{"id":493,"date":"2026-07-08T03:01:16","date_gmt":"2026-07-07T19:01:16","guid":{"rendered":"https:\/\/tech-renaissance.cn\/blog\/?p=493"},"modified":"2026-07-08T22:19:13","modified_gmt":"2026-07-08T14:19:13","slug":"blueprint-dsl%ef%bc%9a%e7%94%a8%e5%a3%b0%e6%98%8e%e5%bc%8f%e8%af%ad%e6%b3%95%e5%ae%9a%e4%b9%89%e7%a5%9e%e7%bb%8f%e7%bd%91%e7%bb%9c","status":"publish","type":"post","link":"https:\/\/tech-renaissance.cn\/blog\/index.php\/2026\/07\/08\/493\/","title":{"rendered":"(12) BluePrint DSL\uff1a\u7528\u58f0\u660e\u5f0f\u8bed\u6cd5\u5b9a\u4e49\u795e\u7ecf\u7f51\u7edc"},"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\u5341\u4e8c<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e0a\u4e00\u7bc7\u6211\u4eec\u8bb2\u4e86 FusedNormalization\uff0c\u628a\u6570\u636e\u589e\u5f3a\u7684\u6700\u540e\u4e00\u6b65\u6536\u5c3e\u64cd\u4f5c\u878d\u8fdb\u4e86\u4e00\u6b21\u5185\u5b58\u904d\u5386\u3002\u4ece\u8fd9\u4e00\u7bc7\u5f00\u59cb\uff0c\u6211\u4eec\u628a\u89c6\u7ebf\u4ece\u6570\u636e\u4fa7\u79fb\u56de\u6a21\u578b\u4fa7\uff0c\u8fdb\u5165 Tech-Renaissance \u7684\u56fe\u7f16\u8bd1\u4e0e\u5185\u5b58\u89c4\u5212\u7bc7\u7ae0\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bad\u7ec3\u4e00\u4e2a\u6df1\u5ea6\u795e\u7ecf\u7f51\u7edc\uff0c\u7b2c\u4e00\u6b65\u5f53\u7136\u662f<strong>\u5b9a\u4e49\u6a21\u578b<\/strong>\u3002\u4f46&#8221;\u5b9a\u4e49\u6a21\u578b&#8221;\u8fd9\u4ef6\u4e8b\u770b\u4f3c\u7b80\u5355\uff0c\u5176\u5b9e\u7275\u52a8\u540e\u7eed\u4e00\u5927\u7247\u8bbe\u8ba1\uff1a\u4f60\u7684\u6a21\u578b\u63cf\u8ff0\u662f\u53ea\u7ed9\u4eba\u7c7b\u770b\u7684\uff0c\u8fd8\u662f\u80fd\u7ed9\u7f16\u8bd1\u5668\u770b\uff1f\u7f16\u8bd1\u5668\u80fd\u4e0d\u80fd\u4ece\u4e2d\u8bc6\u522b\u51fa\u53ef\u4ee5\u878d\u5408\u7684\u7b97\u5b50\u6a21\u5f0f\uff1f\u80fd\u4e0d\u80fd\u81ea\u52a8\u63a8\u5bfc\u5f62\u72b6\u3001\u5206\u914d\u663e\u5b58\u3001\u751f\u6210 CUDA Graph\uff1f\u5982\u679c\u6a21\u578b\u5b9a\u4e49\u5c42\u5199\u5f97\u592a&#8221;\u81ea\u7531&#8221;\uff0c\u7f16\u8bd1\u5668\u5c31\u53ea\u80fd\u770b\u5230\u4e00\u5806\u96f6\u6563\u7684\u5f20\u91cf\u64cd\u4f5c\uff0c\u5f88\u591a\u5168\u5c40\u4f18\u5316\u5c31\u65e0\u4ece\u8c08\u8d77\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e00\u7bc7\u6211\u4eec\u5148\u4e0d\u8bb2\u7f16\u8bd1\u5668\u672c\u8eab\uff0c\u800c\u662f\u8bb2 Tech-Renaissance \u63d0\u4f9b\u7ed9\u7528\u6237\u7684\u6a21\u578b\u5b9a\u4e49\u5165\u53e3\uff1a<strong>BluePrint DSL<\/strong>\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e00\u3001\u6a21\u578b\u5b9a\u4e49\u5c42\uff1a\u6846\u67b6\u4e0e\u7528\u6237\u4e4b\u95f4\u7684&#8221;\u56fe\u7eb8&#8221;<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u51e0\u4e4e\u6240\u6709\u7684\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u90fd\u9700\u8981\u56de\u7b54\u540c\u4e00\u4e2a\u95ee\u9898\uff1a\u7528\u6237\u5e94\u8be5\u4ee5\u4ec0\u4e48\u65b9\u5f0f\u63cf\u8ff0\u4e00\u4e2a\u795e\u7ecf\u7f51\u7edc\uff1f<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u56de\u987e\u5386\u53f2\uff0c\u8fd9\u6761\u8def\u7684\u6f14\u53d8\u975e\u5e38\u6e05\u6670\u3002\u65e9\u671f\u7684 <strong>Caffe<\/strong> \u76f4\u63a5\u7528 Protobuf \u914d\u7f6e\u6587\u4ef6\u63cf\u8ff0\u7f51\u7edc\uff1a\u6bcf\u4e00\u5c42\u662f\u4e00\u4e2a\u5e26\u53c2\u6570\u7684\u8282\u70b9\uff0c\u5c42\u4e0e\u5c42\u4e4b\u95f4\u6309\u987a\u5e8f\u8fde\u63a5\u3002\u8fd9\u79cd\u5199\u6cd5\u975e\u5e38\u58f0\u660e\u5f0f\uff0c\u4e5f\u4fbf\u4e8e\u90e8\u7f72\uff0c\u4f46\u9047\u5230 ResNet \u7684\u6b8b\u5dee\u8fde\u63a5\u3001Inception \u7684\u591a\u5206\u652f\u7ed3\u6784\u65f6\u5c31\u663e\u5f97\u7b28\u62d9\uff0c\u5faa\u73af\u548c\u6761\u4ef6\u5206\u652f\u66f4\u662f\u96be\u4ee5\u8868\u8fbe\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">TensorFlow 1.x \u7ee7\u627f\u4e86\u58f0\u660e\u5f0f\u7684\u601d\u8def\uff0c\u4f46\u628a\u5b83\u4ece\u914d\u7f6e\u6587\u4ef6\u5347\u7ea7\u4e3a Python API\uff1a\u4f60\u5148\u521b\u5efa <code>tf.Graph<\/code>\uff0c\u518d\u7528 <code>tf.layers.conv2d<\/code>\u3001<code>tf.layers.batch_normalization<\/code> \u7b49\u51fd\u6570\u628a\u8282\u70b9\u62fc\u8d77\u6765\u3002\u8fd9\u79cd\u65b9\u5f0f\u6bd4 Caffe \u7075\u6d3b\u5f97\u591a\uff0c\u5374\u4f9d\u7136\u8981\u6c42\u7528\u6237\u5728\u5fc3\u91cc\u533a\u5206&#8221;\u5efa\u56fe\u9636\u6bb5&#8221;\u548c&#8221;\u6267\u884c\u9636\u6bb5&#8221;\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">PyTorch \u5219\u8d70\u4e86\u53e6\u4e00\u6761\u8def\u3002\u5b83\u7528 Python \u7684\u9762\u5411\u5bf9\u8c61\u673a\u5236\u628a\u5c42\u5c01\u88c5\u6210 <code>nn.Module<\/code>\uff0c\u7528\u6237\u5728 <code>forward()<\/code> \u91cc\u7528\u666e\u901a Python \u4ee3\u7801\u63cf\u8ff0\u8ba1\u7b97\u8fc7\u7a0b\u3002\u8fd9\u79cd\u547d\u4ee4\u5f0f\u98ce\u683c\u5f7b\u5e95\u62b9\u5e73\u4e86&#8221;\u5efa\u56fe&#8221;\u4e0e&#8221;\u6267\u884c&#8221;\u7684\u8fb9\u754c\uff0c\u8c03\u8bd5\u4f53\u9a8c\u6781\u4f73\uff0c\u4e5f\u8ba9 PyTorch \u5728\u5b66\u672f\u754c\u8fc5\u901f\u53cd\u8d85 TensorFlow\u3002\u4f46\u4ee3\u4ef7\u662f\u6846\u67b6\u5728\u8fd0\u884c\u65f6\u770b\u5230\u7684\u4e0d\u662f\u4e00\u5f20\u5b8c\u6574\u7684\u56fe\uff0c\u800c\u662f\u4e00\u6bb5\u5df2\u7ecf\u8dd1\u5b8c\u7684\u4ee3\u7801\uff1b\u8981\u505a\u5168\u5c40\u4f18\u5316\uff0c\u53ea\u80fd\u540e\u671f\u901a\u8fc7 <code>torch.compile<\/code> \u6216 TorchScript \u53bb\u6355\u83b7\u548c\u91cd\u5efa\u56fe\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u540e\u6765\u51fa\u73b0\u7684\u65b0\u6846\u67b6\u5219\u5c1d\u8bd5\u878d\u5408\u4e24\u8005\u3002JAX \u7684 <strong>Flax<\/strong>\u3001<strong>Haiku<\/strong>\u3001Google \u7684 <strong>Keras Functional API<\/strong>\uff0c\u90fd\u503e\u5411\u4e8e\u7528\u58f0\u660e\u5f0f\u7ec4\u5408\u63cf\u8ff0\u56fa\u5b9a\u7ed3\u6784\uff0c\u540c\u65f6\u4fdd\u7559 Python \u7684\u7075\u6d3b\u6027\u5904\u7406\u52a8\u6001\u903b\u8f91\u3002ONNX \u66f4\u8fdb\u4e00\u6b65\uff0c\u76f4\u63a5\u628a\u8bad\u7ec3\u597d\u7684\u6a21\u578b\u5bfc\u51fa\u6210\u4e00\u79cd\u4e0e\u6846\u67b6\u65e0\u5173\u7684\u58f0\u660e\u5f0f\u56fe\u8868\u793a\uff0c\u7528\u4e8e\u63a8\u7406\u90e8\u7f72\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e9b\u8def\u7ebf\u7684\u5206\u6b67\uff0c\u672c\u8d28\u4e0a\u662f<strong>\u7075\u6d3b\u6027\u4e0e\u53ef\u4f18\u5316\u6027<\/strong>\u7684\u6743\u8861\u3002\u58f0\u660e\u5f0f\u5199\u6cd5\u628a\u6a21\u578b\u7684\u5b8c\u6574\u7ed3\u6784\u4e00\u6b21\u6027\u4ea4\u7ed9\u6846\u67b6\uff0c\u6846\u67b6\u56e0\u6b64\u53ef\u4ee5\uff1a\u5728\u7f16\u8bd1\u671f\u5b8c\u6210\u6240\u6709\u5f62\u72b6\u63a8\u5bfc\uff0c\u907f\u514d\u8fd0\u884c\u65f6\u518d\u4e34\u65f6\u5206\u914d\u5f20\u91cf\uff1b\u8bc6\u522b Conv+BN+ReLU\u3001Bottleneck \u7b49\u53ef\u878d\u5408\u6a21\u5f0f\uff0c\u8c03\u7528\u4e13\u7528\u878d\u5408 kernel\uff1b\u6309\u5168\u5c40\u751f\u547d\u5468\u671f\u89c4\u5212\u663e\u5b58\uff0c\u628a\u6743\u91cd\u3001\u68af\u5ea6\u3001\u4e2d\u95f4\u7279\u5f81\u56fe\u3001\u4e34\u65f6\u7f13\u51b2\u533a\u5206\u95e8\u522b\u7c7b\u5730\u653e\u8fdb\u56fa\u5b9a Region\uff1b\u4e00\u6b21\u6027\u6355\u83b7\u6574\u4e2a\u8bad\u7ec3\u5faa\u73af\u7684 CUDA Graph\uff0c\u6d88\u9664 kernel launch \u5f00\u9500\u3002PyTorch \u7684 <code>torch.compile<\/code> \u505a\u7684\u4e8b\u60c5\uff0c\u672c\u8d28\u4e0a\u4e5f\u662f\u5728\u52a8\u6001\u56fe\u6267\u884c\u4e4b\u540e\uff0c\u52aa\u529b\u628a\u4fe1\u606f\u91cd\u65b0\u6536\u96c6\u6210\u9759\u6001\u56fe\uff1b\u800c Tech-Renaissance \u4ece\u4e00\u5f00\u59cb\u5c31\u628a\u9759\u6001\u7ed3\u6784\u66b4\u9732\u7ed9\u7528\u6237\uff0c\u8df3\u8fc7\u4e86&#8221;\u6355\u83b7&#8221;\u8fd9\u4e00\u6b65\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u91cc\u8fd8\u6709\u4e00\u4e2a\u503c\u5f97\u533a\u5206\u7684\u6982\u5ff5\uff1a<strong>\u5185\u90e8 DSL\uff08embedded DSL\uff09<\/strong>\u4e0e<strong>\u5916\u90e8 DSL<\/strong>\u3002Caffe \u7684 Protobuf \u914d\u7f6e\u662f\u5916\u90e8 DSL\uff0c\u5b83\u9700\u8981\u72ec\u7acb\u7684\u89e3\u6790\u5668\u548c\u7f16\u8bd1\u5668\uff1bPyTorch \u7684 Python API \u5219\u4e0d\u7b97 DSL\uff0c\u5b83\u5c31\u662f Python \u672c\u8eab\u3002Tech-Renaissance \u7684 BluePrint \u5c5e\u4e8e\u5185\u90e8 DSL\uff1a\u5b83\u4ecd\u7136\u4f7f\u7528 C++ \u7684\u8bed\u6cd5\u548c\u7c7b\u578b\u7cfb\u7edf\uff0c\u4f46\u901a\u8fc7\u4e00\u7ec4\u7cbe\u5fc3\u8bbe\u8ba1\u7684\u5de5\u5382\u51fd\u6570\u548c\u7ec4\u5408\u7b97\u5b50\uff0c\u5f62\u6210\u4e86\u4e00\u79cd\u4e13\u95e8\u7528\u4e8e\u63cf\u8ff0\u795e\u7ecf\u7f51\u7edc\u7684&#8221;\u5b50\u8bed\u8a00&#8221;\u3002\u7528\u6237\u5199\u7684\u662f\u5408\u6cd5 C++ \u4ee3\u7801\uff0c\u5374\u611f\u89c9\u50cf\u5728\u5199\u6a21\u578b\u7ed3\u6784\u56fe\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Tech-Renaissance \u7684\u76ee\u6807\u662f\u5728\u4e3b\u6d41 CV \u8bad\u7ec3\u4efb\u52a1\u4e0a\u505a\u5230\u6781\u81f4\u6027\u80fd\uff0c\u56e0\u6b64\u5b83\u9009\u62e9\u4e86\u58f0\u660e\u5f0f\uff1a\u4ece\u7528\u6237\u5199\u4e0b BluePrint \u7684\u90a3\u4e00\u523b\u8d77\uff0c\u6846\u67b6\u5c31\u62ff\u5230\u4e86\u6574\u5f20\u7f51\u7edc\u7684&#8221;\u56fe\u7eb8&#8221;\uff0c\u53ef\u4ee5\u5728\u7f16\u8bd1\u671f\u505a\u5f62\u72b6\u63a8\u5bfc\u3001\u7b97\u5b50\u878d\u5408\u3001\u663e\u5b58\u89c4\u5212\u548c CUDA Graph \u6355\u83b7\u3002BluePrint DSL \u5c31\u662f\u8fd9\u6761\u8bbe\u8ba1\u8def\u7ebf\u7684\u7528\u6237\u4fa7\u5165\u53e3\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e8c\u3001BluePrint \u957f\u4ec0\u4e48\u6837<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5148\u770b\u4e00\u4e2a\u6700\u76f4\u89c2\u7684\u4f8b\u5b50\u3002\u5728 Tech-Renaissance \u91cc\u5b9a\u4e49\u4e00\u4e2a 4 \u5c42 MLP\uff0c\u53ea\u9700\u8981\u8fd9\u51e0\u884c\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=\"\">BluePrint mlp = seq(\n    fc(1024, true), relu(),\n    fc(512, true),  relu(),\n    fc(256, true),  relu(),\n    fc(10, true)\n);<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6ca1\u6709\u7c7b\u5b9a\u4e49\uff0c\u6ca1\u6709 <code>forward()<\/code>\uff0c\u6ca1\u6709\u663e\u5f0f\u7684\u5f20\u91cf\u5bf9\u8c61\u3002<code>seq<\/code> \u8868\u793a\u987a\u5e8f\u7ec4\u5408\uff0c<code>fc<\/code> \u662f\u5168\u8fde\u63a5\u5c42\uff0c<code>relu<\/code> \u662f\u6fc0\u6d3b\u51fd\u6570\u3002\u6574\u4e2a\u6a21\u578b\u5c31\u662f\u4e00\u68f5<strong>\u5c42\u7ec4\u5408\u6811<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u518d\u770b VGG-16-BN \u7684\u9aa8\u5e72\u90e8\u5206\u3002PyTorch \u91cc\u4f60\u9700\u8981\u5199\u4e00\u5806 <code>nn.Conv2d<\/code>\u3001<code>nn.BatchNorm2d<\/code>\u3001<code>nn.ReLU<\/code>\u3001\u6a21\u5757\u5d4c\u5957\uff1b\u800c\u5728 Tech-Renaissance \u91cc\uff0c\u5b83\u66f4\u63a5\u8fd1\u4f60\u8111\u5b50\u91cc\u7684\u7ed3\u6784\u56fe\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=\"\">BluePrint vgg16bn = seq(\n    \/\/ Block 1: 224x224 -> 112x112\n    conv(64, 3, 1, 1), bn(), relu(),\n    conv(64, 3, 1, 1), bn(), relu(),\n    maxpool(2, 2, 0),\n\n    \/\/ Block 2: 112x112 -> 56x56\n    conv(128, 3, 1, 1), bn(), relu(),\n    conv(128, 3, 1, 1), bn(), relu(),\n    maxpool(2, 2, 0),\n\n    \/\/ ...\n    flatten(),\n    fc(4096, true), relu(), dropout(0.5),\n    fc(4096, true), relu(), dropout(0.5),\n    fc(1000, true)\n);<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u79cd\u5199\u6cd5\u7684\u597d\u5904\u662f\uff1a<strong>\u4f60\u5199\u7684\u51e0\u4e4e\u5c31\u662f\u7f51\u7edc\u7ed3\u6784\u672c\u8eab<\/strong>\u3002\u6bcf\u4e2a\u5de5\u5382\u51fd\u6570\u7684\u53c2\u6570\u4e5f\u76f4\u63a5\u5bf9\u5e94\u5377\u79ef\u6838\u5927\u5c0f\u3001\u6b65\u957f\u3001\u586b\u5145\u3001\u8f93\u51fa\u901a\u9053\u8fd9\u4e9b\u8d85\u53c2\u6570\uff0c\u6ca1\u6709\u989d\u5916\u7684\u5305\u88c5\u5c42\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">BluePrint \u672c\u8eab\u5219\u662f\u4e00\u4e2a\u66f4\u8f7b\u91cf\u7684\u95e8\u9762\u7c7b\u3002\u5728 <code>blueprint.h<\/code> \u4e2d\uff0c\u5b83\u53ea\u6709\u51e0\u53e5\u8bdd\u7684\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=\"\">class BluePrint {\npublic:\n    BluePrint() = default;\n    BluePrint(Layer root) : root_(std::move(root)) {}\n    BluePrint&amp; operator=(Layer root) { root_ = std::move(root); return *this; }\n    bool empty() const noexcept { return !root_.valid(); }\nprivate:\n    Layer root_;\n    friend class ArchPlan;\n};<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><code>BluePrint<\/code> \u51e0\u4e4e\u4e0d\u589e\u52a0\u4efb\u4f55\u8bed\u4e49\uff0c\u5b83\u53ea\u662f\u628a\u4e00\u68f5 <code>Layer<\/code> \u6811\u6253\u5305\u6210&#8221;\u6a21\u578b&#8221;\u7684\u6982\u5ff5\u3002\u7528\u6237\u53ef\u4ee5\u50cf\u4e0a\u9762\u90a3\u6837\u76f4\u63a5\u6784\u9020\uff0c\u4e5f\u53ef\u4ee5\u5148\u62fc\u597d <code>Layer<\/code> \u518d\u8d4b\u503c\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=\"\">Layer backbone = seq(conv(64,3,1,1), bn(), relu(), maxpool(2,2,0));\nBluePrint model = backbone;<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u968f\u540e <code>DeepLearningTask::model(model)<\/code> \u4f1a\u628a\u5b83\u4ea4\u7ed9 ArchPlan \u548c Compiler\uff0c\u6574\u4e2a\u8fc7\u7a0b\u4e0d\u9700\u8981\u7528\u6237\u624b\u52a8\u6307\u5b9a\u8f93\u5165\u8f93\u51fa\u5f20\u91cf\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9700\u8981\u5f3a\u8c03\u7684\u662f\uff0cBluePrint \u53ea\u8d1f\u8d23\u63cf\u8ff0<strong>\u6a21\u578b\u7ed3\u6784<\/strong>\u3002\u635f\u5931\u51fd\u6570\u3001\u4f18\u5316\u5668\u3001\u5b66\u4e60\u7387\u8c03\u5ea6\u3001\u53c2\u6570\u521d\u59cb\u5316\u3001\u6570\u636e\u9884\u5904\u7406\u8fd9\u4e9b\u8bad\u7ec3\u914d\u7f6e\uff0c\u90fd\u4e0d\u5728 BluePrint \u91cc\uff0c\u800c\u662f\u901a\u8fc7 <code>DeepLearningTask<\/code> \u7684\u94fe\u5f0f API \u5355\u72ec\u8bbe\u7f6e\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=\"\">DeepLearningTask task;\ntask.model(vgg16bn)\n    .loss(CrossEntropyLoss().label_smoothing(0.1f))\n    .optimizer(SGD().momentum(0.9f).weight_decay(1e-4f))\n    .scheduler(CosineAnnealingLR().base_lr(0.36f).warmup(10));<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u79cd\u5206\u79bb\u8ba9\u6a21\u578b\u7ed3\u6784\u53ef\u4ee5\u590d\u7528\uff1a\u540c\u4e00\u4efd BluePrint \u65e2\u53ef\u4ee5\u914d SGD \u8bad ImageNet\uff0c\u4e5f\u53ef\u4ee5\u914d AdamW \u505a\u8fc1\u79fb\u5b66\u4e60\uff0c\u800c\u6a21\u578b\u5b9a\u4e49\u672c\u8eab\u4e0d\u9700\u8981\u6539\u52a8\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5bf9\u4e8e ResNet-50 \u8fd9\u79cd\u5305\u542b\u6b8b\u5dee\u548c\u91cd\u590d\u5806\u53e0\u7684\u7f51\u7edc\uff0cBluePrint \u4e5f\u80fd\u7528\u5f88\u5c11\u7684\u4ee3\u7801\u8868\u8fbe\u3002\u501f\u52a9 <code>block()<\/code> \u548c <code>repeat()<\/code>\uff0c\u6574\u4e2a ResNet-50 \u53ef\u4ee5\u5728\u5341\u51e0\u884c\u5185\u5199\u5b8c\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=\"\">BluePrint resnet50 = seq(\n    conv(64, 7, 2, 3), bn(), relu(),\n    maxpool(3, 2, 1),\n\n    \/\/ stage 1: 256 channels, 3 blocks\n    block(64, 256, BlockStyle::RESNET_1_3_1),\n    repeat(block(64, 256, BlockStyle::RESNET_1_3_1), 2),\n\n    \/\/ stage 2: 512 channels, 4 blocks, downsample\n    block(128, 512, BlockStyle::RESNET_1_3_1_DS),\n    repeat(block(128, 512, BlockStyle::RESNET_1_3_1), 3),\n\n    \/\/ stage 3: 1024 channels, 6 blocks, downsample\n    block(256, 1024, BlockStyle::RESNET_1_3_1_DS),\n    repeat(block(256, 1024, BlockStyle::RESNET_1_3_1), 5),\n\n    \/\/ stage 4: 2048 channels, 3 blocks, downsample\n    block(512, 2048, BlockStyle::RESNET_1_3_1_DS),\n    repeat(block(512, 2048, BlockStyle::RESNET_1_3_1), 2),\n\n    gap(),\n    fc(1000, true)\n);<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><code>block()<\/code> \u63cf\u8ff0\u4e00\u4e2a Bottleneck \u6b8b\u5dee\u5757\uff0c<code>repeat()<\/code> \u628a\u5b83\u91cd\u590d\u591a\u6b21\u3002\u7b2c\u4e00\u4e2a\u53c2\u6570 <code>mid_ch<\/code> \u662f Bottleneck \u7684\u4e2d\u95f4\uff08bottleneck\uff09\u901a\u9053\u6570\uff0c\u7b2c\u4e8c\u4e2a\u53c2\u6570 <code>out_ch<\/code> \u662f\u8f93\u51fa\u901a\u9053\u6570\uff1b\u5f53\u8f93\u5165\u901a\u9053\u6570\u4e0e <code>out_ch<\/code> \u76f8\u540c\u65f6\uff0cArchPlan \u4f1a\u81ea\u52a8\u8bc6\u522b\u4e3a identity shortcut\uff0c\u5426\u5219\u4f7f\u7528 1&#215;1 \u6295\u5f71 shortcut\u3002\u6b8b\u5dee\u8fde\u63a5\u548c downsample \u90fd\u88ab\u5c01\u88c5\u5728 <code>BlockStyle<\/code> \u91cc\uff0c\u4e0d\u9700\u8981\u7528\u6237\u624b\u52a8\u5199 <code>add2<\/code>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">MobileNetV2 \u7684\u5199\u6cd5\u540c\u6837\u975e\u5e38\u7d27\u51d1\u3002\u5012\u7f6e\u6b8b\u5dee\u5757\u901a\u8fc7 <code>mbconv(expand_ratio, out_ch, stride)<\/code> \u8868\u8fbe\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=\"\">BluePrint mobilenetv2 = seq(\n    conv(32, 3, 2, 1), bn(), relu6(),\n\n    \/\/ stage 1\n    mbconv(1, 16, 1),\n\n    \/\/ stage 2\n    mbconv(6, 24, 2),\n    mbconv(6, 24, 1),\n\n    \/\/ stage 3\n    mbconv(6, 32, 2),\n    mbconv(6, 32, 1),\n    mbconv(6, 32, 1),\n\n    \/\/ ...\n    conv(1280, 1, 1, 0), bn(), relu6(),\n    gap(),\n    fc(1000, true)\n);<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6bcf\u4e2a <code>mbconv<\/code> \u5728\u5f00\u542f AMP \u878d\u5408\u65f6\u4f1a\u88ab\u6620\u5c04\u4e3a <code>InvResidualIdentity<\/code> \u6216 <code>InvResidualNoShortcut<\/code> \u878d\u5408\u5c42\uff0c\u8c03\u7528\u540e\u7aef\u7684 <code>INVRESIDUAL_AMP_FWD\/BWD<\/code> \u4e13\u7528 kernel\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e09\u3001Layer\uff1a\u4e00\u68f5\u4e0d\u53ef\u53d8\u7684\u6811<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">BluePrint \u7684\u5e95\u5c42\u7c7b\u578b\u662f <code>Layer<\/code>\u3002\u5728 <code>include\/renaissance\/graph\/blueprint.h<\/code> \u4e2d\uff0c\u5b83\u7684\u5b9e\u73b0\u975e\u5e38\u8f7b\u91cf\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=\"\">class Layer {\npublic:\n    Layer() = default;\n    bool valid() const noexcept { return static_cast&lt;bool>(node_); }\nprivate:\n    std::shared_ptr&lt;const Node> node_;\n    explicit Layer(std::shared_ptr&lt;const Node> n) : node_(std::move(n)) {}\n    \/\/ ...\n};\n\nstruct Layer::Node {\n    detail::NodeKind kind;\n    detail::Payload  payload;\n};<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><code>Layer<\/code> \u662f\u4e00\u4e2a Pimpl \u98ce\u683c\u7684\u5c0f\u628a\u624b\uff0c\u5185\u90e8\u6301\u6709\u4e00\u4e2a <code>const Node<\/code> \u7684\u5171\u4eab\u6307\u9488\u3002<code>Node<\/code> \u7531\u4e24\u90e8\u5206\u7ec4\u6210\uff1a<code>NodeKind<\/code> \u8868\u793a\u5b83\u662f\u54ea\u79cd\u5c42\uff0c<code>Payload<\/code> \u662f <code>std::variant<\/code> \u627f\u8f7d\u7684\u5177\u4f53\u53c2\u6570\u3002\u8fd9\u68f5\u6811\u6709\u4e09\u4e2a\u91cd\u8981\u7279\u6027\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u7b2c\u4e00\uff0c\u4e0d\u53ef\u53d8<\/strong>\u3002\u8282\u70b9\u4e00\u65e6\u521b\u5efa\u5c31\u4e0d\u4f1a\u88ab\u4fee\u6539\uff0c\u7236\u5b50\u5173\u7cfb\u5728\u6784\u9020\u65f6\u786e\u5b9a\u3002\u8fd9\u6d88\u9664\u4e86\u5f88\u591a\u7531\u53ef\u53d8\u72b6\u6001\u5e26\u6765\u7684\u9690\u60a3\uff1a\u4f60\u4e0d\u7528\u62c5\u5fc3\u67d0\u4e2a\u5c42\u5728\u7f16\u8bd1\u8fc7\u7a0b\u4e2d\u88ab\u6084\u6084\u6539\u6389\uff0c\u4e5f\u4e0d\u7528\u62c5\u5fc3\u591a\u4e2a\u6a21\u578b\u526f\u672c\u4e4b\u95f4\u4e92\u76f8\u6c61\u67d3\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u7b2c\u4e8c\uff0c\u503c\u8bed\u4e49<\/strong>\u3002<code>Layer<\/code> \u53ef\u4ee5\u50cf\u6574\u6570\u4e00\u6837\u88ab\u62f7\u8d1d\u3001\u4f20\u9012\u3001\u653e\u8fdb <code>std::vector<\/code>\uff0c\u4f46\u62f7\u8d1d\u53ea\u662f\u5171\u4eab\u6307\u9488\uff0c\u6ca1\u6709\u6df1\u5c42\u590d\u5236\u3002\u7ec4\u5408 <code>seq(a, b, c)<\/code> \u65f6\uff0c\u4e0d\u4f1a\u4ea7\u751f\u5927\u91cf\u4e34\u65f6\u5bf9\u8c61\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u7b2c\u4e09\uff0c\u7ec4\u5408\u4f18\u5148<\/strong>\u3002\u57fa\u7840\u5c42\uff08<code>conv<\/code>\u3001<code>bn<\/code>\u3001<code>relu<\/code>\u3001<code>fc<\/code> \u7b49\uff09\u662f\u53f6\u5b50\u8282\u70b9\uff1b\u7ec4\u5408\u5c42\uff08<code>seq<\/code>\u3001<code>add2<\/code>\u3001<code>repeat<\/code>\uff09\u662f\u5185\u90e8\u8282\u70b9\u3002<code>add2<\/code> \u8868\u8fbe\u6b8b\u5dee\u8fde\u63a5\uff0c<code>repeat<\/code> \u8868\u8fbe\u91cd\u590d\u5806\u53e0\uff0c\u4e24\u8005\u914d\u5408\u5c31\u80fd\u63cf\u8ff0 ResNet\u3001MobileNet \u8fd9\u7c7b\u590d\u6742\u62d3\u6251\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 BluePrint \u91cc\uff0c\u6b8b\u5dee\u8fde\u63a5\u4f7f\u7528 <code>add2(lhs, rhs)<\/code> \u8868\u8fbe\uff0c\u5176\u4e2d <code>lhs<\/code> \u901a\u5e38\u662f shortcut \u5206\u652f\uff0c<code>rhs<\/code> \u662f\u4e3b\u5206\u652f\u3002\u4e3a\u4e86\u4fdd\u8bc1 shortcut \u5206\u652f\u53ef\u4ee5\u88ab\u7f16\u8bd1\u5668\u8bc6\u522b\uff0c<code>identity()<\/code> \u88ab\u7528\u4f5c\u5360\u4f4d\u7b26\uff0c\u8868\u793a&#8221;\u76f4\u63a5\u4f20\u9012\u8f93\u5165&#8221;\u3002\u4f8b\u5982\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=\"\">Layer residual = add2(\n    identity(),                                    \/\/ shortcut\n    seq(conv(64,3,1,1), bn(), relu(),\n        conv(64,3,1,1), bn())                      \/\/ main branch\n);<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u5c55\u5f00\u5230 ArchPlan \u65f6\uff0c<code>add2<\/code> \u4f1a\u88ab\u7ffb\u8bd1\u6210 <code>Add2Start<\/code>\u3001<code>Add2ShortcutEnd<\/code> \u548c <code>Add2End<\/code> \u4e09\u4e2a\u6807\u8bb0\u5c42\uff0c\u7528\u6765\u5728\u540e\u7eed\u6784\u5efa\u8ba1\u7b97\u56fe\u65f6\u6b63\u786e\u63d2\u5165\u5f20\u91cf\u52a0\u6cd5\u3002\u914d\u5408 <code>block()<\/code> \u4f7f\u7528\u65f6\uff0c\u8fd9\u4e9b\u6807\u8bb0\u7531\u6846\u67b6\u81ea\u52a8\u751f\u6210\uff0c\u7528\u6237\u65e0\u9700\u5173\u5fc3\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e9b\u5de5\u5382\u51fd\u6570\u90fd\u5b9a\u4e49\u5728\u5934\u6587\u4ef6\u91cc\uff0c\u5168\u90e8\u662f <code>inline<\/code>\uff0c\u56e0\u6b64 BluePrint DSL \u662f<strong>header-only<\/strong>\u7684\u3002\u7528\u6237\u53ea\u9700\u8981 <code>#include \"renaissance\/graph\/blueprint.h\"<\/code>\uff0c\u4e0d\u9700\u8981\u94fe\u63a5\u989d\u5916\u7684\u5e93\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><code>Layer::Node<\/code> \u5185\u90e8\u7528 <code>std::variant<\/code> \u627f\u8f7d\u6240\u6709\u53ef\u80fd\u7684\u53c2\u6570\u7c7b\u578b\u2014\u2014<code>ConvParam<\/code>\u3001<code>BNParam<\/code>\u3001<code>SequentialParam<\/code>\u3001<code>BlockParam<\/code> \u7b49\u7b49\u3002\u76f8\u6bd4\u4f20\u7edf\u7684\u7ee7\u627f\u4f53\u7cfb\uff0cvariant \u6709\u4e24\u4e2a\u597d\u5904\uff1a\u4e00\u662f\u7c7b\u578b\u5b89\u5168\uff0c\u7f16\u8bd1\u5668\u4f1a\u68c0\u67e5\u4f60\u8bbf\u95ee\u7684\u662f\u5426\u662f\u6b63\u786e\u7684\u53c2\u6570\uff1b\u4e8c\u662f\u6ca1\u6709\u865a\u51fd\u6570\u6307\u9488\u5f00\u9500\uff0c\u6bcf\u4e2a <code>Layer<\/code> \u90fd\u53ea\u662f\u4e00\u4e2a <code>shared_ptr<\/code> \u52a0\u4e00\u4e2a\u6307\u5411\u4e0d\u53ef\u53d8\u8282\u70b9\u7684\u5f15\u7528\uff0c\u5185\u5b58\u5e03\u5c40\u7d27\u51d1\u3002\u6bcf\u4e2a\u5de5\u5382\u51fd\u6570\u672c\u8d28\u4e0a\u5c31\u662f\u6784\u9020\u4e00\u4e2a <code>Layer::Node<\/code> \u5e76\u5305\u88c5\u6210 <code>Layer<\/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=\"\">inline Layer conv(int out_ch, int k, int s = 1, int p = 0) {\n    return Layer(std::make_shared&lt;Layer::Node>(\n        detail::NodeKind::Conv2d,\n        detail::ConvParam{out_ch, k, s, p}));\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u7ec4\u5408\u51fd\u6570\u5219\u5229\u7528 C++ \u53d8\u53c2\u6a21\u677f\u628a\u4efb\u610f\u591a\u4e2a <code>Layer<\/code> \u6536\u96c6\u6210 <code>std::vector&lt;Layer&gt;<\/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=\"\">template &lt;typename... Args>\ninline Layer seq(Args&amp;&amp;... args) {\n    auto layers = detail::make_layers(std::forward&lt;Args>(args)...);\n    if (layers.empty()) throw std::invalid_argument(\"seq: empty sequential is not allowed\");\n    return Layer(std::make_shared&lt;Layer::Node>(\n        detail::NodeKind::Sequential,\n        detail::SequentialParam{std::move(layers)}));\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u610f\u5473\u7740 <code>seq<\/code> \u5728\u7f16\u8bd1\u671f\u5c31\u80fd\u68c0\u67e5\u7c7b\u578b\uff1a\u4f60\u628a\u4e00\u4e2a\u4e0d\u662f <code>Layer<\/code> \u7684\u4e1c\u897f\u585e\u8fdb\u53bb\uff0c\u4f1a\u76f4\u63a5\u7f16\u8bd1\u62a5\u9519\u3002\u76f8\u6bd4 Python \u6846\u67b6\u7684\u8fd0\u884c\u65f6\u68c0\u67e5\uff0c\u8fd9\u79cd\u9759\u6001\u7c7b\u578b\u7ea6\u675f\u53ef\u4ee5\u628a\u5f88\u591a\u4f4e\u7ea7\u9519\u8bef\u627c\u6740\u5728\u7f16\u8bd1\u9636\u6bb5\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u56db\u3001BluePrint \u7684\u6838\u5fc3\u8bbe\u8ba1\u539f\u5219<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">BluePrint DSL \u7684\u8bbe\u8ba1\u53ef\u4ee5\u6982\u62ec\u4e3a\u4e09\u6761\u539f\u5219\uff0c\u5b83\u4eec\u5171\u540c\u51b3\u5b9a\u4e86 DSL \u65e2\u597d\u5199\u53c8\u597d\u4f18\u5316\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>1. \u4e0d\u53ef\u53d8\u4e0e\u58f0\u660e\u5f0f<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6240\u6709 <code>Layer<\/code> \u5bf9\u8c61\u4e00\u65e6\u521b\u5efa\u5c31\u4e0d\u53ef\u4fee\u6539\uff0c\u5185\u90e8\u901a\u8fc7 <code>std::shared_ptr&lt;const Node&gt;<\/code> \u5171\u4eab\u4e0d\u53ef\u53d8\u6570\u636e\u3002\u7528\u6237\u63cf\u8ff0\u7684\u662f&#8221;\u7f51\u7edc\u662f\u4ec0\u4e48&#8221;\uff0c\u800c\u4e0d\u662f&#8221;\u7f51\u7edc\u600e\u4e48\u7b97&#8221;\u3002\u6846\u67b6\u6839\u636e\u8fd9\u4efd\u58f0\u660e\u6765\u51b3\u5b9a\u6267\u884c\u987a\u5e8f\u3001\u5185\u5b58\u5e03\u5c40\u3001\u7b97\u5b50\u878d\u5408\u7b56\u7565\u3002\u8fd9\u4e3a\u540e\u7eed\u7684\u9759\u6001\u56fe\u7f16\u8bd1\u3001CUDA Graph \u5168\u6355\u83b7\u3001MemoryPlan \u9759\u6001\u663e\u5b58\u89c4\u5212\u6253\u4e0b\u4e86\u57fa\u7840\u2014\u2014\u6240\u6709\u4f18\u5316\u90fd\u5efa\u7acb\u5728&#8221;\u7f16\u8bd1\u671f\u77e5\u9053\u4e00\u5207&#8221;\u7684\u524d\u63d0\u4e0a\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>2. \u4e3a\u878d\u5408\u800c\u751f<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">BluePrint \u4e0d\u4ec5\u63d0\u4f9b\u57fa\u7840\u5c42\u5de5\u5382\u51fd\u6570\uff0c\u8fd8\u5185\u7f6e\u4e86\u9884\u878d\u5408\u6a21\u5757\uff08\u5982 <code>cbr<\/code>\u3001<code>gap_fc<\/code>\uff09\u548c\u9ad8\u5c42\u5757\uff08\u5982 <code>block<\/code>\u3001<code>mbconv<\/code>\uff09\u3002\u8fd9\u4e9b\u4e0d\u662f\u7b80\u5355\u7684\u8bed\u6cd5\u7cd6\u2014\u2014\u5b83\u4eec\u5411\u7f16\u8bd1\u5668\u4f20\u9012\u4e86\u660e\u786e\u7684\u878d\u5408\u610f\u56fe\u3002\u5f53\u7528\u6237\u5199 <code>cbr(64, 3, 1, 1)<\/code> \u800c\u4e0d\u662f <code>seq(conv(64, 3, 1, 1), bn(), relu())<\/code> \u65f6\uff0c\u7f16\u8bd1\u5668\u4e0d\u9700\u8981\u505a\u6a21\u5f0f\u5339\u914d\u5c31\u80fd\u786e\u5b9a\u8fd9\u662f\u4e00\u4e2a CBR \u878d\u5408\u8282\u70b9\uff0c\u53ef\u4ee5\u76f4\u63a5\u6620\u5c04\u5230\u540e\u7aef\u7684\u9ad8\u6548\u878d\u5408\u7b97\u5b50\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>3. \u8d34\u8fd1\u786c\u4ef6\u7ec6\u8282<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">BluePrint \u5de5\u5382\u51fd\u6570\u7684\u53c2\u6570\u8bbe\u8ba1\u523b\u610f\u5730\u4fdd\u6301\u514b\u5236\u3002\u4f8b\u5982 <code>conv<\/code> \u4e0d\u63a5\u53d7 bias \u53c2\u6570\uff0c\u56e0\u4e3a Tech-Renaissance \u7684\u5377\u79ef\u5c42\u7edf\u4e00\u4e0d\u4f7f\u7528 bias\uff1a\u5377\u79ef\u540e\u7684 BN \u5c42\u672c\u8eab\u5c31\u6709 <code>beta<\/code> \u53c2\u6570\uff08shift\uff09\uff0cconv \u7684 bias \u4e0e BN \u7684 beta \u5728\u6570\u5b66\u4e0a\u662f\u5197\u4f59\u7684\u3002\u53bb\u6389 conv bias \u65e2\u51cf\u5c11\u4e86\u53c2\u6570\u91cf\uff0c\u4e5f\u7b80\u5316\u4e86 Conv+BN \u878d\u5408\u65f6\u7684\u8ba1\u7b97\u903b\u8f91\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u518d\u6bd4\u5982 <code>channel_padding()<\/code>\uff0c\u5b83\u4f1a\u628a\u5f53\u524d\u901a\u9053\u6570\u5411\u4e0a\u5bf9\u9f50\u5230 8 \u7684\u500d\u6570\u3002\u5728 FP16 AMP \u6a21\u5f0f\u4e0b\uff0cTensor Core \u5bf9\u901a\u9053\u7ef4\u5ea6\u6709\u5bf9\u9f50\u8981\u6c42\uff1b\u867d\u7136\u6846\u67b6\u4f1a\u81ea\u52a8\u5904\u7406\u9996\u5c42 Conv\/CBR \u7684\u8f93\u5165 C \u5bf9\u9f50\u5230 4\uff0c\u4f46\u5728\u7f51\u7edc\u4e2d\u95f4\u67d0\u4e9b\u7279\u6b8a\u4f4d\u7f6e\uff0c\u7528\u6237\u4ecd\u53ef\u4ee5\u663e\u5f0f\u63d2\u5165 <code>channel_padding()<\/code> \u6765\u786e\u4fdd\u901a\u9053\u6570\u6ee1\u8db3\u540e\u7aef kernel \u7684\u5bf9\u9f50\u7ea6\u675f\u3002\u4e0b\u9762\u662f\u4e00\u6bb5\u7b80\u5316\u7684\u5f62\u72b6\u63a8\u5bfc\u4f2a\u4ee3\u7801\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=\"\">\/\/ \u8f93\u5165\u7279\u5f81\u56fe\u5f62\u72b6\u4e3a [N, H, W, C]\n\/\/ \u5bf9\u4e8e 2D \u5377\u79ef\uff0c\u8f93\u51fa H\u3001W\u3001C \u6309\u4e0b\u5f0f\u8ba1\u7b97\uff1a\nint out_h = (H + 2 * pad - k) \/ stride + 1;\nint out_w = (W + 2 * pad - k) \/ stride + 1;\nint out_c = out_ch;                       \/\/ \u7531 conv\/cbr \u7684\u53c2\u6570\u6307\u5b9a\n\n\/\/ channel_padding \u628a C \u5bf9\u9f50\u5230 8\uff1a\nint aligned_c = (C + 7) \/ 8 * 8;<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u5176\u4e2d <code>H<\/code>\u3001<code>W<\/code>\u3001<code>C<\/code> \u5206\u522b\u662f\u8f93\u5165\u7279\u5f81\u56fe\u7684\u9ad8\u3001\u5bbd\u3001\u901a\u9053\u6570\uff0c<code>k<\/code>\u3001<code>stride<\/code>\u3001<code>pad<\/code> \u662f\u5377\u79ef\u6838\u5927\u5c0f\u3001\u6b65\u957f\u548c\u586b\u5145\uff0c<code>out_ch<\/code> \u662f\u8f93\u51fa\u901a\u9053\u6570\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e94\u3001\u9884\u878d\u5408\u5757\uff1a\u628a\u4f18\u5316\u610f\u56fe\u5199\u8fdb DSL<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">BluePrint \u91cc\u6709\u4e00\u6279\u4e13\u95e8\u7528\u6765\u8868\u8fbe&#8221;<strong>\u8fd9\u91cc\u5e94\u8be5\u88ab\u878d\u5408<\/strong>&#8220;\u7684\u5de5\u5382\u51fd\u6570\u3002\u6700\u5e38\u7528\u7684\u662f <code>cbr<\/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=\"\">inline Layer cbr(int out_ch, int k, int s, int p,\n                 double momentum = 0.1, double eps = 1e-5) {\n    return Layer(std::make_shared&lt;Layer::Node>(\n        detail::NodeKind::CBR,\n        detail::CBRParam{out_ch, k, s, p,\n                         static_cast&lt;float>(momentum),\n                         static_cast&lt;float>(eps)}));\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><code>cbr<\/code> \u4ee3\u8868 Conv + BatchNorm + ReLU \u4e09\u5143\u878d\u5408\u3002\u7528\u6237\u5199\u4e00\u4e2a <code>cbr(64, 3, 1, 1)<\/code>\uff0c\u7f16\u8bd1\u5668\u5c31\u77e5\u9053\u8fd9\u662f\u4e00\u4e2a\u6574\u4f53\uff0c\u4e0d\u4f1a\u628a\u5b83\u62c6\u6210\u4e09\u4e2a\u72ec\u7acb\u5c42\u3002\u5982\u679c\u7528\u6237\u663e\u5f0f\u5199\u6210 <code>seq(conv(...), bn(), relu())<\/code>\uff0cArchPlan \u7684 <code>step9_merge_triple<\/code> \u4e5f\u4f1a\u5c1d\u8bd5\u628a\u5b83\u4eec\u5408\u5e76\u4e3a <code>CBR<\/code> \u878d\u5408\u5c42\uff0c\u4f46\u76f4\u63a5\u5199 <code>cbr<\/code> \u7b49\u4e8e\u628a\u4f18\u5316\u610f\u56fe\u63d0\u524d\u8868\u8fbe\u6e05\u695a\uff0c\u51cf\u5c11\u7f16\u8bd1\u5668\u731c\u6d4b\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7c7b\u4f3c\u5730\uff0c\u8fd8\u6709\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><code>cbrp<\/code>\uff1aConv + BN + ReLU + MaxPool \u56db\u5143\u7ec4\u5408\u3002\u5b83\u76ee\u524d\u5728 <code>blueprint.h<\/code> \u4e2d\u4ecd\u4f5c\u4e3a\u8bed\u6cd5\u7cd6\u5b58\u5728\uff0c\u4f46 <code>ArchPlan::step8_merge_quadruple<\/code> \u5df2\u4e3a\u7a7a\u5b9e\u73b0\u2014\u2014\u4e5f\u5c31\u662f\u8bf4\uff0c\u5f53\u524d\u7248\u672c\u91cc <code>cbrp<\/code> \u4f1a\u88ab\u5c55\u5f00\u4e3a\u56db\u4e2a\u72ec\u7acb\u5c42\uff0c\u800c\u4e0d\u4f1a\u88ab\u878d\u5408\u6210\u5355\u4e2a kernel\u3002\u4fdd\u7559\u8fd9\u4e2a\u5de5\u5382\u51fd\u6570\u662f\u4e3a\u4e86\u8ba9\u7528\u6237\u63cf\u8ff0 VGG \u98ce\u683c\u7684&#8221;\u5377\u79ef\u5757+\u6c60\u5316&#8221;\u7ed3\u6784\u65f6\u66f4\u7b80\u6d01\u3002<\/li>\n\n\n\n<li><code>gap_fc<\/code>\uff1aGlobal Average Pooling + FC \u4e8c\u5143\u878d\u5408\u3002\u5728 ResNet \u7684 classifier \u5934\u90e8\uff0cGAP \u5c06 7&#215;7 \u7684\u7279\u5f81\u56fe\u6c60\u5316\u4e3a 1&#215;1\uff0c\u7136\u540e\u63a5\u4e00\u4e2a FC\uff1b\u8fd9\u4e24\u4e2a\u64cd\u4f5c\u53ef\u4ee5\u878d\u5408\u6210\u4e00\u6b21\u5185\u6838\u8c03\u7528\u3002<\/li>\n\n\n\n<li><code>block(...)<\/code>\uff1aResNet Bottleneck \/ BasicBlock \u878d\u5408\u5757\u3002<\/li>\n\n\n\n<li><code>mbconv(...)<\/code>\uff1aMobileNetV2 \u5012\u7f6e\u6b8b\u5dee\u5757\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u4ee5 Bottleneck \u4e3a\u4f8b\uff0c<code>block(64, 256, BlockStyle::RESNET_1_3_1)<\/code> \u751f\u6210\u7684\u662f\u4e00\u4e2a <code>NodeKind::Block<\/code> \u8282\u70b9\u3002\u5728 <code>ArchPlan::expand_tree()<\/code> \u4e2d\uff0c\u5982\u679c\u5f53\u524d\u5f00\u542f\u4e86\u878d\u5408\u6a21\u5f0f\uff0c\u5b83\u4f1a\u8c03\u7528 <code>expand_block_impl()<\/code>\uff0c\u628a\u8282\u70b9\u76f4\u63a5\u6620\u5c04\u4e3a <code>BottleneckProjection<\/code> \u6216 <code>BottleneckIdentity<\/code> \u8fd9\u79cd\u878d\u5408\u5c42\uff1b\u5982\u679c\u5173\u95ed\u878d\u5408\uff0c\u5219\u8c03\u7528 <code>expand_block_unfused()<\/code>\uff0c\u628a\u5b83\u62c6\u6210 <code>conv<\/code>\u3001<code>bn<\/code>\u3001<code>relu<\/code>\u3001<code>add2<\/code> \u7b49\u539f\u59cb\u7b97\u5b50\u5e8f\u5217\u3002\u540c\u4e00\u4efd BluePrint\uff0c\u65e2\u80fd\u8dd1\u9ad8\u6027\u80fd\u878d\u5408\u8def\u5f84\uff0c\u4e5f\u80fd\u8dd1\u7eaf\u539f\u8bed\u8def\u5f84\uff0c\u4fbf\u4e8e\u8c03\u8bd5\u6216 CPU \u6a21\u5f0f\u8fd0\u884c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e9b\u9884\u878d\u5408\u5757\u7684\u8bbe\u8ba1\u7406\u5ff5\u662f\uff1a<strong>DSL \u4e0d\u5e94\u8be5\u53ea\u63cf\u8ff0&#8221;\u6709\u4ec0\u4e48&#8221;\uff0c\u8fd8\u5e94\u8be5\u6697\u793a&#8221;\u600e\u4e48\u6267\u884c\u6700\u597d&#8221;<\/strong>\u3002\u7528\u6237\u5728\u9ad8\u5c42\u5199\u4e0b\u7684\u6bcf\u4e00\u4e2a <code>block()<\/code>\u3001<code>cbr()<\/code>\uff0c\u90fd\u4f1a\u539f\u5c01\u4e0d\u52a8\u5730\u8fdb\u5165 ArchPlan\uff0c\u6210\u4e3a\u540e\u7aef\u9009\u62e9\u4e13\u7528\u878d\u5408 kernel \u7684\u4f9d\u636e\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u516d\u3001\u4ece BluePrint \u5230\u53ef\u6267\u884c\u56fe<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">BluePrint \u672c\u8eab\u53ea\u662f\u6a21\u578b\u5b9a\u4e49\uff0c\u4e0d\u6301\u6709\u4efb\u4f55\u6743\u91cd\u3001\u4e0d\u5206\u914d\u4efb\u4f55\u663e\u5b58\u3002\u771f\u6b63\u628a\u5b83\u53d8\u6210\u53ef\u6267\u884c\u56fe\u7684\u662f Graph \u6a21\u5757\u7684\u7f16\u8bd1\u7ba1\u7ebf\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7b2c\u4e00\u6b65\u662f <code>ArchPlan::from_blueprint()<\/code>\u3002\u5b83\u628a <code>Layer<\/code> \u6811\u5c55\u5f00\u6210\u4e00\u4e32 <code>ArchLayer<\/code>\uff0c\u6bcf\u4e2a <code>ArchLayer<\/code> \u5305\u542b\u5f52\u4e00\u5316\u540e\u7684\u5c42\u7c7b\u578b\u3001\u53c2\u6570\u3001\u8f93\u5165\u8f93\u51fa\u5f62\u72b6\u5360\u4f4d\uff0c\u4ee5\u53ca\u662f\u5426\u9996\u5c42\u7b49\u6807\u8bb0\u3002\u5c55\u5f00\u8fc7\u7a0b\u5728 <code>src\/graph\/arch_plan_expand.cpp<\/code> \u4e2d\u9012\u5f52\u5b8c\u6210\uff0c\u6838\u5fc3\u903b\u8f91\u53ef\u4ee5\u7528\u4e0b\u9762\u8fd9\u6bb5\u7b80\u5316\u4ee3\u7801\u6982\u62ec\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=\"\">void ArchPlan::expand_tree(const Layer&amp; root,\n                           std::vector&lt;ArchLayer>&amp; out,\n                           int&amp; current_c, bool fuse) {\n    auto&amp; node = *root.node_;\n    int src_id = static_cast&lt;int>(out.size());   \/\/ \u5f53\u524d\u5c42\u5728\u5e8f\u5217\u4e2d\u7684\u4f4d\u7f6e\n    switch (node.kind) {\n        case NodeKind::Conv2d:\n        case NodeKind::ReLU:\n            \/\/ \u57fa\u7840\u5c42\u76f4\u63a5\u6620\u5c04\u5230\u5bf9\u5e94\u7684 ArchLayer\n            expand_primitive_impl(node, out, current_c, src_id, fuse);\n            break;\n        case NodeKind::Sequential:\n            for (const auto&amp; child : node.children)\n                expand_tree(child, out, current_c, fuse);\n            break;\n        case NodeKind::Block:\n            if (fuse) expand_block_impl(node, out, current_c, current_c, src_id);\n            else      expand_block_unfused(node, out, current_c, current_c, src_id);\n            break;\n        case NodeKind::Add2:\n            \/\/ \u5c55\u5f00\u6b8b\u5dee\u8fde\u63a5\uff0c\u63d2\u5165 Add2Start \/ Add2ShortcutEnd \/ Add2End \u6807\u8bb0\n            break;\n        case NodeKind::Repeat:\n            for (int t = 0; t &lt; node.times; ++t)\n                expand_tree(node.body, out, current_c, fuse);\n            break;\n    }\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6ce8\u610f <code>fuse<\/code> \u53c2\u6570\u3002\u5b83\u901a\u5e38\u4e0e <code>GlobalRegistry<\/code> \u91cc\u7684 AMP \u5f00\u5173\u8054\u52a8\uff1a<code>from_blueprint()<\/code> \u9ed8\u8ba4\u4f20\u5165 <code>fuse=true<\/code>\uff0c\u4f46\u5982\u679c\u6b64\u65f6 AMP \u672a\u5f00\u542f\uff0c\u6846\u67b6\u4f1a\u76f4\u63a5\u629b\u9519\uff0c\u56e0\u4e3a\u5f53\u524d\u878d\u5408 kernel \u53ea\u5728 AMP \u6a21\u5f0f\u4e0b\u5b9e\u73b0\u3002\u5f00\u542f AMP \u65f6\uff0c\u9ed8\u8ba4\u505a\u878d\u5408\uff1b\u5173\u95ed AMP\uff0c\u6216\u663e\u5f0f\u4f20\u5165 <code>fuse=false<\/code> \u65f6\uff0c<code>block<\/code>\u3001<code>cbr<\/code> \u7b49\u8282\u70b9\u4f1a\u5c55\u5f00\u6210\u539f\u59cb\u7b97\u5b50\u5e8f\u5217\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5c55\u5f00\u4e4b\u540e\uff0cArchPlan \u8fd8\u4f1a\u505a\u4e00\u7cfb\u5217\u5f52\u4e00\u5316\uff1a\u91cd\u547d\u540d BN\u3001\u5904\u7406 SoftmaxCE\u3001\u628a identity\/flatten \u6807\u51c6\u5316\u3001\u63a8\u5bfc\u5f62\u72b6\u3001\u5408\u5e76 Bottleneck\/BasicBlock\/InvResidual\u3001\u5408\u5e76\u4e8c\u5143\/\u4e09\u5143\u6a21\u5f0f\u3001\u6807\u8bb0\u9996\u5c42\u3002\u6700\u7ec8\u5f97\u5230\u4e00\u4efd\u6807\u51c6\u5316\u7684\u67b6\u6784\u63cf\u8ff0\uff0c\u4ea4\u7ed9 Compiler \u8fdb\u5165\u4e94\u9636\u6bb5\u7f16\u8bd1\u7ba1\u7ebf\uff1a\u5f62\u72b6\u63a8\u5bfc \u2192 \u6700\u5927\u69fd\u4f4d\u8ba1\u7b97 \u2192 MemoryPlan \u5206\u914d \u2192 ComputationGraph \u6784\u5efa \u2192 \u53d8\u4f53\u5171\u4eab\u53bb\u91cd\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e9b\u6b65\u9aa4\u6211\u4eec\u4f1a\u5728\u4e0b\u4e00\u7bc7\u300aArchPlan \u4e0e\u7f16\u8bd1\u7ba1\u7ebf\u300b\u91cc\u8be6\u7ec6\u8bb2\u3002\u8fd9\u91cc\u53ea\u9700\u8981\u8bb0\u4f4f\u4e00\u70b9\uff1a<strong>BluePrint \u662f\u7528\u6237\u4e0e\u7f16\u8bd1\u5668\u4e4b\u95f4\u7684\u5951\u7ea6<\/strong>\u3002\u5b83\u8db3\u591f\u7b80\u6d01\uff0c\u8ba9\u4eba\u5199\u5f97\u8212\u670d\uff1b\u53c8\u8db3\u591f\u7ed3\u6784\u5316\uff0c\u8ba9\u7f16\u8bd1\u5668\u80fd\u8bfb\u61c2\u5e76\u69a8\u53d6\u4f18\u5316\u7a7a\u95f4\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e03\u3001\u4e0e PyTorch \u7684\u5bf9\u6bd4<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5f88\u591a\u4eba\u7b2c\u4e00\u6b21\u770b Tech-Renaissance \u7684 BluePrint \u4f1a\u95ee\uff1a\u8fd9\u548c PyTorch \u7684 <code>nn.Sequential<\/code> \u6709\u4ec0\u4e48\u533a\u522b\uff1f<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8868\u9762\u7684\u76f8\u4f3c\u662f\u90fd\u7528\u4e86\u987a\u5e8f\u7ec4\u5408\uff0c\u4f46\u5e95\u5c42\u601d\u8def\u622a\u7136\u4e0d\u540c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">PyTorch \u7684 <code>nn.Sequential<\/code> \u672c\u8d28\u4e0a\u662f\u4e00\u7ec4<strong>\u6709\u72b6\u6001\u7684 Module \u5bf9\u8c61<\/strong>\u3002\u6bcf\u4e2a <code>nn.Conv2d<\/code> \u5185\u90e8\u5df2\u7ecf\u6301\u6709\u6743\u91cd\u5f20\u91cf\u3001\u504f\u7f6e\u5f20\u91cf\u3001\u68af\u5ea6\u5f20\u91cf\uff1b\u6a21\u578b\u5b9a\u4e49\u548c\u53c2\u6570\u5206\u914d\u662f\u540c\u65f6\u53d1\u751f\u7684\u3002\u4e5f\u6b63\u56e0\u5982\u6b64\uff0cPyTorch \u770b\u5230\u7684\u662f\u4e00\u5806\u5bf9\u8c61\uff0c\u800c\u4e0d\u662f\u4e00\u5f20\u7eaf\u7cb9\u7684\u56fe\u3002<code>torch.compile<\/code> \u8981\u82b1\u5f88\u5927\u529b\u6c14\u53bb\u6355\u83b7\u3001\u53bb\u91cd\u3001\u5206\u6790\uff0c\u624d\u80fd\u4ece\u8fd9\u4e9b\u5bf9\u8c61\u7684\u6267\u884c\u8f68\u8ff9\u4e2d\u6062\u590d\u51fa\u53ef\u4f18\u5316\u7684\u9759\u6001\u56fe\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Tech-Renaissance \u7684 <code>Layer<\/code> \u548c <code>BluePrint<\/code> \u662f<strong>\u65e0\u72b6\u6001\u7684\u62bd\u8c61\u8bed\u6cd5\u6811\uff08AST\uff09<\/strong>\u3002\u5b83\u53ea\u63cf\u8ff0\u7f51\u7edc\u7ed3\u6784\uff0c\u4e0d\u6301\u6709\u4efb\u4f55\u53c2\u6570\u3002\u4f60\u6ca1\u6cd5\u50cf PyTorch \u90a3\u6837\u5728\u5b9a\u4e49\u540e\u7acb\u523b <code>print(model.conv1.weight)<\/code>\uff0c\u56e0\u4e3a\u90a3\u65f6\u5019\u6743\u91cd\u8fd8\u6ca1\u6709\u5206\u914d\u3002\u6743\u91cd\u3001\u68af\u5ea6\u3001\u4e2d\u95f4\u7279\u5f81\u56fe\u7684\u5f20\u91cf\u63cf\u8ff0\u8981\u5230 <code>Compiler<\/code> \u9636\u6bb5\u624d\u7531 <code>MemoryPlan<\/code> \u6839\u636e\u7ed3\u6784\u7edf\u4e00\u5206\u914d\u3002\u8fd9\u610f\u5473\u7740\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u7f16\u8bd1\u5668\u62ff\u5230\u7684\u662f\u5b8c\u6574\u84dd\u56fe\uff0c\u4e0d\u662f\u6267\u884c\u8f68\u8ff9<\/strong>\u3002\u4e0d\u9700\u8981\u56fe\u6355\u83b7\uff0c\u4e5f\u4e0d\u9700\u8981\u4ece Python \u52a8\u6001\u884c\u4e3a\u4e2d\u53cd\u63a8\u7ed3\u6784\u3002<\/li>\n\n\n\n<li><strong>\u9884\u878d\u5408\u5757\u5929\u7136\u4fdd\u7559<\/strong>\u3002<code>cbr<\/code>\u3001<code>block<\/code> \u8fd9\u4e9b\u8282\u70b9\u76f4\u63a5\u8fdb\u5165 ArchPlan\uff0c\u6210\u4e3a\u540e\u7aef\u878d\u5408\u7b97\u5b50\u7684\u5019\u9009\uff0c\u800c\u4e0d\u662f\u4e8b\u540e\u7531 pattern matcher \u53bb\u8bc6\u522b\u3002<\/li>\n\n\n\n<li><strong>\u591a shape \u53d8\u4f53\u53ef\u4ee5\u5171\u4eab\u540c\u4e00\u4efd\u56fe<\/strong>\u3002\u56e0\u4e3a BluePrint \u4e0d\u7ed1\u5b9a\u5177\u4f53 batch size \u548c\u5206\u8fa8\u7387\uff0c\u540c\u4e00\u5f20 <code>ComputationGraph<\/code> \u53ef\u4ee5\u670d\u52a1\u8bad\u7ec3\u9ad8\u5206\u8fa8\u7387\u3001\u8bad\u7ec3\u4f4e\u5206\u8fa8\u7387\u3001\u9a8c\u8bc1\u7b49\u591a\u4e2a\u53d8\u4f53\uff0c\u53ea\u6709 <code>MemoryPlan<\/code> \u4e0d\u540c\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u4ece\u66f4\u5b8f\u89c2\u7684\u89c6\u89d2\u770b\uff0cBluePrint \u4e0e JAX\/Flax\u3001Keras Functional API\u3001ONNX \u540c\u5c5e\u4e8e&#8221;\u58f0\u660e\u5f0f&#8221;\u9635\u8425\uff0c\u4f46\u5b83\u662f\u7eaf C++ \u7684\uff0c\u6ca1\u6709 Python \u8fd0\u884c\u65f6\u548c GIL \u7684\u675f\u7f1a\u3002\u6a21\u578b\u5b9a\u4e49\u5b8c\u6210\u540e\uff0c\u7f16\u8bd1\u7ba1\u7ebf\u53ef\u4ee5\u76f4\u63a5\u5728 C++ \u4fa7\u5b8c\u6210\u5c55\u5f00\u3001\u5f62\u72b6\u63a8\u5bfc\u3001\u5185\u5b58\u89c4\u5212\u548c\u56fe\u6784\u5efa\uff0c\u4e0d\u9700\u8981\u8de8\u8bed\u8a00\u5e8f\u5217\u5316\uff0c\u4e5f\u4e0d\u9700\u8981\u7b49\u5f85 Python \u89e3\u91ca\u5668\u9010\u884c\u6267\u884c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">BluePrint \u4e5f\u5177\u6709\u8f83\u597d\u7684\u53ef\u6269\u5c55\u6027\u3002\u5982\u679c\u9700\u8981\u652f\u6301\u4e00\u79cd\u65b0\u7684\u5c42\u6216\u65b0\u7684\u7ec4\u5408\u6a21\u5f0f\uff0c\u901a\u5e38\u53ea\u9700\u8981\u505a\u4e09\u4ef6\u4e8b\uff1a\u5728 <code>NodeKind<\/code> \u679a\u4e3e\u4e2d\u589e\u52a0\u4e00\u4e2a\u6761\u76ee\uff0c\u5728 <code>Payload<\/code> variant \u4e2d\u589e\u52a0\u4e00\u4e2a\u53c2\u6570\u7ed3\u6784\u4f53\uff0c\u5728 <code>ArchPlan<\/code> \u7684 <code>expand_tree<\/code> \u91cc\u589e\u52a0\u5bf9\u5e94\u7684\u5c55\u5f00\u903b\u8f91\u3002\u540e\u7aef\u662f\u5426\u652f\u6301\u8be5\u5c42\u5219\u901a\u8fc7 <code>LayerDescriptor<\/code> \u6ce8\u518c\u8868\u6765\u7ba1\u7406\uff0cDSL \u672c\u8eab\u4e0d\u9700\u8981\u5173\u5fc3 CUDA kernel \u7684\u5177\u4f53\u5b9e\u73b0\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5f53\u7136\uff0c\u8fd9\u79cd\u8bbe\u8ba1\u4e5f\u4ed8\u51fa\u4e86\u7075\u6d3b\u6027\u4e0a\u7684\u4ee3\u4ef7\u3002\u5982\u679c\u4f60\u7684\u6a21\u578b\u9700\u8981\u5728\u8fd0\u884c\u65f6\u6839\u636e\u8f93\u5165\u52a8\u6001\u6539\u53d8\u7ed3\u6784\uff08\u6bd4\u5982\u8f93\u5165\u76f8\u5173\u63a7\u5236\u6d41\u3001\u52a8\u6001\u56fe\u7ed3\u6784\uff09\uff0cBluePrint \u5e76\u4e0d\u64c5\u957f\u3002Tech-Renaissance \u7684\u76ee\u6807\u573a\u666f\u662f\u4e3b\u6d41 CV \u8bad\u7ec3\u4efb\u52a1\u2014\u2014ResNet\u3001VGG\u3001MobileNet \u8fd9\u7c7b\u7ed3\u6784\u56fa\u5b9a\u3001\u53ef\u9ad8\u5ea6\u4f18\u5316\u7684\u7f51\u7edc\uff0c\u5728\u8fd9\u4e9b\u573a\u666f\u4e0b\uff0c\u58f0\u660e\u5f0f DSL \u7684\u6536\u76ca\u8fdc\u5927\u4e8e\u4ee3\u4ef7\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u516b\u3001\u5c0f\u7ed3<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">BluePrint DSL \u662f Tech-Renaissance \u9759\u6001\u56fe\u7f16\u8bd1\u8def\u7ebf\u7684\u8d77\u70b9\u3002\u5b83\u7528\u4e00\u7ec4\u8f7b\u91cf\u7684\u5de5\u5382\u51fd\u6570\u2014\u2014<code>seq<\/code>\u3001<code>add2<\/code>\u3001<code>repeat<\/code>\u3001<code>conv<\/code>\u3001<code>bn<\/code>\u3001<code>relu<\/code>\u3001<code>cbr<\/code>\u3001<code>block<\/code>\u3001<code>mbconv<\/code> \u7b49\u2014\u2014\u628a\u590d\u6742\u7684\u795e\u7ecf\u7f51\u7edc\u7ed3\u6784\u8868\u8fbe\u6210\u4e00\u68f5\u4e0d\u53ef\u53d8\u7684 <code>Layer<\/code> \u6811\u3002\u7528\u6237\u5199\u7684\u662f\u7ed3\u6784\uff0c\u7f16\u8bd1\u5668\u770b\u5230\u7684\u662f\u4f18\u5316\u673a\u4f1a\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u68f5\u6811\u7684\u8bbe\u8ba1\u8981\u70b9\u503c\u5f97\u8bb0\u4f4f\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u4e0d\u53ef\u53d8\u3001\u503c\u8bed\u4e49\u7684 Layer<\/strong>\uff0c\u8ba9\u6a21\u578b\u5b9a\u4e49\u6ca1\u6709\u9690\u5f0f\u72b6\u6001\uff0c\u53ef\u4ee5\u5b89\u5168\u5730\u62f7\u8d1d\u548c\u590d\u7528\uff1b<\/li>\n\n\n\n<li><strong>\u7ec4\u5408\u5f0f\u8bed\u6cd5<\/strong>\uff0c\u7528\u987a\u5e8f\u3001\u6b8b\u5dee\u3001\u91cd\u590d\u4e09\u79cd\u57fa\u672c\u7ed3\u6784\u8986\u76d6\u7edd\u5927\u591a\u6570\u7ecf\u5178\u7f51\u7edc\uff1b<\/li>\n\n\n\n<li><strong>\u9884\u878d\u5408\u9ad8\u5c42\u5757<\/strong>\uff0c\u628a Conv+BN+ReLU\u3001Bottleneck\u3001InvResidual \u7b49\u6027\u80fd\u5173\u952e\u6a21\u5f0f\u5728 DSL \u5c42\u5c31\u4fdd\u7559\u4e0b\u6765\uff1b<\/li>\n\n\n\n<li><strong>\u7eaf C++\u3001header-only<\/strong>\uff0c\u6ca1\u6709 Python \u8fd0\u884c\u65f6\u548c\u8de8\u8bed\u8a00\u5e8f\u5217\u5316\u7684\u5f00\u9500\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">BluePrint \u672c\u8eab\u4e0d\u505a\u4efb\u4f55\u8ba1\u7b97\uff0c\u5b83\u53ea\u662f\u628a&#8221;\u7528\u6237\u60f3\u8bad\u4ec0\u4e48\u6a21\u578b&#8221;\u7ffb\u8bd1\u6210&#8221;\u7f16\u8bd1\u5668\u80fd\u4f18\u5316\u4ec0\u4e48\u56fe&#8221;\u3002\u771f\u6b63\u628a\u8fd9\u4e9b\u5c42\u53d8\u6210\u5f20\u91cf\u3001\u7b97\u5b50\u3001CUDA Graph \u7684\uff0c\u662f\u4e0b\u4e00\u7bc7\u8981\u8bb2\u7684 ArchPlan \u4e0e\u7f16\u8bd1\u7ba1\u7ebf\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\u5341\u4e8c \u4e0a\u4e00\u7bc7\u6211\u4eec\u8bb2\u4e86 FusedN [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":696,"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-493","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\/493","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=493"}],"version-history":[{"count":3,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/posts\/493\/revisions"}],"predecessor-version":[{"id":726,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/posts\/493\/revisions\/726"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/media\/696"}],"wp:attachment":[{"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/media?parent=493"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/categories?post=493"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/tags?post=493"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}