u4ece <code>GlobalRegistry<\/code> \u8bfb\u53d6 <code>label_smoothing<\/code>\uff0c\u628a\u5b83\u5199\u5165 <code>MemoryPlan<\/code> \u4e2d <code>Region::S_SCALAR_FP32<\/code> \u533a\u57df\u7684\u4e00\u4e2a\u6807\u91cf DTensor\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ src\/graph\/compiler.cpp\nfloat ls_val = GlobalRegistry::instance().label_smoothing();\nmemory_plans[s]->set_init_config(\n    memory_plans[s]->baseline().label_smoothing, kInitConstant(ls_val));<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a\u6807\u91cf\u968f\u540e\u88ab <code>SOFTMAX_CE_FP32_FWD<\/code>\u3001<code>SOFTMAX_CE_AMP_FWD<\/code> \u4ee5\u53ca\u5bf9\u5e94\u7684\u53cd\u5411\u7b97\u5b50\u4f5c\u4e3a\u8f93\u5165\u8bfb\u53d6\u3002\u5b9e\u9645\u7684\u524d\u5411 softmax\u3001\u4ea4\u53c9\u71b5\u3001\u53cd\u5411\u6c42\u5bfc\u90fd\u88ab\u878d\u5408\u6210\u4e00\u4e2a\u540e\u7aef\u7b97\u5b50\u6267\u884c\uff0c\u4e2d\u95f4\u4e0d\u4f1a\u628a\u5b8c\u6574 logits \u6216\u6982\u7387\u77e9\u9635\u5199\u56de\u5168\u5c40\u663e\u5b58\u3002\u8fd9\u4e2a\u8bbe\u8ba1\u4e0e PyTorch \u7684 fused cross entropy \u5728\u6570\u5b66\u4e0a\u7b49\u4ef7\uff0c\u4f46\u66f4\u7b26\u5408 Tech-Renaissance \u9759\u6001\u56fe\u7f16\u8bd1\u3001\u51cf\u5c11\u4e2d\u95f4\u5f20\u91cf\u7684\u6574\u4f53\u98ce\u683c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6807\u7b7e\u5e73\u6ed1\u7684\u8303\u56f4\u5728\u4ee3\u7801\u91cc\u88ab\u9650\u5236\u5728\u7ea6 <code>[0, 0.2]<\/code>\uff08\u4e0a\u754c\u5199\u6210 <code>0.20001f<\/code>\uff0c\u4fdd\u7559\u4e86\u4e00\u70b9\u6d6e\u70b9\u5bb9\u5dee\uff09\uff0c\u6ce8\u91ca\u91cc\u4e5f\u63d0\u5230 MLPerf Closed Division \u7684\u5e38\u89c1\u53d6\u503c\u662f <code>0<\/code> \u6216 <code>0.1<\/code>\u3002\u8fd9\u4e2a\u9650\u5236\u770b\u8d77\u6765\u5f88\u5c0f\uff0c\u4f46\u5b83\u53cd\u6620\u4e86\u6846\u67b6\u7684\u4e00\u4e2a\u6001\u5ea6\uff1a<strong>\u8bad\u7ec3\u914d\u7f6e\u4e0d\u662f\u4efb\u610f\u503c\u90fd\u53ef\u4ee5\uff0c\u800c\u662f\u8981\u548c\u4e3b\u6d41\u57fa\u51c6\u3001\u53ef\u590d\u73b0\u6027\u8981\u6c42\u5bf9\u9f50<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6709\u4eba\u53ef\u80fd\u4f1a\u95ee\uff0c\u4e3a\u4ec0\u4e48\u4e0d\u591a\u5b9e\u73b0\u51e0\u79cd\u635f\u5931\u51fd\u6570\uff1f\u7b54\u6848\u662f Tech-Renaissance \u76ee\u524d\u4e13\u6ce8\u4e8e\u5206\u7c7b\u4efb\u52a1\u7684\u7aef\u5230\u7aef\u541e\u5410\u4f18\u5316\uff0c\u635f\u5931\u51fd\u6570\u672c\u8eab\u5e76\u4e0d\u662f\u74f6\u9888\u3002\u771f\u6b63\u5f71\u54cd\u6027\u80fd\u7684\u662f\u635f\u5931\u4e0e\u524d\u5411\u3001\u53cd\u5411\u7b97\u5b50\u4e4b\u95f4\u7684\u878d\u5408\u65b9\u5f0f\u3002\u5148\u628a\u4ea4\u53c9\u71b5\u505a\u900f\u3001\u548c softmax \u4e00\u8d77\u878d\u8fdb\u540e\u7aef\uff0c\u6bd4\u4ed3\u4fc3\u652f\u6301\u5341\u79cd\u635f\u5931\u51fd\u6570\u66f4\u6709\u4ef7\u503c\u3002\u672a\u6765\u5982\u679c\u9700\u8981\u56de\u5f52\u3001\u68c0\u6d4b\u6216\u8bed\u8a00\u6a21\u578b\u4efb\u52a1\uff0c\u518d\u5728\u6b64\u57fa\u7840\u4e0a\u6269\u5c55\u4e5f\u6c34\u5230\u6e20\u6210\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u56db\u3001\u5b66\u4e60\u7387\u8c03\u5ea6\u5668\uff1a\u516b\u79cd\u7b56\u7565\uff0c\u4e00\u5957\u7eaf\u51fd\u6570\u6846\u67b6<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Tech-Renaissance \u5b9e\u73b0\u4e86\u516b\u79cd\u5b66\u4e60\u7387\u8c03\u5ea6\u5668\uff0c\u90fd\u7ee7\u627f\u81ea\u540c\u4e00\u4e2a\u62bd\u8c61\u57fa\u7c7b <code>LRScheduler<\/code>\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u8c03\u5ea6\u5668<\/th><th>\u6838\u5fc3\u884c\u4e3a<\/th><th>\u5173\u952e\u53c2\u6570<\/th><\/tr><\/thead><tbody><tr><td><code>ConstantLR<\/code><\/td><td>\u5168\u7a0b\u4fdd\u6301 <code>base_lr<\/code><\/td><td>\u2014<\/td><\/tr><tr><td><code>StepLR<\/code><\/td><td>\u6bcf <code>step_size<\/code> \u4e2a epoch \u628a\u5b66\u4e60\u7387\u4e58\u4ee5 <code>gamma<\/code><\/td><td><code>step_size<\/code>, <code>gamma<\/code><\/td><\/tr><tr><td><code>MultiStepLR<\/code><\/td><td>\u5728\u9884\u8bbe\u7684 <code>milestones<\/code> epoch \u5904\u4e58\u4ee5 <code>gamma<\/code><\/td><td><code>milestones<\/code>, <code>gamma<\/code><\/td><\/tr><tr><td><code>ExponentialLR<\/code><\/td><td>\u6bcf\u4e2a epoch \u628a\u5b66\u4e60\u7387\u4e58\u4ee5 <code>gamma<\/code><\/td><td><code>gamma<\/code><\/td><\/tr><tr><td><code>PolynomialLR<\/code><\/td><td>\u6309 <code>(1 - progress)^power<\/code> \u4ece <code>base_lr<\/code> \u8870\u51cf\u5230 <code>end_lr<\/code><\/td><td><code>power<\/code>, <code>end_lr<\/code><\/td><\/tr><tr><td><code>CosineAnnealingLR<\/code><\/td><td>\u6309\u4f59\u5f26\u66f2\u7ebf\u4ece <code>base_lr<\/code> \u8870\u51cf\u5230 <code>eta_min<\/code><\/td><td><code>eta_min<\/code><\/td><\/tr><tr><td><code>CosineAnnealingWithWarmRestartsLR<\/code><\/td><td>\u5468\u671f\u6027\u4f59\u5f26\u9000\u706b\uff0c\u5468\u671f\u957f\u5ea6\u6309 <code>T_mult<\/code> \u9012\u589e<\/td><td><code>T_0<\/code>, <code>T_mult<\/code>, <code>eta_min<\/code><\/td><\/tr><tr><td><code>WSDLR<\/code><\/td><td>\u524d <code>decay_start<\/code> \u6bd4\u4f8b\u4fdd\u6301\u7a33\u5b9a\uff0c\u4e4b\u540e\u7ebf\u6027\u8870\u51cf\u5230 <code>end_lr<\/code><\/td><td><code>decay_start<\/code>, <code>end_lr<\/code><\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u57fa\u7c7b\u7684\u8bbe\u8ba1\u975e\u5e38\u660e\u786e\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ include\/renaissance\/algo\/scheduler.h\nclass LRScheduler {\npublic:\n    virtual ~LRScheduler() = default;\n\n    virtual LRScheduler&amp; base_lr(float lr);\n    virtual LRScheduler&amp; warmup(int epochs);\n    virtual LRScheduler&amp; warmup_start_lr(float start_lr);\n    virtual LRScheduler&amp; warmup_start_factor(float factor);\n    virtual LRScheduler&amp; step_by_batch(bool v = true);\n    virtual LRScheduler&amp; step_by_epoch();\n\n    void prepare(int total_epochs, int steps_per_epoch);\n    float get_lr_by_batch(int batch_id) const;\n    float get_lr_by_epoch(int epoch_id) const;\n    bool is_step_by_batch() const noexcept { return step_by_batch_; }\n\nprotected:\n    virtual float compute_decay_lr(int decay_step, int total_decay) const = 0;\n    float compute_lr_at_step(int effective_step) const;\n    \/\/ ...\n};<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u57fa\u7c7b\u8d1f\u8d23\u7edf\u4e00\u88c1\u526a\u8fb9\u754c\u3001\u6267\u884c\u7ebf\u6027 Warmup\uff0c\u7136\u540e\u628a\u8870\u51cf\u9636\u6bb5\u59d4\u6258\u7ed9\u6d3e\u751f\u7c7b\u5b9e\u73b0\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ src\/algo\/scheduler.cpp\nfloat LRScheduler::compute_lr_at_step(int effective_step) const {\n    \/\/ effective_step : \u4ece\u8bad\u7ec3\u5f00\u59cb\u7d2f\u8ba1\u7684\u6b65\u6570\uff08\u5df2\u88ab\u88c1\u526a\u5230 [0, total_steps_]\uff09\n    \/\/ total_steps_   : \u8bad\u7ec3\u603b\u6b65\u6570 = total_epochs_ * steps_per_epoch_\n    \/\/ warmup_steps_  : Warmup \u603b\u6b65\u6570 = warmup_epochs_ * steps_per_epoch_\n    if (effective_step &lt; 0) effective_step = 0;\n    if (effective_step >= total_steps_) effective_step = total_steps_;\n\n    \/\/ \u7ebf\u6027 Warmup\uff1a\u5cf0\u503c\u51fa\u73b0\u5728 effective_step == warmup_steps_\n    if (warmup_steps_ > 0 &amp;&amp; effective_step &lt;= warmup_steps_) {\n        float progress = static_cast&lt;float>(effective_step)\n                       \/ static_cast&lt;float>(warmup_steps_);\n        float start_lr = resolve_warmup_start_lr();\n        return start_lr + (base_lr_ - start_lr) * progress;\n    }\n\n    \/\/ \u8870\u51cf\u9636\u6bb5\uff1a\u6d3e\u751f\u7c7b\u53ea\u9700\u63d0\u4f9b compute_decay_lr\n    int decay_step  = effective_step - warmup_steps_;\n    int total_decay = total_steps_ - warmup_steps_;\n    float lr = compute_decay_lr(decay_step, total_decay);\n    if (lr &lt; 0.0f || std::isnan(lr)) lr = 0.0f;\n    return lr;\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6d3e\u751f\u7c7b\u53ea\u9700\u8981\u5b9e\u73b0\u4e00\u4e2a\u7eaf\u51fd\u6570\u3002\u4f8b\u5982 <code>CosineAnnealingLR<\/code>\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ src\/algo\/scheduler.cpp\nfloat CosineAnnealingLR::compute_decay_lr(int decay_step, int total_decay) const {\n    if (total_decay &lt;= 0) return base_lr_;\n\n    int effective_total = step_by_batch_ ? total_decay : (total_decay - steps_per_epoch_);\n    if (effective_total &lt;= 0) return base_lr_;\n\n    float progress = static_cast&lt;float>(decay_step)\n                   \/ static_cast&lt;float>(effective_total);\n    if (progress > 1.0f) progress = 1.0f;\n\n    return eta_min_ + (base_lr_ - eta_min_)\n           * (1.0f + std::cos(static_cast&lt;float>(M_PI) * progress)) * 0.5f;\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><code>PolynomialLR<\/code> \u5219\u5bf9\u9f50\u4e86 TensorFlow \u7684 <code>polynomial_decay<\/code> \u884c\u4e3a\uff0c\u5206\u6bcd\u4f7f\u7528 <code>total_decay + 1<\/code>\uff1b<code>StepLR<\/code> \u548c <code>MultiStepLR<\/code> \u6309 epoch \u6570\u6253\u6298\uff1b<code>WSDLR<\/code> \u5148\u7a33\u5b9a\u540e\u7ebf\u6027\u8870\u51cf\u3002\u6bcf\u4e00\u79cd\u7b56\u7565\u90fd\u662f\u7eaf\u51fd\u6570\uff0c\u6ca1\u6709\u4efb\u4f55\u5185\u90e8\u72b6\u6001\u63a8\u8fdb\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.1 \u9010 epoch \u8fd8\u662f\u9010 batch<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\"><code>LRScheduler<\/code> \u652f\u6301\u4e24\u79cd\u6b65\u8fdb\u6a21\u5f0f\u3002<code>step_by_epoch()<\/code> \u8868\u793a\u6bcf\u4e2a epoch \u5185\u5b66\u4e60\u7387\u4e0d\u53d8\uff1b<code>step_by_batch()<\/code> \u5219\u628a <code>epoch * steps_per_epoch + batch_id<\/code> \u4f5c\u4e3a\u5168\u5c40\u6b65\u6570\uff0c\u8ba9\u5b66\u4e60\u7387\u5728\u6bcf\u4e2a batch \u90fd\u53d8\u5316\u3002<code>DeepLearningTask::fetch_lr_for_batch()<\/code> \u91cc\u7684\u903b\u8f91\u5f88\u76f4\u63a5\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ src\/task\/deep_learning_task.cpp\nfloat DeepLearningTask::fetch_lr_for_batch(int batch_id) const {\n    return std::visit([this, batch_id](auto&amp;&amp; sch) -> float {\n        using T = std::decay_t&lt;decltype(sch)>;\n        if constexpr (std::is_same_v&lt;T, std::monostate>) {\n            return 0.0f;\n        } else {\n            if (sch.is_step_by_batch()) {\n                int global_step = current_epoch_ * sch.steps_per_epoch() + batch_id;\n                return sch.get_lr_by_batch(global_step);\n            } else {\n                return sch.get_lr_by_epoch(current_epoch_);\n            }\n        }\n    }, sched_cfg_);\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u91cc <code>std::variant<\/code> \u4fdd\u5b58\u4e86\u5177\u4f53\u8c03\u5ea6\u5668\u7c7b\u578b\u3002\u7f16\u8bd1\u671f\u786e\u5b9a\u7c7b\u578b\uff0c\u8fd0\u884c\u671f\u65e0\u865a\u51fd\u6570\u5f00\u9500\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u53e6\u5916\uff0c<code>step_by_batch()<\/code> \u548c <code>step_by_epoch()<\/code> \u4e00\u65e6\u8c03\u7528\u5c31\u4f1a\u9501\u5b9a\u6b65\u8fdb\u6a21\u5f0f\uff0c\u540e\u7eed\u51b2\u7a81\u8c03\u7528\u4f1a\u76f4\u63a5\u629b\u9519\u3002\u8fd9\u662f\u4e00\u79cd\u9632\u5fa1\u6027\u8bbe\u8ba1\uff0c\u9632\u6b62\u7528\u6237\u94fe\u5f0f\u914d\u7f6e\u65f6\u5148\u5199 <code>.step_by_batch()<\/code> \u53c8\u5199 <code>.step_by_epoch()<\/code>\uff0c\u5bfc\u81f4\u5b66\u4e60\u7387\u66f4\u65b0\u8282\u594f\u51fa\u73b0\u6b67\u4e49\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.2 \u5982\u4f55\u628a\u5b66\u4e60\u7387\u9001\u8fdb CUDA Graph<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Tech-Renaissance \u7684\u8bad\u7ec3\u5faa\u73af\u6574\u4f53\u88ab\u6355\u83b7\u6210 CUDA Graph\u3002\u5982\u679c\u5b66\u4e60\u7387\u53d8\u5316\u9700\u8981\u91cd\u6784\u6574\u4e2a\u56fe\uff0c\u90a3\u5f00\u9500\u5c31\u592a\u5927\u4e86\u3002\u6846\u67b6\u7684\u89e3\u51b3\u529e\u6cd5\u662f\uff1a<strong>\u5b66\u4e60\u7387\u672c\u8eab\u662f\u4e00\u4e2a <code>S_SCALAR_FP32<\/code> \u6807\u91cf DTensor\uff0c\u8fd0\u884c\u65f6\u901a\u8fc7\u9501\u9875\u5185\u5b58\u505a\u4e00\u6b21 4 \u5b57\u8282\u7684 <code>cudaMemcpyAsync<\/code><\/strong>\u3002<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ src\/task\/deep_learning_task.cpp\nfloat* lr_dev_ptr = static_cast&lt;float*>(ctx.ptr_at(lr_dtensor_id_));\n\/\/ ...\nbool need_lr = is_step_by_batch_mode() || batch == 0;\nif (need_lr) {\n    lr = fetch_lr_for_batch(batch);\n    *lr_pinned_[rank] = lr;\n    cudaMemcpyAsync(lr_dev_ptr, lr_pinned_[rank], sizeof(float),\n                    cudaMemcpyHostToDevice, s_trans);\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><code>lr_pinned_<\/code> \u5728 <code>compile()<\/code> \u9636\u6bb5\u901a\u8fc7 <code>cudaMallocHost<\/code> \u5206\u914d\u3002\u8fd9\u4e2a\u5c0f\u62f7\u8d1d\u53d1\u751f\u5728\u4f20\u8f93\u6d41\u4e0a\uff0c\u548c\u8ba1\u7b97\u6d41\u5e76\u884c\uff0c4 \u5b57\u8282\u7684\u6570\u636e\u91cf\u51e0\u4e4e\u53ef\u4ee5\u5ffd\u7565\u3002\u4f18\u5316\u5668\u7b97\u5b50\u4ece\u540c\u4e00\u4e2a\u6807\u91cf\u5f20\u91cf\u8bfb\u53d6\u5f53\u524d\u5b66\u4e60\u7387\uff0c\u56e0\u6b64\u56fe\u7ed3\u6784\u65e0\u9700\u6539\u53d8\u3002\u8fd9\u662f\u9759\u6001\u56fe\u6846\u67b6\u4e0b\u5b9e\u73b0\u52a8\u6001\u5b66\u4e60\u7387\u7684\u4e00\u79cd\u52a1\u5b9e\u65b9\u6848\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u503c\u5f97\u8865\u5145\u7684\u662f\uff0c\u6846\u67b6\u91cc\u8fd8\u5b58\u5728\u4e00\u4e2a <code>StagingParamPool<\/code> \u5c0f\u53c2\u6570\u533a\uff08\u9ed8\u8ba4 256 \u5b57\u8282\uff0c64 \u4e2a FP32 \u7684\u9501\u9875 staging \u533a\uff09\uff0c\u5b83\u914d\u5408 <code>RANGE_H2D_COPY_DTENSOR<\/code> \u7b49\u8303\u56f4\u7b97\u5b50\uff0c\u53ef\u4ee5\u628a\u6781\u5c0f\u53c2\u6570\u7684 H2D \u62f7\u8d1d\u4e5f\u8868\u8fbe\u4e3a\u56fe\u4e2d\u7684\u4e00\u4e2a\u8282\u70b9\u3002\u4e0d\u8fc7\u5728\u5f53\u524d <code>DeepLearningTask<\/code> \u7684 GPU \u8bad\u7ec3\u5faa\u73af\u91cc\uff0c\u5b66\u4e60\u7387\u8d70\u7684\u662f\u76f4\u63a5\u7684 <code>cudaMemcpyAsync<\/code> \u8def\u5f84\uff1b<code>StagingParamPool<\/code> \u5728 <code>SimpleTask<\/code> \u7b49\u5176\u5b83\u4efb\u52a1\u8def\u5f84\u4e2d\u53d1\u6325\u4f5c\u7528\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e94\u3001\u53c2\u6570\u521d\u59cb\u5316\uff1a\u4ece\u7b56\u7565\u5230\u6bcf\u4e2a Region \u7684 8 \u5b57\u8282\u914d\u7f6e<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u53c2\u6570\u521d\u59cb\u5316\u5728 Tech-Renaissance \u91cc\u7531\u4e09\u4e2a\u6587\u4ef6\u5171\u540c\u627f\u62c5\uff1a<code>init_config.h<\/code> \u5b9a\u4e49\u7b56\u7565\u679a\u4e3e\u548c\u7d27\u51d1\u914d\u7f6e\uff0c<code>initializer.h\/.cpp<\/code> \u8d1f\u8d23\u7b56\u7565\u63a8\u5bfc\u548c\u6570\u5b66\u5b9e\u73b0\uff0c<code>TaskBase::init_all()<\/code> \u8d1f\u8d23\u5b9e\u9645\u586b\u5145\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5.1 InitConfig\uff1a8 \u5b57\u8282\u7684\u521d\u59cb\u5316\u7b56\u7565\u63cf\u8ff0\u7b26<\/h3>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ include\/renaissance\/core\/init_config.h\nenum class InitKind : uint8_t {\n    NONE, ZEROS, CONSTANTS,\n    KAIMING_NORMAL, KAIMING_UNIFORM,\n    XAVIER_NORMAL, XAVIER_UNIFORM,\n    TRUNC_NORMAL, FIXED_NORMAL,\n    STANDARD, ZERO_GAMMA\n};\n\nenum class FanMode : uint8_t { FAN_IN, FAN_OUT, FAN_AVG };\n\nstruct InitConfig {\n    float    scale = 1.0f;           \/\/ CONSTANTS=\u586b\u5145\u503c\uff0cKaiming\/Xavier=gain\uff0cFIXED_NORMAL=\u6807\u51c6\u5dee\n    InitKind kind  = InitKind::NONE; \/\/ \u521d\u59cb\u5316\u65b9\u6cd5\u79cd\u7c7b\n    FanMode  fan   = FanMode::FAN_IN; \/\/ fan \u8ba1\u7b97\u6a21\u5f0f\n    \/\/ \u7531\u4e8e\u5bf9\u9f50\uff0c\u7f16\u8bd1\u5668\u4f1a\u5728\u5c3e\u90e8\u8865 2 \u5b57\u8282\uff0c\u6574\u4f53\u6070\u597d 8 bytes\n};\n\nstatic_assert(sizeof(InitConfig) == 8, \"InitConfig must be exactly 8 bytes\");<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><code>InitConfig<\/code> \u88ab\u8bbe\u8ba1\u6210 8 \u5b57\u8282\u7684\u503c\u7c7b\u578b\uff0c\u53ef\u4ee5\u5728 CPU \u4fa7\u4f5c\u4e3a\u666e\u901a\u503c\u81ea\u7531\u62f7\u8d1d\u3002<code>scale<\/code> \u5b57\u6bb5\u6839\u636e <code>kind<\/code> \u6709\u4e0d\u540c\u542b\u4e49\uff1a\u5bf9 <code>CONSTANTS<\/code> \u5b83\u662f\u586b\u5145\u503c\uff0c\u5bf9 <code>KAIMING<\/code>\/<code>XAVIER<\/code> \u5b83\u662f gain\uff0c\u5bf9 <code>FIXED_NORMAL<\/code> \u5b83\u662f\u6807\u51c6\u5dee\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5.2 Initializer\uff1a\u94fe\u5f0f\u914d\u7f6e\u4e0e Region \u63a8\u5bfc<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u7528\u6237\u4fa7 API \u662f\u5178\u578b\u7684\u94fe\u5f0f\u98ce\u683c\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">Initializer init;\ninit.conv(InitKind::TRUNC_NORMAL)\n    .fc(InitKind::KAIMING_UNIFORM)\n    .bn(InitKind::STANDARD)\n    .fan(FanMode::FAN_IN)\n    .scale(1.0f);<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><code>Initializer<\/code> \u672c\u8eab\u53ea\u8bb0\u5f55\u5c42\u7c7b\u578b\u7ea7\u522b\u7684\u7b56\u7565\u3002\u771f\u6b63\u51b3\u5b9a\u67d0\u4e2a DTensor \u5982\u4f55\u521d\u59cb\u5316\uff0c\u662f\u5728\u7f16\u8bd1\u671f\u901a\u8fc7 <code>derive(Region)<\/code> \u5b8c\u6210\u7684\u3002<code>Region<\/code> \u662f MemoryPlan \u91cc\u6bcf\u4e2a\u5f20\u91cf\u7684\u8bed\u4e49\u5206\u533a\uff0c<code>Initializer::derive()<\/code> \u91c7\u7528\u4e09\u6bb5\u5f0f\u7ed3\u6784\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ src\/core\/initializer.cpp\nInitConfig Initializer::derive(Region region) const {\n    \/\/ \u7b2c\u4e00\u6bb5\uff1a\u504f\u7f6e\u533a \u2192 ZEROS\n    if (is_bias_region(region)) {\n        return InitConfig{0.0f, InitKind::ZEROS, FanMode::FAN_IN};\n    }\n\n    \/\/ \u52a8\u91cf\/\u901f\u5ea6\u6743\u91cd\u533a \u2192 ZEROS\n    if (region == Region::M_BN_WEIGHT  || region == Region::M_FC_WEIGHT  ||\n        region == Region::M_FIRST_CONV || region == Region::M_DEEP_CONV  ||\n        region == Region::V_BN_WEIGHT  || region == Region::V_FC_WEIGHT  ||\n        region == Region::V_FIRST_CONV || region == Region::V_DEEP_CONV) {\n        return InitConfig{0.0f, InitKind::ZEROS, FanMode::FAN_IN};\n    }\n\n    \/\/ BN running mean \/ variance \u7684\u521d\u59cb\u503c\n    if (region == Region::B_PREV_MEAN || region == Region::B_NEXT_MEAN) {\n        return InitConfig{0.0f, InitKind::CONSTANTS, FanMode::FAN_IN};\n    }\n    if (region == Region::B_PREV_VAR || region == Region::B_NEXT_VAR) {\n        return InitConfig{1.0f, InitKind::CONSTANTS, FanMode::FAN_IN};\n    }\n\n    \/\/ \u975e\u53c2\u6570\u533a \u2192 NONE\uff08\u663e\u5b58\u6c60\u5df2\u5168\u5c40\u7f6e\u96f6\uff09\n    if (!is_param_region(region)) {\n        return InitConfig{1.0f, InitKind::NONE, FanMode::FAN_IN};\n    }\n\n    \/\/ \u7b2c\u4e8c\u6bb5\uff1a\u6743\u91cd\u533a\u6309\u5c42\u7c7b\u578b\u5206\u53d1\n    if (region == Region::W_EQ_SCALE) {\n        return InitConfig{1.0f, InitKind::CONSTANTS, FanMode::FAN_IN};\n    }\n    if (is_bn_weight(region)) {\n        return InitConfig{1.0f, InitKind::CONSTANTS, FanMode::FAN_IN};\n    }\n    if (is_conv_weight(region)) {\n        float gain = global_scale_ * (conv_kind_ == InitKind::KAIMING_NORMAL ||\n                                      conv_kind_ == InitKind::KAIMING_UNIFORM\n                                      ? std::sqrt(2.0f \/ (1.0f + kaiming_a_ * kaiming_a_))\n                                      : 1.0f);\n        return InitConfig{gain, conv_kind_, fan_mode_};\n    }\n    if (is_fc_weight(region)) {\n        if (fc_kind_ == InitKind::FIXED_NORMAL) {\n            return InitConfig{fc_param_, InitKind::FIXED_NORMAL, fan_mode_};\n        }\n        float gain = global_scale_ * (fc_kind_ == InitKind::KAIMING_NORMAL ||\n                                      fc_kind_ == InitKind::KAIMING_UNIFORM\n                                      ? std::sqrt(2.0f \/ (1.0f + kaiming_a_ * kaiming_a_))\n                                      : 1.0f);\n        return InitConfig{gain, fc_kind_, fan_mode_};\n    }\n    \/\/ ...\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u6bb5\u4ee3\u7801\u4f53\u73b0\u4e86 Tech-Renaissance \u521d\u59cb\u5316\u7cfb\u7edf\u7684\u6838\u5fc3\u601d\u8def\uff1a<strong>\u4e0d\u662f\u6309\u201c\u5c42\u201d\u53bb\u521d\u59cb\u5316\uff0c\u800c\u662f\u6309\u201cRegion\u201d\u53bb\u521d\u59cb\u5316<\/strong>\u3002\u56e0\u4e3a MemoryPlan \u5df2\u7ecf\u628a\u540c\u8bed\u4e49\u7684\u5f20\u91cf\u96c6\u4e2d\u6392\u653e\uff0c\u504f\u7f6e\u3001BN weight\u3001Conv weight\u3001FC weight \u5404\u81ea\u843d\u5728\u4e0d\u540c\u7684 Region \u4e0a\u3002<code>derive()<\/code> \u8d1f\u8d23\u8986\u76d6\u6240\u6709\u9700\u8981\u663e\u5f0f\u521d\u59cb\u5316\u7684\u53c2\u6570\u533a\uff1b\u50cf\u68af\u5ea6\u3001\u52a8\u91cf\u3001EMA \u7b49\u533a\u8981\u4e48\u5728\u5168\u5c40 <code>memset<\/code> \u4e2d\u5df2\u7ecf\u4e3a\u96f6\uff0c\u8981\u4e48\u7531\u540e\u7eed\u7684\u8303\u56f4\u7b97\u5b50\uff08RangeOp\uff09\u6279\u91cf\u5904\u7406\u3002<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u987a\u5e26\u8bf4\u660e\uff1a<code>Region<\/code> \u679a\u4e3e\u76ee\u524d\u5171\u6709 69 \u4e2a\u69fd\u4f4d\uff08\u542b\u8fb9\u754c\u54e8\u5175\uff09\uff0c\u5176\u4e2d\u5b9e\u9645\u547d\u540d\u7684\u8bed\u4e49 Region \u4e3a 68 \u4e2a\uff0c\u4f46 <code>derive()<\/code> \u53ea\u9700\u5904\u7406\u771f\u6b63\u8fdb\u5165\u521d\u59cb\u5316\u7ba1\u7ebf\u7684\u53c2\u6570\u533a\u5373\u53ef\u3002<\/p>\n<\/blockquote>\n\n\n\n<p class=\"wp-block-paragraph\">\u5bf9\u4e8e ResNet \u4e2d\u7684 Zero Gamma \u7b56\u7565\uff0c\u6846\u67b6\u5df2\u7ecf\u9884\u7559\u4e86 <code>mark_bn3()<\/code> \u63a5\u53e3\uff1a\u7f16\u8bd1\u5668\u5728\u5c55\u5f00\u6b8b\u5dee\u5757\u65f6\u4f1a\u8bb0\u5f55\u4e3b\u5206\u652f\u6700\u540e\u4e00\u4e2a BN \u7684 weight DTensor id\u3002\u7406\u8bba\u4e0a <code>init_all()<\/code> \u9636\u6bb5\u5e94\u628a\u8fd9\u4e9b\u7279\u5b9a id \u7684 BN weight \u4ece\u9ed8\u8ba4\u7684 1.0 \u8986\u76d6\u4e3a 0.0\uff0c\u4ece\u800c\u8ba9\u6574\u4e2a\u6b8b\u5dee\u5206\u652f\u5728\u8bad\u7ec3\u521d\u671f\u8fd1\u4f3c\u6052\u7b49\u6620\u5c04\u3002\u4e0d\u8fc7\u5f53\u524d\u7248\u672c\u4e2d\uff0cZero Gamma \u7684\u6700\u7ec8\u8986\u76d6\u903b\u8f91\u56e0\u4e3a\u9a8c\u8bc1\u5c1a\u672a\u5b8c\u5168\u901a\u8fc7\u800c\u88ab\u663e\u5f0f\u7981\u7528\uff1b<code>mark_bn3<\/code> \u7684\u6807\u8bb0\u94fe\u8def\u5df2\u7ecf\u5c31\u7eea\uff0c\u540e\u7eed\u9a8c\u8bc1\u5b8c\u6210\u540e\u5373\u53ef\u6253\u5f00\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5.3 \u6570\u5b66\u5b9e\u73b0\uff1a\u4ece fan \u5230\u968f\u673a\u6570<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\"><code>Initializer::apply_to_tensor()<\/code> \u662f\u5b9e\u9645\u751f\u6210\u968f\u673a\u6570\u7684\u5165\u53e3\uff0c\u5168\u90e8\u5728 CPU \u7aef\u5b8c\u6210\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ src\/core\/initializer.cpp\nvoid Initializer::apply_to_tensor(Tensor&amp; t, const Shape&amp; shape, InitConfig cfg) {\n    switch (cfg.kind) {\n        case InitKind::TRUNC_NORMAL: {\n            int64_t fan = compute_fan(shape, cfg.fan);\n            float std = std::sqrt(cfg.scale \/ static_cast&lt;float>(fan));\n            t.truncated_normal(0.0f, std, -2.0f * std, 2.0f * std);\n            return;\n        }\n        case InitKind::KAIMING_NORMAL: {\n            int64_t fan = compute_fan(shape, cfg.fan);\n            float std = cfg.scale \/ std::sqrt(static_cast&lt;float>(fan));\n            t.normal(0.0f, std);\n            return;\n        }\n        case InitKind::KAIMING_UNIFORM: {\n            int64_t fan = compute_fan(shape, cfg.fan);\n            float bound = cfg.scale * std::sqrt(3.0f \/ static_cast&lt;float>(fan));\n            t.uniform(-bound, bound);\n            return;\n        }\n        case InitKind::XAVIER_NORMAL: {\n            int64_t fi = compute_fan(shape, FanMode::FAN_IN);\n            int64_t fo = compute_fan(shape, FanMode::FAN_OUT);\n            float std = cfg.scale * std::sqrt(2.0f \/ static_cast&lt;float>(fi + fo));\n            t.normal(0.0f, std);\n            return;\n        }\n        case InitKind::XAVIER_UNIFORM: {\n            int64_t fi = compute_fan(shape, FanMode::FAN_IN);\n            int64_t fo = compute_fan(shape, FanMode::FAN_OUT);\n            float bound = cfg.scale * std::sqrt(6.0f \/ static_cast&lt;float>(fi + fo));\n            t.uniform(-bound, bound);\n            return;\n        }\n        \/\/ ...\n    }\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u5377\u79ef\u6743\u91cd\u7684 fan \u8ba1\u7b97\u57fa\u4e8e Tech-Renaissance \u7edf\u4e00\u7684 <strong>KRSC \u5e03\u5c40<\/strong>\uff1a<code>[K=outC, R=kH, S=kW, C=inC]<\/code>\uff0c\u56e0\u6b64 <code>fan_in = C \u00d7 R \u00d7 S<\/code>\uff0c<code>fan_out = K \u00d7 R \u00d7 S<\/code>\u3002FC \u5c42\u5219\u6309\u666e\u901a\u4e8c\u7ef4\u6743\u91cd\u5904\u7406\u3002\u8fd9\u4e9b\u516c\u5f0f\u548c PyTorch <code>nn.init.kaiming_normal_<\/code>\u3001<code>nn.init.xavier_uniform_<\/code> \u7684\u5b9a\u4e49\u4e00\u81f4\uff0c\u53ea\u662f\u5b9e\u73b0\u88ab\u6574\u5408\u8fdb\u4e86\u9759\u6001\u521d\u59cb\u5316\u7ba1\u7ebf\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5.4 \u521d\u59cb\u5316\u6d41\u7a0b\uff1arank 0 \u751f\u6210\uff0cH2D\uff0c\u5e7f\u64ad<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u7f16\u8bd1\u5b8c\u6210\u540e\uff0c<code>TaskBase::init_all()<\/code> \u4f1a\u904d\u5386\u6240\u6709 DTensor\uff0c\u6839\u636e <code>init_config<\/code> \u9010\u4e2a\u521d\u59cb\u5316\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ src\/task\/task_base.cpp\nvoid TaskBase::init_all() {\n    for (const auto&amp; dtensor : active_memory_plan_->dtensors()) {\n        init(dtensor);\n    }\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><code>TaskBase::init()<\/code> \u7684\u5b9e\u9645\u6d41\u7a0b\u662f\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u5728 CPU \u7aef\u7533\u8bf7\u4e00\u4e2a\u540c\u6837 shape \u7684 <code>Tensor<\/code>\uff1b<\/li>\n\n\n\n<li>\u8c03\u7528 <code>Initializer::apply_to_tensor()<\/code> \u751f\u6210\u968f\u673a\u6570\uff1b<\/li>\n\n\n\n<li>\u628a\u8fd9\u4e2a CPU \u5f20\u91cf\u4f20\u8f93\u5230 rank 0 \u7684 GPU\uff1b<\/li>\n\n\n\n<li>\u5982\u679c\u6709\u591a\u5361\uff0c\u901a\u8fc7 NCCL <code>broadcast_from_rank0<\/code> \u628a rank 0 \u7684\u6743\u91cd\u5e7f\u64ad\u5230\u5176\u4ed6 rank\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u6837\u505a\u7684\u597d\u5904\u662f\uff1a<strong>\u6240\u6709 GPU \u4e0a\u7684\u521d\u59cb\u6743\u91cd\u5b8c\u5168\u4e00\u81f4<\/strong>\uff0c\u907f\u514d\u5206\u5e03\u5f0f\u8bad\u7ec3\u56e0\u4e3a\u521d\u59cb\u503c\u4e0d\u540c\u5bfc\u81f4\u68af\u5ea6 AllReduce \u540e\u4ecd\u7136\u4e0d\u4e00\u81f4\u3002\u968f\u673a\u6570\u751f\u6210\u7531 Philox \u8ba1\u6570\u5668 RNG \u8d1f\u8d23\uff0c\u4e0b\u4e00\u7bc7\u6587\u7ae0\u4f1a\u4e13\u95e8\u8bb2\u5b83\u5982\u4f55\u4fdd\u8bc1\u591a\u7ebf\u7a0b\u3001\u591a\u5361\u4e0b\u7684\u53ef\u590d\u73b0\u6027\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9700\u8981\u5f3a\u8c03\u7684\u662f\uff0c\u521d\u59cb\u5316\u867d\u7136\u53d1\u751f\u5728\u8bad\u7ec3\u5faa\u73af\u4e4b\u524d\uff0c\u4f46\u5b83\u5bf9\u6574\u4e2a\u5206\u5e03\u5f0f\u7cfb\u7edf\u7684\u6b63\u786e\u6027\u81f3\u5173\u91cd\u8981\u3002\u5982\u679c\u6bcf\u4e2a rank \u5404\u81ea\u72ec\u7acb\u751f\u6210\u968f\u673a\u521d\u59cb\u503c\uff0c\u5373\u4f7f\u540e\u7eed AllReduce \u540c\u6b65\u68af\u5ea6\uff0c\u5404 rank \u7684\u53c2\u6570\u66f4\u65b0\u8d77\u70b9\u4e5f\u4e0d\u540c\uff0c\u8bad\u7ec3\u7ed3\u679c\u5728\u6570\u5b66\u4e0a\u5c31\u4e0d\u518d\u7b49\u4ef7\u4e8e\u5355\u5361\u5927 batch\u3002Tech-Renaissance \u9009\u62e9 rank 0 \u751f\u6210\u3001\u5168\u5361\u5e7f\u64ad\u7684\u65b9\u6848\uff0c\u6b63\u662f\u4e3a\u4e86\u786e\u4fdd\u8fd9\u4e2a\u8d77\u70b9\u7684\u4e00\u81f4\u6027\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u516d\u3001\u4ece\u7528\u6237\u914d\u7f6e\u5230\u8bad\u7ec3\u56fe\uff1a\u4e00\u6761\u5b8c\u6574\u7684\u94fe\u8def<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u628a\u8fd9\u4e9b\u4e1c\u897f\u4e32\u8d77\u6765\u770b\uff0cTech-Renaissance \u7684\u8bad\u7ec3\u7b97\u6cd5\u914d\u7f6e\u5c42\u5927\u81f4\u662f\u8fd9\u6837\u5de5\u4f5c\u7684\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ tests\/example\/mlp_mnist.cpp\nDeepLearningTask task;\ntask.model(mlp)\n    .loss(CrossEntropyLoss().label_smoothing(0.1f))\n    .optimizer(AdamW().weight_decay(1e-4f))\n    .scheduler(CosineAnnealingLR().base_lr(0.001f).warmup(5))\n    .initializer(Initializer()\n        .fc(InitKind::KAIMING_UNIFORM)\n        .fan(FanMode::FAN_IN))\n    .total_epochs(kTotalEpochs);\n\ntask.compile();\nauto result = task.run();<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u7528\u6237\u9762\u5bf9\u7684 API \u975e\u5e38\u8584\u3002<code>DeepLearningTask<\/code> \u5728 <code>on_prepare()<\/code> \u91cc\u5b8c\u6210\u51e0\u4ef6\u4e8b\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u5982\u679c\u6ca1\u6709\u663e\u5f0f\u6307\u5b9a\u521d\u59cb\u5316\u5668\uff0c\u5c31\u4f7f\u7528\u9ed8\u8ba4 <code>Initializer()<\/code>\uff1b<\/li>\n\n\n\n<li>\u6839\u636e\u4f18\u5316\u5668\u7c7b\u578b\u8bbe\u7f6e <code>PlanConfig<\/code>\uff1b<\/li>\n\n\n\n<li>\u4ece BluePrint \u751f\u6210 <code>ArchPlan<\/code>\uff0c\u518d\u8c03\u7528 <code>Compiler::compile(..., initializer_, variant_specs)<\/code>\uff1b<\/li>\n\n\n\n<li>\u7f16\u8bd1\u5b8c\u6210\u540e\uff0c\u67e5\u627e <code>Region::S_SCALAR_FP32<\/code> \u4e2d\u7684\u5b66\u4e60\u7387\u6807\u91cf DTensor\uff1b<\/li>\n\n\n\n<li>\u4e3a\u4f18\u5316\u5668\u6807\u91cf\uff08momentum\u3001weight_decay\u3001beta2\u3001eps \u7b49\uff09\u8bbe\u7f6e <code>InitConfig<\/code>\uff1b<\/li>\n\n\n\n<li>\u628a\u8bad\u7ec3\u56fe\u548c\u63a8\u7406\u56fe\u52a0\u5165 <code>TaskBase<\/code>\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">Compiler \u5728 <code>create_memory_plans()<\/code> \u9636\u6bb5\u4f1a\u5bf9\u6bcf\u4e2a\u5206\u914d\u51fa\u7684 DTensor \u8c03\u7528 <code>initializer.derive(desc.region)<\/code>\uff0c\u628a\u521d\u59cb\u5316\u7b56\u7565\u5199\u5165 MemoryPlan\u3002\u968f\u540e <code>TaskBase::init_all()<\/code> \u5728\u8bad\u7ec3\u5f00\u59cb\u524d\u4e00\u6b21\u6027\u6267\u884c\u521d\u59cb\u5316\u3002\u6574\u4e2a\u8fc7\u7a0b\u6ca1\u6709\u52a8\u6001\u5206\u914d\uff0c\u6ca1\u6709\u8fd0\u884c\u65f6\u72b6\u6001\u7ade\u4e89\uff0c\u6240\u6709\u914d\u7f6e\u90fd\u5728\u7f16\u8bd1\u671f\u843d\u8fdb\u56fe\u548c\u5185\u5b58\u5e03\u5c40\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e03\u3001\u5c0f\u7ed3<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u635f\u5931\u51fd\u6570\u3001\u5b66\u4e60\u7387\u8c03\u5ea6\u3001\u53c2\u6570\u521d\u59cb\u5316\uff0c\u8fd9\u4e9b\u5e38\u5e38\u88ab\u521d\u5b66\u8005\u5f53\u4f5c\u201c\u8c03\u53c2\u4e09\u677f\u65a7\u201d\u7684\u4e1c\u897f\uff0c\u5728\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u5185\u90e8\u5176\u5b9e\u662f\u4e00\u6574\u5957\u9700\u8981\u7cbe\u5fc3\u8bbe\u8ba1\u7684\u57fa\u7840\u8bbe\u65bd\u3002\u5b83\u4eec\u5fc5\u987b\u548c\u8ba1\u7b97\u56fe\u3001\u663e\u5b58\u89c4\u5212\u3001CUDA Graph \u6355\u83b7\u3001\u5206\u5e03\u5f0f\u4e00\u81f4\u6027\u7d27\u5bc6\u914d\u5408\uff0c\u624d\u80fd\u5728\u4fdd\u8bc1\u6b63\u786e\u7684\u524d\u63d0\u4e0b\u4e0d\u6210\u4e3a\u6027\u80fd\u74f6\u9888\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Tech-Renaissance \u7684\u505a\u6cd5\u53ef\u4ee5\u6982\u62ec\u4e3a\u4e09\u70b9\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u914d\u7f6e\u5bf9\u8c61\u65e0\u72b6\u6001<\/strong>\uff1a\u635f\u5931\u3001\u8c03\u5ea6\u5668\u3001\u521d\u59cb\u5316\u5668\u90fd\u4e0d\u7ef4\u62a4\u8fd0\u884c\u65f6\u72b6\u6001\uff0c\u7ed9\u5b9a\u8f93\u5165\u76f4\u63a5\u8fd4\u56de\u7ed3\u679c\uff0c\u5929\u7136\u9002\u5408\u591a\u5361\u5e76\u884c\u3002<\/li>\n\n\n\n<li><strong>\u914d\u7f6e\u8fdb\u56fe\u3001\u8fdb Region<\/strong>\uff1a\u635f\u5931\u53c2\u6570\u548c\u5b66\u4e60\u7387\u53d8\u6210 MemoryPlan \u6807\u91cf\u5f20\u91cf\uff0c\u521d\u59cb\u5316\u7b56\u7565\u53d8\u6210\u6bcf\u4e2a DTensor \u7684 8 \u5b57\u8282 <code>InitConfig<\/code>\uff0c\u907f\u514d\u8fd0\u884c\u65f6\u53cd\u590d\u67e5\u8be2\u5bf9\u8c61\u5c5e\u6027\u3002<\/li>\n\n\n\n<li><strong>\u4fdd\u6301\u4e0e\u4e3b\u6d41\u505a\u6cd5\u7684\u6570\u5b66\u4e00\u81f4\u6027<\/strong>\uff1a\u4ea4\u53c9\u71b5\u52a0\u6807\u7b7e\u5e73\u6ed1\u3001\u516b\u79cd\u5e38\u89c1\u5b66\u4e60\u7387\u8c03\u5ea6\u3001Kaiming\/Xavier\/Truncated Normal \u521d\u59cb\u5316\uff0c\u516c\u5f0f\u90fd\u4e0e PyTorch\u3001TensorFlow \u5bf9\u9f50\uff0c\u786e\u4fdd\u8fc1\u79fb\u6a21\u578b\u65f6\u4e0d\u9700\u8981\u91cd\u65b0\u8c03\u53c2\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e0b\u4e00\u7bc7\uff0c\u6211\u4eec\u5c06\u8fdb\u5165\u8bad\u7ec3\u6d41\u7a0b\u4e2d\u53e6\u4e00\u4e2a\u5bb9\u6613\u88ab\u5ffd\u89c6\u5374\u81f3\u5173\u91cd\u8981\u7684\u4e3b\u9898\u2014\u2014\u968f\u673a\u6570\u751f\u6210\uff0c\u4ee5\u53ca Tech-Renaissance \u5982\u4f55\u7528 Philox \u8ba1\u6570\u5668\u5b9e\u73b0\u7aef\u5230\u7aef\u7684\u786e\u5b9a\u6027\u8bad\u7ec3\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u2014\u2014\u201c\u4e00\u4e2a\u4eba\u7528AI\u5982\u4f55\u5199\u51fa\u6bd4PyTorch\u66f4\u5feb\u7684\u81ea\u7814\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u201d\u7cfb\u5217\u6587\u7ae0\u4e4b\u4e8c\u5341\u4e8c \u524d\u9762\u7684\u6587\u7ae0\u82b1\u4e86\u5f88\u5927\u7bc7\u5e45\u8bb2  [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":650,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-container-style":"default","site-container-layout":"default","site-sidebar-layout":"default","disable-article-header":"default","disable-site-header":"default","disable-site-footer":"default","disable-content-area-spacing":"default","footnotes":""},"categories":[15],"tags":[],"class_list":["post-553","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-15"],"_links":{"self":[{"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/posts\/553","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/comments?post=553"}],"version-history":[{"count":3,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/posts\/553\/revisions"}],"predecessor-version":[{"id":706,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/posts\/553\/revisions\/706"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/media\/650"}],"wp:attachment":[{"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/media?parent=553"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/categories?post=553"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/tags?post=553"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}