{"id":405,"date":"2025-11-07T18:20:37","date_gmt":"2025-11-07T10:20:37","guid":{"rendered":"https:\/\/icalkzhangzihao.com\/?p=405"},"modified":"2026-04-12T13:28:01","modified_gmt":"2026-04-12T05:28:01","slug":"%e6%89%8b%e5%86%99%e6%95%b0%e5%ad%97%e8%af%86%e5%88%ab%e9%97%ae%e9%a2%98","status":"publish","type":"post","link":"https:\/\/icalkzhangzihao.com\/?p=405","title":{"rendered":"\u624b\u5199\u6570\u5b57\u8bc6\u522b\u95ee\u9898"},"content":{"rendered":"\n<h1 class=\"wp-block-heading\">0 \u56fe\u7247\u5206\u7c7b\u95ee\u9898<\/h1>\n\n\n\n<p>\u56fe\u7247\u5206\u7c7b\u95ee\u9898\u5c31\u662f\u8fa8\u8ba4\u8f93\u5165\u7684\u56fe\u7247\u7c7b\u522b\u7684\u95ee\u9898\uff0c\u4e14\u56fe\u7247\u7684\u7c7b\u522b\u5c5e\u4e8e\u4e8b\u5148\u7ed9\u5b9a\u7684\u4e00\u4e2a\u7c7b\u522b\u7ec4\u4e2d\u3002\u5c3d\u7ba1\u8fd9\u770b\u8d77\u6765\u5f88\u7b80\u5355\uff0c\u4f46\u8fd9\u662f\u8ba1\u7b97\u673a\u89c6\u89c9\u7684\u4e00\u4e2a\u6838\u5fc3\u95ee\u9898\uff0c\u4e14\u6709\u5f88\u5e7f\u6cdb\u7684\u5b9e\u9645\u5e94\u7528\u3002\u5e76\u4e14\uff0c\u6709\u5f88\u591a\u7684\u8ba1\u7b97\u673a\u89c6\u89c9\u7684\u95ee\u9898\u6700\u7ec8\u4f1a\u5316\u7b80\u4e3a\u56fe\u7247\u5206\u7c7b\u95ee\u9898\u3002<\/p>\n\n\n\n<p>\u4e3e\u4f8b\u6765\u8bf4\uff0c\u5047\u8bbe\u6709\u4e00\u4e2a\u56fe\u7247\u5206\u7c7b\u6a21\u578b\uff0c\u5b83\u5bf9\u4e8e\u8f93\u5165\u7684\u4e09\u901a\u9053\u7684\u56fe\u7247\u4f1a\u9884\u6d4b\u5176\u5c5e\u4e8e\u56db\u4e2a\u6807\u7b7e\uff08label\uff09\u7684\u6982\u7387\uff08\u56db\u4e2a\u6807\u7b7e\u4e3a cat\uff0c dog\uff0chat\uff0cmug\uff09\u3002\u4e0b\u56fe\u6240\u793a\u7684\u56fe\u7247\u662f\u4e00\u5f20248\u50cf\u7d20\u5bbd\u5ea6\uff0c400\u50cf\u7d20\u9ad8\u5ea6\u7684\u56fe\u7247\uff0c\u5e76\u4e14\u6709RGB\u4e09\u901a\u9053\uff0c\u90a3\u4e48\u8fd9\u5f20\u56fe\u7247\u53ef\u4ee5\u7528&nbsp;3*248*400&nbsp;\u4e2a\u6570\u5b57\u8868\u793a\uff0c\u6bcf\u4e2a\u6570\u5b57\u8303\u56f4\u4ece 0\u5230255\uff0c\u6a21\u578b\u7684\u4efb\u52a1\u5c31\u662f\u63a5\u53d7\u8fd9\u4e9b\u6570\u5b57\uff0c\u7136\u540e\u9884\u6d4b\u51fa\u8fd9\u4e9b\u6570\u5b57\u4ee3\u8868\u7684\u6807\u7b7e\uff08label\uff09\u3002<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/icalkzhangzihao.com\/wp-content\/uploads\/2025\/11\/image.png'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  loading=\"lazy\" decoding=\"async\" width=\"872\" height=\"619\" data-original=\"https:\/\/icalkzhangzihao.com\/wp-content\/uploads\/2025\/11\/image.png\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"\" class=\"wp-image-406\"  sizes=\"auto, (max-width: 872px) 100vw, 872px\" \/><\/div><\/figure>\n<\/div>\n\n\n<h1 class=\"wp-block-heading\">1 \u6570\u636e\u9a71\u52a8\u65b9\u6cd5<\/h1>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"header-id-4\">1.1 \u5f53\u524d\u7684\u6311\u6218<\/h2>\n\n\n\n<p>\u867d\u7136\u56fe\u7247\u8bc6\u522b\u5bf9\u4e8e\u4eba\u6765\u8bf4\u662f\u4e00\u4ef6\u8f7b\u677e\u7684\u4e8b\u60c5\uff0c\u4f46\u662f\u5bf9\u4e8e\u8ba1\u7b97\u673a\u6765\u8bf4\uff0c\u7531\u4e8e\u63a5\u53d7\u7684\u662f\u4e00\u4e32\u6570\u5b57\uff0c\u5bf9\u4e8e\u540c\u4e00\u4e2a\u7269\u4f53\uff0c\u8868\u793a\u8fd9\u4e2a\u7269\u4f53\u7684\u6570\u5b57\u53ef\u80fd\u4f1a\u6709\u5f88\u5927\u7684\u4e0d\u540c\uff0c\u6240\u4ee5\u4f7f\u7528\u7b97\u6cd5\u6765\u5b9e\u73b0\u8fd9\u4e00\u4efb\u52a1\u8fd8\u662f\u6709\u5f88\u591a\u6311\u6218\u7684\uff0c\u5177\u4f53\u6765\u8bf4\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u89c2\u5bdf\u89d2\u5ea6\u7684\u53d8\u5316&nbsp;<strong>Viewpoint variation<\/strong>\uff1a\u4e00\u53e5\u8bd7\u53ef\u4ee5\u5f88\u597d\u6982\u62ec\uff0c\u201c\u4e0d\u8bc6\u5e90\u5c71\u771f\u9762\u76ee\uff0c\u53ea\u7f18\u8eab\u5728\u6b64\u5c71\u4e2d\u201d\u3002<\/li>\n\n\n\n<li>\u5c3a\u5ea6\u53d8\u6362&nbsp;<strong>Scale variation<\/strong>\uff1a\u56fe\u7247\u5927\u5c0f\u6bd4\u4f8b\u7684\u53d8\u5316\u4e5f\u4f1a\u4f7f\u5f97\u6570\u636e\u53d1\u751f\u6539\u53d8\u3002<\/li>\n\n\n\n<li>\u53d8\u5f62&nbsp;<strong>Deformation<\/strong>\uff1a\u5f88\u591a\u7269\u4f53\u7684\u5916\u5f62\u4e0d\u662f\u4e00\u6210\u4e0d\u53d8\u7684\uff0c\u6bd4\u5982\u4f17\u6240\u5468\u77e5\uff0c\u732b\u662f\u6db2\u4f53\u3002<\/li>\n\n\n\n<li>\u906e\u6321&nbsp;<strong>Occlusion<\/strong>\uff1a\u8981\u88ab\u8bc6\u522b\u7684\u7269\u4f53\u53ef\u80fd\u88ab\u906e\u6321\uff0c\u53ea\u9732\u51fa\u4e00\u90e8\u5206\u3002<\/li>\n\n\n\n<li>\u5149\u7ebf\u6761\u4ef6&nbsp;<strong>Illumination conditions<\/strong>\uff1a\u73af\u5883\u5149\u7ebf\u7684\u53d8\u5316\u5bf9\u7269\u4f53\u7684\u56fe\u7247\u4e5f\u4f1a\u6709\u5f88\u5927\u7684\u5f71\u54cd\u3002<\/li>\n\n\n\n<li>\u80cc\u666f\u5e72\u6270&nbsp;<strong>Background clutter<\/strong>\uff1a\u5982\u679c\u7269\u4f53\u548c\u80cc\u666f\u6709\u5f88\u76f8\u4f3c\u7684\u989c\u8272\u548c\u7eb9\u8def\uff0c\u90a3\u4e48\u5c31\u5f88\u96be\u88ab\u8bc6\u522b\u3002<\/li>\n\n\n\n<li>\u7269\u79cd\u53d8\u5f02&nbsp;<strong>Intra-class variation<\/strong>\uff1a\u540c\u4e00\u7269\u79cd\u53ef\u80fd\u4e5f\u6709\u5dee\u5f02\u5f88\u5927\u7684\u5f62\u6001\u3002<\/li>\n<\/ul>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-large\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/icalkzhangzihao.com\/wp-content\/uploads\/2025\/11\/image-1-1024x385.png'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"385\" data-original=\"https:\/\/icalkzhangzihao.com\/wp-content\/uploads\/2025\/11\/image-1-1024x385.png\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"\" class=\"wp-image-409\"  sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/div><\/figure>\n<\/div>\n\n\n<h2 class=\"wp-block-heading\">1.2 \u6570\u636e\u9a71\u52a8\u65b9\u6cd5<\/h2>\n\n\n\n<p>\u90a3\u4e48\u6211\u4eec\u5982\u4f55\u8bbe\u8ba1\u7b97\u6cd5\u53bb\u5206\u8fa8\u4e0d\u540c\u7684\u7c7b\u522b\u5462\uff1f\u6211\u4eec\u4e0d\u4f1a\u53bb\u8bbe\u8ba1\u4e00\u4e2a\u7279\u5b9a\u7684\u7b97\u6cd5\u6765\u89e3\u51b3\u8fd9\u6837\u7684\u95ee\u9898\uff0c\u800c\u662f\u5c06\u5927\u91cf\u5e26\u6709\u6807\u7b7e\u7684\u6570\u636e\u9001\u7ed9\u4e00\u4e2a\u6a21\u578b\uff0c\u8ba9\u6a21\u578b\u81ea\u5df1\u5b66\u4e60\uff0c\u8fd9\u79cd\u65b9\u5f0f\u5c31\u6210\u4e3a\u6570\u636e\u9a71\u52a8\u65b9\u6cd5\uff0c\u56e0\u4e3a\u5b83\u4f9d\u8d56\u4e8e\u4e00\u4e2a\u5e26\u6709\u6807\u7b7e\u7684\u6570\u636e\u96c6\u5408\u3002<\/p>\n\n\n\n<p>\u6240\u4ee5\u901a\u5e38\u56fe\u7247\u8bc6\u522b\u4efb\u52a1\u7684\u6d41\u6c34\u7ebf\u5982\u4e0b\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u8f93\u5165\uff1a\u8f93\u5165&nbsp;N&nbsp;\u5f20\u56fe\u7247\uff0c\u56fe\u7247\u7684\u603b\u7c7b\u522b\u6570\u91cf\u4e3a&nbsp;K\uff0c\u6211\u4eec\u79f0\u8fd9\u4e00\u90e8\u5206\u7684\u6570\u636e\u4e3a\u8bad\u7ec3\u96c6\u3002<\/li>\n\n\n\n<li>\u5b66\u4e60\uff1a\u4f7f\u7528\u6a21\u578b\u5728\u8bad\u7ec3\u96c6\u4e2d\u5b66\u4e60\uff0c\u63d0\u53d6\u6bcf\u4e00\u4e2a\u79cd\u7c7b\u7684\u7279\u5f81\u3002\u6211\u4eec\u79f0\u4e4b\u4e3a \u8bad\u7ec3\u4e00\u4e2a\u6a21\u578b\u6216\u8005\u8bad\u7ec3\u4e00\u4e2a\u5206\u7c7b\u5668\u3002<\/li>\n\n\n\n<li>\u8bc4\u4f30\uff1a\u5728\u6700\u540e\uff0c\u6211\u4eec\u9700\u8981\u8bc4\u4f30\u8fd9\u4e2a\u8bad\u7ec3\u7684\u6a21\u578b\u597d\u574f\u3002\u8fd9\u65f6\u9700\u8981\u4e00\u4e2a\u4e4b\u524d\u4ece\u6765\u90fd\u6ca1\u4f7f\u7528\u8fc7\u7684\u65b0\u6570\u636e\u96c6\uff08\u4fdd\u8bc1\u7c7b\u522b\u4e5f\u5728&nbsp;K&nbsp;\u7c7b\u4e4b\u4e2d\uff09\uff0c\u7136\u540e\u5728\u65b0\u7684\u6570\u636e\u96c6\u4e0a\u9884\u6d4b\u6bcf\u5f20\u56fe\u7247\u7684\u79cd\u7c7b\uff0c\u6211\u4eec\u671f\u671b\u7684\u662f\u5206\u7c7b\u6b63\u786e\u7684\u56fe\u7247\u8d8a\u591a\u8d8a\u597d\u3002<\/li>\n<\/ul>\n\n\n\n<h1 class=\"wp-block-heading\" id=\"header-id-6\">2. \u6700\u8fd1\u90bb\u57df\u5206\u7c7b\u5668 NN<\/h1>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"header-id-7\">2.1 \u6570\u636e\u96c6\u548c\u539f\u7406<\/h2>\n\n\n\n<p>\u9996\u5148\u6211\u4eec\u6765\u4ecb\u7ecd\u4e00\u4e0b\u6700\u8fd1\u90bb\u57df\u5206\u7c7b\u5668\uff0c\u8fd9\u662f\u4e00\u4e2a\u5341\u5206\u7b80\u5355\u5e76\u4e14\u4e0d\u5e38\u7528\u4e8e\u5206\u7c7b\u7684\u7b97\u6cd5\uff0c\u4f46\u662f\u901a\u8fc7\u8fd9\u4e2a\u7b97\u6cd5\uff0c \u6211\u4eec\u4e5f\u53ef\u4ee5\u5927\u81f4\u4e86\u89e3\u89e3\u51b3\u56fe\u7247\u5206\u7c7b\u95ee\u9898\u7684\u5927\u81f4\u65b9\u6cd5\u3002\u672c\u6b21\u4f7f\u7528\u7684\u6570\u636e\u96c6\u662fMNIST\uff08\u624b\u5199\u6570\u5b57\u8bc6\u522b\uff09 \u548c<\/p>\n\n\n\n<p>  CIFAR-10\uff08\u8fd9\u662f\u4e00\u4e2a\u6709\u540d\u7684\u516c\u5f00\u56fe\u7247\u6570\u636e\u96c6\uff0c\u753160000\u5f20&nbsp;&nbsp;\u7684\u56fe\u7247\u7ec4\u6210\uff0c\u4e00\u5171\u670910\u4e2a\u79cd\u7c7b\uff0c\u4e00\u822c\u6211\u4eec\u5c06\u5176\u4e2d\u768450000\u5f20\u4f5c\u4e3a\u8bad\u7ec3\u96c6\uff0c10000\u5f20\u4f5c\u4e3a\u6d4b\u8bd5\u96c6\uff0c\u4e0b\u56fe\u5de6\u5c31\u662f10\u4e2a\u7c7b\u522b\u7684\u90e8\u5206\u56fe\u7247\u3002\uff09<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-large\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/icalkzhangzihao.com\/wp-content\/uploads\/2025\/11\/image-2-1024x414.png'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"414\" data-original=\"https:\/\/icalkzhangzihao.com\/wp-content\/uploads\/2025\/11\/image-2-1024x414.png\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"\" class=\"wp-image-410\"  sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/div><\/figure>\n<\/div>\n\n\n<p>\u73b0\u5728\u6211\u4eec\u7684\u8bad\u7ec3\u96c6\u4e2d\u5c31\u6709\u4e8650000\u5f20\u56fe\u7247\uff0c\u6bcf\u4e2a\u7c7b\u522b5000\u5f20\uff0c\u5bf9\u4e8e\u6d4b\u8bd5\u96c610000\u5f20\u56fe\u7247\u4e2d\u7684\u6bcf\u4e00\u5f20\u56fe\u7247\uff0c\u6211\u4eec\u8981\u505a\u7684\u662f\u5c06\u5176\u4e0e\u8bad\u7ec3\u96c6\u4e2d\u7684\u6bcf\u4e00\u5f20\u56fe\u7247\u8fdb\u884c\u6bd4\u8f83\uff0c\u7136\u540e\u5c06\u8fd9\u79cd\u56fe\u7247\u4e0e\u8bad\u7ec3\u96c6\u4e2d\u6700\u76f8\u4f3c\u7684\u56fe\u7247\u5f52\u4e3a\u4e00\u7c7b\uff0c\u4e0a\u56fe\u53f3\u5c31\u662f\u90e8\u5206\u5206\u7c7b\u540e\u7684\u7ed3\u679c\uff0c\u53ef\u4ee5\u53d1\u73b0\uff0c\u5b58\u5728\u5f88\u591a\u7684\u8bef\u5206\u7c7b\uff0c\u539f\u56e0\u5728\u4e8e\u867d\u7136\u56fe\u7247\u7684\u79cd\u7c7b\u4e0d\u540c\uff0c\u4f46\u662f\u4e24\u79cd\u56fe\u7247\u7684\u989c\u8272\u56fe\u6848\u7b49\u975e\u5e38\u7c7b\u4f3c\uff0c\u5c31\u5bb9\u6613\u88ab\u5f52\u4e3a\u4e00\u7c7b\u3002<\/p>\n\n\n\n<p>\u90a3\u4e48\u5728\u6700\u8fd1\u90bb\u57df\u7b97\u6cd5\u4e2d\uff0c\u6211\u4eec\u8861\u91cf\u4e24\u5f20\u56fe\u7247\u662f\u5426\u76f8\u8fd1\u7684\u6807\u51c6\u662f\u4ec0\u4e48\u5462\uff1f\u4e00\u79cd\u6700\u7b80\u5355\u7684\u6807\u51c6\u5c31\u662f L1\u8ddd\u79bb\uff0c\u5047\u8bbe\u6211\u4eec\u5c06\u4e24\u5f20\u56fe\u7247\u5206\u522b\u8868\u793a\u4e3a\u4e24\u4e2a\u5411\u91cf&nbsp;I1\u3001I2\uff0c\u90a3\u4e48L1\u8ddd\u79bb\u7684\u5b9a\u4e49\u5982\u4e0b\uff1a<\/p>\n\n\n\n<p>$$L1(I1, I2)=\\sum_p|I1-I2|$$<\/p>\n\n\n\n<p>\u4e00\u4e2a\u7b80\u5355\u7684\u8ba1\u7b97\u6d41\u7a0b\u6f14\u793a\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/icalkzhangzihao.com\/wp-content\/uploads\/2025\/11\/image-3-1024x314.png'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"314\" data-original=\"https:\/\/icalkzhangzihao.com\/wp-content\/uploads\/2025\/11\/image-3-1024x314.png\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"\" class=\"wp-image-413\"  sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/div><\/figure>\n\n\n\n<p>\u5bf9\u4e8e\u4e24\u79cd\u56fe\u7247\u7684\u8861\u91cf\u6807\u51c6\u8fd8\u6709 L2 \u8ddd\u79bb\uff0c\u5b9a\u4e49\u5982\u4e0b\uff1a<\/p>\n\n\n\n<p>$$L2(I1, I2)=\\sum_p(I1-I2)^2$$<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"header-id-8\">2.2 \u4ee3\u7801\u5b9e\u73b0<\/h2>\n\n\n\n<h4 class=\"wp-block-heading\">\u8bfe\u7a0b\u4e00\u5185\u5bb9<\/h4>\n\n\n\n<p>\u5b89\u88c5\u865a\u62df\u73af\u5883\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>pip install numpy matplotlib torchvision<\/code><\/pre>\n\n\n\n<p>\u5bfc\u5165\u76ee\u6807\u5904\u7406\u7684\u5e93<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>import os\nimport numpy as np\nimport matplotlib.pyplot as plt\nfrom torchvision.datasets import MNIST<\/code><\/pre>\n\n\n\n<p>\u9996\u5148\u6211\u4eec\u9700\u8981\u5904\u7406 MNIST \u6570\u636e\u96c6\uff0c\u5229\u7528torchvision\u4e0b\u8f7d\u6570\u636e\u96c6\u3002<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>data_path = r'C:\\code\\ccs2\\check\\data'\n\ndef downloaddataset():\n# \u4e0b\u8f7d\u6570\u636e\u96c6\n    train_dataset_no = MNIST(data_path, train=True, download=True)\n    test_dataset_no = MNIST(data_path, train=False, download=True)\ndownloaddataset()<\/code><\/pre>\n\n\n\n<p>\u5c06\u5176\u4ee5\u56db\u4e2a\u6570\u7ec4\u7684\u5f62\u5f0f\u8868\u793a\uff0c\u5206\u522b\u4e3a \u8bad\u7ec3\u96c6\u6570\u636e\u3001\u8bad\u7ec3\u96c6\u6807\u7b7e\u3001\u6d4b\u8bd5\u96c6\u6570\u636e\u3001\u6d4b\u8bd5\u96c6\u6807\u7b7e\uff0c\u4e0b\u9762\u7684\u4ee3\u7801\u4e2d&nbsp;<code>Xtr<\/code>&nbsp;\u8868\u793a\u8bad\u7ec3\u96c6\u6570\u636e\uff0c<code>Ytr<\/code>&nbsp;\u8868\u793a\u8bad\u7ec3\u96c6\u6807\u7b7e\uff0c\u5f97\u5230\u6570\u636e\u540e\u5c06\u5176\u62c9\u6210\u4e00\u6761\u5411\u91cf\uff0c\u4fbf\u4e8e\u8ba1\u7b97\u3002<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>def load_mnist(root=data_path, download=False):\n    <em>\"\"\"\n    \u8bfb MNIST \u624b\u5199\u6570\u5b57\u3002\n    \u8fd4\u56de\u56db\u4e2a\u4e1c\u897f\uff1a\n      Xtr: \u8bad\u7ec3\u56fe\u7247\uff0c\u5f62\u72b6(\u8bad\u7ec3\u6570\u91cf, 28, 28)\n      Ytr: \u8bad\u7ec3\u6807\u7b7e\uff0c\u5f62\u72b6(\u8bad\u7ec3\u6570\u91cf,)\n      Xte: \u6d4b\u8bd5\u56fe\u7247\uff0c\u5f62\u72b6(\u6d4b\u8bd5\u6570\u91cf, 28, 28)\n      Yte: \u6d4b\u8bd5\u6807\u7b7e\uff0c\u5f62\u72b6(\u6d4b\u8bd5\u6570\u91cf,)\n    \"\"\"\n    <\/em># 1) \u8bfb\u53d6\u6570\u636e\u96c6\uff08\u5982\u679c\u672c\u5730\u6ca1\u6709\u4e14 download=True\uff0c\u5c31\u4f1a\u81ea\u52a8\u4e0b\u8f7d\uff09\n    train_set = MNIST(root=root, train=True,  download=download)   # 6\u4e07\u5f20\n    test_set  = MNIST(root=root, train=False, download=download)   # 1\u4e07\u5f20\n\n    # 2) \u628a\u8bad\u7ec3\u96c6\u91cc\u7684\u56fe\u7247\u548c\u6807\u7b7e\u4e00\u4e2a\u4e00\u4e2a\u53d6\u51fa\u6765\uff0c\u88c5\u8fdb\u5217\u8868\n    train_images = &#91;]   # \u5b58\u56fe\u7247\n    train_labels = &#91;]   # \u5b58\u6807\u7b7e\n    for img, lbl in train_set:\n        # img \u662f\u4e00\u5f20 28x28 \u7684\u7070\u5ea6\u56fe\uff08PIL \u56fe\u7247\uff09\uff0c\u5148\u628a\u5b83\u53d8\u6210 numpy \u6570\u7ec4\n        img_array = np.array(img, dtype=np.uint8)\n        # print(img_array.shape)\n        train_images.append(img_array)\n        train_labels.append(int(lbl))\n\n    # 3) \u540c\u6837\u7684\u65b9\u6cd5\u5904\u7406\u6d4b\u8bd5\u96c6\n    test_images = &#91;]\n    test_labels = &#91;]\n    for img, lbl in test_set:\n        img_array = np.array(img, dtype=np.uint8)\n        test_images.append(img_array)\n        test_labels.append(int(lbl))\n\n    # 4) \u628a\u5217\u8868\u8f6c\u6210 numpy \u6570\u7ec4\uff0c\u65b9\u4fbf\u540e\u9762\u8ba1\u7b97\n    Xtr = np.array(train_images)   # (60000, 28, 28)\n    Ytr = np.array(train_labels)   # (60000,)\n    Xte = np.array(test_images)   # (10000, 28, 28)\n    Yte = np.array(test_labels)    # (10000,)\n\n    return Xtr, Ytr, Xte, Yte\n\n# ---------- \u4e3b\u6d41\u7a0b ----------\nXtr, Ytr, Xte, Yte = load_mnist(download=False)   # \u82e5\u672c\u5730\u6ca1\u6570\u636e\uff0c\u53ef\u6539\u4e3a True<\/code><\/pre>\n\n\n\n<p><\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u8bfe\u7a0b\u4e8c\u5185\u5bb9<\/h4>\n\n\n\n<p>\u53ef\u89c6\u5316\u4e00\u5f20\u56fe\u7247<\/p>\n\n\n\n<p><\/p>\n\n\n\n<pre class=\"wp-block-code\"><code># \u53ef\u9009\uff1a\u67e5\u770b\u4e00\u5f20\u8bad\u7ec3\u56fe\u7247\nSAMPLE_INDEX = 0\nimg = Xtr&#91;SAMPLE_INDEX]\nlabel = Ytr&#91;SAMPLE_INDEX]\nprint(\"Label:\", label)\nplt.imshow(img, cmap='gray')\nplt.title(f\"Label = {label}\")\nplt.axis('off')\nplt.show()\nplt.imsave('mnist_sample.png', img, cmap='gray')\nprint(\"\u56fe\u7247\u5df2\u901a\u8fc7 matplotlib \u4fdd\u5b58\u4e3a mnist_sample.png\")\n<\/code><\/pre>\n\n\n\n<p>\u6a21\u578b\u6784\u5efa\u8fc7\u7a0b<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>class NearestNeighbor(object):\n    def __init__(self):\n        pass\n\n    def train(self, X, y):\n        self.Xtr = X\n        self.ytr = y\n\n    def predict(self, X, k=1):\n        <em>\"\"\" X is N x D where each row is an example we wish to predict label for \"\"\"\n<\/em><em>        <\/em>num_test = X.shape&#91;0]\n        Ypred = np.zeros(num_test, dtype=self.ytr.dtype)\n\n        for i in range(num_test):\n            distances = np.sum(np.abs(self.Xtr - X&#91;i, :]), axis=1)\n\n            # \u627e\u5230\u8ddd\u79bb\u6700\u5c0f\u7684 k \u4e2a\u8bad\u7ec3\u6837\u672c \n            knn_indices = np.argsort(distances)&#91;:k]\n            knn_labels = self.ytr&#91;knn_indices]\n\n            # \u591a\u6570\u6295\u7968\uff08k=1 \u65f6\u7b49\u4ef7\u4e8e\u539f\u59cb\u7248\u672c\uff09 \n            Ypred&#91;i] = np.bincount(knn_labels).argmax()\n\n        return Ypred<\/code><\/pre>\n\n\n\n<p><\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u8bfe\u7a0b\u4e09\u5185\u5bb9<\/h4>\n\n\n\n<p>\u6574\u4f53pipline\u642d\u5efa<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code># ---------- \u4e3b\u6d41\u7a0b ----------\nXtr, Ytr, Xte, Yte = load_mnist(download=False)   # \u82e5\u672c\u5730\u6ca1\u6570\u636e\uff0c\u53ef\u6539\u4e3a True\nprint(\"Train\/Test shapes:\", Xtr.shape, Xte.shape) # (60000, 28, 28) \/ (10000, 28, 28)\n\n# \u62c9\u5e73\u6210\u5411\u91cf\u4ee5\u4f9b KNN \u4f7f\u7528\nXtr_rows = Xtr.reshape(Xtr.shape&#91;0], -1)&#91;:1000]  # (60000, 784)\nYtr = Ytr&#91;:1000]\nXte_rows = Xte.reshape(Xte.shape&#91;0], -1)&#91;:50]  # (10000, 784)\nYte = Yte&#91;:50]\n# \u9009\u53d6\u9a8c\u8bc1\u96c6\u6700\u4f18 k\n# \u627e\u51fa\u9a8c\u8bc1\u96c6\u91cc\u54ea\u4e2a k \u7684\u51c6\u786e\u7387\u6700\u9ad8\nbest_k = 1\n\n# \u7528\u6700\u4f18 k \u5728\u6d4b\u8bd5\u96c6\u8bc4\u4f30\uff08\u53ef\u4e0e\u66f4\u5927\u7684\u8bad\u7ec3\u96c6\u7ec4\u5408\uff09\n# \u8fd9\u91cc\u5c06\u8bad\u7ec3\u96c6\u6269\u5927\u4e3a\u5168\u90e8 60000\uff0c\u7528\u4e8e\u6700\u7ec8\u6d4b\u8bd5\u8bc4\u4f30\uff08\u53ef\u80fd\u8f83\u6162\uff0c\u914c\u60c5\u6539\u4e3a\u90e8\u5206\uff09\nnn = NearestNeighbor()\nnn.train(Xtr_rows, Ytr)\nYte_pred = nn.predict(Xte_rows, k=best_k)\ntest_acc = np.mean(Yte_pred == Yte)\nprint(f'test acc (k={best_k}): {test_acc*100:.2f}')<\/code><\/pre>\n\n\n\n<p>\u53ef\u89c6\u5316\u6548\u679c<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>import os\nimport numpy as np\nimport matplotlib.pyplot as plt\nfrom class_all.class2 import NearestNeighbor  # \u9700\u4e3a\u652f\u6301 k \u7684\u7248\u672c\nfrom class_all.class1 import load_mnist\n\n# ---------- \u4e3b\u6d41\u7a0b ----------\n# \u52a0\u8f7d MNIST \u6570\u636e\u96c6\uff08\u82e5\u672c\u5730\u65e0\u6570\u636e\uff0c\u53ef\u628a download=True \u4ee5\u81ea\u52a8\u4e0b\u8f7d\uff09\nXtr, Ytr, Xte, Yte = load_mnist(download=False)\nprint(\"Train\/Test shapes:\", Xtr.shape, Xte.shape)  # (60000, 28, 28) \/ (10000, 28, 28)\n\n# \u5c06\u56fe\u50cf\u7531 28x28 \u62c9\u5e73\u6210 784 \u7ef4\u5411\u91cf\uff0c\u4fbf\u4e8e\u6700\u8fd1\u90bb\u5206\u7c7b\u5668\u8fdb\u884c\u6b27\u6c0f\u8ddd\u79bb\u8ba1\u7b97\n# \u8fd9\u91cc\u4e3a\u4e86\u8fd0\u884c\u901f\u5ea6\uff0c\u53ea\u53d6\u524d 1000 \u4e2a\u8bad\u7ec3\u6837\u672c\uff0c\u4ee5\u53ca\u524d 50 \u4e2a\u6d4b\u8bd5\u6837\u672c\nXtr_rows = Xtr.reshape(Xtr.shape&#91;0], -1)&#91;:1000]\nYtr = Ytr&#91;:1000]\nXte_rows = Xte.reshape(Xte.shape&#91;0], -1)&#91;:50]\nYte = Yte&#91;:50]\n\n# \u8bbe\u5b9a k \u503c\uff0c\u8fd9\u91cc\u56fa\u5b9a\u4e3a 1\uff08\u5373\u6700\u8fd1\u90bb\u6cd5\uff09\nbest_k = 1\n\n# \u8bad\u7ec3\u5e76\u8bc4\u4f30\uff1a\u5c06\u8bad\u7ec3\u7279\u5f81\u4e0e\u6807\u7b7e\u201c\u5582\u7ed9\u201d\u6700\u8fd1\u90bb\u5206\u7c7b\u5668\uff0c\u7136\u540e\u5728\u6d4b\u8bd5\u96c6\u505a\u9884\u6d4b\u5e76\u8ba1\u7b97\u51c6\u786e\u7387\nnn = NearestNeighbor()\nnn.train(Xtr_rows, Ytr)\nYte_pred = nn.predict(Xte_rows, k=best_k)\ntest_acc = np.mean(Yte_pred == Yte)\nprint(f'test acc (k={best_k}): {test_acc*100:.2f}')\n\n# ---------- \u4ec5\u53ef\u89c6\u5316\u201c\u6700\u540e 50 \u4e2a\u6837\u672c\u201d\u7684\u9884\u6d4b\u7ed3\u679c\u4e0e\u5176\u5bf9\u5e94\u7684 Top-1 \u8bad\u7ec3\u90bb\u5c45 ----------\nimport numpy as np\nimport matplotlib.pyplot as plt\n\n# \u6307\u5b9a\u9700\u8981\u5c55\u793a\u7684\u6837\u672c\u6570\u91cf\uff08\u8fd9\u91cc\u4e3a 50\uff09\n# \u82e5\u5f53\u524d\u6d4b\u8bd5\u5b50\u96c6\u672c\u8eab\u5c11\u4e8e 50\uff0c\u5219\u6309\u5b9e\u9645\u6570\u91cf\u5c55\u793a\nn = 50\n\n# \u4e3a\u4e86\u907f\u514d\u4e0e\u4e0a\u6e38\u88c1\u526a\u4e0d\u4e00\u81f4\uff0c\u8fd9\u91cc\u76f4\u63a5\u57fa\u4e8e\u5f53\u524d\u5b50\u96c6\u53d6\u524d n \u4e2a\uff08\u7b49\u4ef7\u4e8e\u201c\u6700\u540e 50\u201d\u5728\u4f60\u5f53\u524d\u7684\u5207\u7247\u914d\u7f6e\u4e0b\uff09\n# \u4e09\u8005\u5207\u7247\u5fc5\u987b\u4e00\u81f4\uff1a\u6d4b\u8bd5\u7279\u5f81\u5411\u91cf\u3001\u6d4b\u8bd5\u6807\u7b7e\u3001\u4ee5\u53ca\u7528\u4e8e\u663e\u793a\u7684\u6d4b\u8bd5\u56fe\u50cf\nXte_last_rows = Xte_rows\nimgs_last = Xte_rows.reshape(n, 28, 28)     # \u5c06\u5411\u91cf\u8fd8\u539f\u4e3a 28x28 \u4fbf\u4e8e\u663e\u793a\nYte_last = Yte\nYte_pred_last = Yte_pred\n\n# \u9010\u4e2a\u8ba1\u7b97\u6bcf\u4e2a\u6d4b\u8bd5\u6837\u672c\u4e0e\u5168\u90e8\u8bad\u7ec3\u6837\u672c\u4e4b\u95f4\u7684\u6b27\u6c0f\u8ddd\u79bb\uff0c\u627e\u5230\u8ddd\u79bb\u6700\u5c0f\u7684\u90a3\u4e2a\u8bad\u7ec3\u6837\u672c\n# \u4fdd\u5b58\u5176\u7d22\u5f15\u3001\u8ddd\u79bb\u503c\uff0c\u7528\u4e8e\u540e\u7eed\u5c55\u793a\u6700\u8fd1\u90bb\u56fe\u50cf\u4e0e\u6807\u7b7e\nnn_indices = np.empty(n, dtype=int)\nnn_dists = np.empty(n, dtype=float)\nfor i in range(n):\n    x = Xte_last_rows&#91;i:i+1]                     # \u53d6\u7b2c i \u4e2a\u6d4b\u8bd5\u6837\u672c\uff08\u5f62\u72b6\uff1a(1, 784)\uff09\n    dists = np.linalg.norm(Xtr_rows - x, axis=1) # \u8ba1\u7b97\u4e0e\u6240\u6709\u8bad\u7ec3\u6837\u672c\u7684\u6b27\u6c0f\u8ddd\u79bb\uff08\u5f62\u72b6\uff1a(N_train,)\uff09\n    j = int(np.argmin(dists))                    # \u6700\u5c0f\u8ddd\u79bb\u5bf9\u5e94\u7684\u8bad\u7ec3\u6837\u672c\u7d22\u5f15\n    nn_indices&#91;i] = j\n    nn_dists&#91;i] = float(dists&#91;j])\n\n# \u6839\u636e\u6700\u8fd1\u90bb\u7d22\u5f15\u53d6\u51fa\u5bf9\u5e94\u7684\u8bad\u7ec3\u56fe\u50cf\u4e0e\u6807\u7b7e\uff08\u7528\u4e8e\u53ef\u89c6\u5316 Top-1 \u90bb\u5c45\uff09\nnn_imgs = Xtr&#91;nn_indices]                        # \u5f62\u72b6\uff1a(n, 28, 28)\nnn_labels = Ytr&#91;nn_indices]                      # \u5f62\u72b6\uff1a(n,)\n\n# \u7ed8\u5236\u51fd\u6570 1\uff1a\u4ee5\u7f51\u683c\u5f62\u5f0f\u5c55\u793a\u6d4b\u8bd5\u56fe\u50cf\uff0c\u5e76\u5728\u6807\u9898\u6807\u6ce8\u9884\u6d4b\u503c\/\u771f\u5b9e\u503c\n# \u7eff\u8272\u8868\u793a\u9884\u6d4b\u6b63\u786e\uff0c\u7ea2\u8272\u8868\u793a\u9884\u6d4b\u9519\u8bef\ndef show_test_grid(images, y_true, y_pred, rows=5, cols=10, title=\"Last-N test samples (green=correct, red=wrong)\"):\n    n_show = min(len(images), rows*cols)         # \u9632\u6b62\u8d85\u51fa\u7f51\u683c\u5bb9\u91cf\n    plt.figure(figsize=(cols*1.6, rows*1.6))\n    for i in range(n_show):\n        ax = plt.subplot(rows, cols, i+1)\n        ax.imshow(images&#91;i], cmap='gray')        # \u4f7f\u7528\u7070\u5ea6\u663e\u793a\u624b\u5199\u6570\u5b57\u56fe\u50cf\n        ax.axis('off')                           # \u5173\u95ed\u5750\u6807\u8f74\n        ok = (y_pred&#91;i] == y_true&#91;i])            # \u5224\u65ad\u662f\u5426\u9884\u6d4b\u6b63\u786e\n        ax.set_title(\n            f\"P:{int(y_pred&#91;i])} \/ T:{int(y_true&#91;i])}\",\n            fontsize=9,\n            color=('green' if ok else 'red')     # \u6b63\u786e\u4e3a\u7eff\uff0c\u9519\u8bef\u4e3a\u7ea2\n        )\n    plt.suptitle(title, fontsize=14, y=1.02)\n    plt.tight_layout()\n    plt.show()\n\n# \u7ed8\u5236\u51fd\u6570 2\uff1a\u4ee5\u7f51\u683c\u5f62\u5f0f\u5c55\u793a\u4e0e\u6bcf\u4e2a\u6d4b\u8bd5\u6837\u672c\u5bf9\u5e94\u7684\u201c\u6700\u76f8\u4f3c\u8bad\u7ec3\u56fe\u50cf\uff08Top-1 \u90bb\u5c45\uff09\u201d\n# \u6807\u9898\u4e2d\u9644\u5e26\u8be5\u90bb\u5c45\u7684\u6807\u7b7e\u4e0e\u5230\u6d4b\u8bd5\u6837\u672c\u7684\u6b27\u6c0f\u8ddd\u79bb\uff0c\u4fbf\u4e8e\u7406\u89e3\u5224\u522b\u4f9d\u636e\ndef show_neighbor_grid(images, labels, dists, rows=5, cols=10, title=\"Top-1 training neighbor per test sample\"):\n    n_show = min(len(images), rows*cols)\n    plt.figure(figsize=(cols*1.6, rows*1.6))\n    for i in range(n_show):\n        ax = plt.subplot(rows, cols, i+1)\n        ax.imshow(images&#91;i], cmap='gray')\n        ax.axis('off')\n        ax.set_title(f\"NN:{int(labels&#91;i])} \/ D:{dists&#91;i]:.1f}\", fontsize=9)\n    plt.suptitle(title, fontsize=14, y=1.02)\n    plt.tight_layout()\n    plt.show()\n\n# \u4f9d\u6b21\u7ed8\u5236\uff1a\n# 1) \u6d4b\u8bd5\u6837\u672c\u7f51\u683c\uff08\u5c55\u793a\u9884\u6d4b\u503c\u4e0e\u771f\u5b9e\u503c\uff09\n# 2) \u5bf9\u5e94\u7684 Top-1 \u8bad\u7ec3\u90bb\u5c45\u7f51\u683c\uff08\u5c55\u793a\u90bb\u5c45\u6807\u7b7e\u4e0e\u8ddd\u79bb\uff09\nshow_test_grid(imgs_last, Yte_last, Yte_pred_last, rows=5, cols=10)\nshow_neighbor_grid(nn_imgs, nn_labels, nn_dists, rows=5, cols=10)<\/code><\/pre>\n\n\n\n<p><\/p>\n\n\n\n<p><\/p>\n\n\n\n<p>\u6df1\u5ea6\u5b66\u4e60\u4ee3\u7801\u4ecb\u7ecd<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>import torch\nimport torch.nn as nn\nimport torch.optim as optim\nfrom torchvision import datasets, transforms\nfrom torch.utils.data import DataLoader\n\n# 1. \u6570\u636e\u51c6\u5907 (\u5305\u542b\u4e0b\u8f7d\u3001\u9884\u5904\u7406\u4e0e\u52a0\u8f7d)\ntransform = transforms.Compose(&#91;\n    transforms.ToTensor(),\n    transforms.Normalize((0.1307,), (0.3081,))\n])\ndata_path = r'C:\\code\\ccs2\\check\\data'\n# \u6839\u76ee\u5f55\u5efa\u8bae\u6539\u4e3a\u4f60\u4e60\u60ef\u7684\u8def\u5f84\uff0c\u4f8b\u5982 '.\/data'\ntrain_set = datasets.MNIST(root=data_path, train=True, download=True, transform=transform)\ntest_set = datasets.MNIST(root=data_path, train=False, download=True, transform=transform)\n\ntrain_loader = DataLoader(train_set, batch_size=64, shuffle=True)\ntest_loader = DataLoader(test_set, batch_size=1000, shuffle=False)\n\n# 2. \u5b9a\u4e49\u7cbe\u70bc\u7684 CNN \u6a21\u578b\nclass TinyCNN(nn.Module):\n    def __init__(self):\n        super().__init__()\n        self.net = nn.Sequential(\n            nn.Conv2d(1, 16, 3, padding=1), nn.ReLU(), nn.MaxPool2d(2),\n            nn.Conv2d(16, 32, 3, padding=1), nn.ReLU(), nn.MaxPool2d(2),\n            nn.Flatten(),\n            nn.Linear(32 * 7 * 7, 128), nn.ReLU(),\n            nn.Linear(128, 10)\n        )\n    def forward(self, x): return self.net(x)\n\n# 3. \u8bad\u7ec3\u4e0e\u6d4b\u8bd5\u903b\u8f91\ndevice = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\nmodel = TinyCNN().to(device)\noptimizer = optim.Adam(model.parameters(), lr=0.001)\ncriterion = nn.CrossEntropyLoss()\n\nprint(f\"Using device: {device}\")\n\n# \u8bad\u7ec3 1 \u8f6e\u5373\u53ef\u8fbe\u5230\u7ea6 98% \u51c6\u786e\u7387\nmodel.train()\nfor data, target in train_loader:\n    data, target = data.to(device), target.to(device)\n    optimizer.zero_grad()\n    criterion(model(data), target).backward()\n    optimizer.step()\n\n# \u5efa\u8bae\u540e\u7f00\u4f7f\u7528 .pth \u6216 .pt\ntorch.save(model.state_dict(), \"mnist_model.pth\")\nprint(\"\u6a21\u578b\u6743\u91cd\u5df2\u6210\u529f\u4fdd\u5b58\u81f3 mnist_model.pth\")\n\n# 4. \u6d4b\u8bd5\u51c6\u786e\u7387\nmodel.eval()\ncorrect = 0\nwith torch.no_grad():\n    for data, target in test_loader:\n        data, target = data.to(device), target.to(device)\n        output = model(data)\n        correct += (output.argmax(1) == target).sum().item()\n\nprint(f\"\\nFinal Test Accuracy: {100. * correct \/ len(test_loader.dataset):.2f}%\")<\/code><\/pre>\n\n\n\n<p>\u6df1\u5ea6\u5b66\u4e60\u9884\u6d4b\u4ee3\u7801<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>import torch\nimport torch.nn as nn\nimport torch.optim as optim\nfrom torchvision import datasets, transforms\nfrom torch.utils.data import DataLoader\n\nclass TinyCNN(nn.Module):\n    def __init__(self):\n        super().__init__()\n        self.net = nn.Sequential(\n            nn.Conv2d(1, 16, 3, padding=1), nn.ReLU(), nn.MaxPool2d(2),\n            nn.Conv2d(16, 32, 3, padding=1), nn.ReLU(), nn.MaxPool2d(2),\n            nn.Flatten(),\n            nn.Linear(32 * 7 * 7, 128), nn.ReLU(),\n            nn.Linear(128, 10)\n        )\n    def forward(self, x): return self.net(x)\n\ntransform = transforms.Compose(&#91;\n    transforms.ToTensor(),\n    transforms.Normalize((0.1307,), (0.3081,))\n])\ndevice = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n\n\ndef predict_single_image(image_path):\n    # 1. \u5b9e\u4f8b\u5316\u6a21\u578b\u5e76\u52a0\u8f7d\u6743\u91cd\n    model = TinyCNN().to(device)\n    model.load_state_dict(torch.load(\"mnist_model.pth\"))\n    model.eval()  # \u5fc5\u987b\u5207\u6362\u5230\u8bc4\u4f30\u6a21\u5f0f\n\n    # 2. \u56fe\u50cf\u9884\u5904\u7406 (\u5fc5\u987b\u4e0e\u8bad\u7ec3\u65f6\u7684\u5904\u7406\u5b8c\u5168\u4e00\u81f4)\n    from PIL import Image\n    # \u8bfb\u53d6\u56fe\u7247 -&gt; \u8f6c\u4e3a\u7070\u5ea6 -&gt; \u7f29\u653e\u5230 28x28\n    img = Image.open(image_path).convert('L').resize((28, 28))\n\n    # \u8f6c\u6362\u903b\u8f91\uff1aPIL -&gt; Tensor -&gt; \u5f52\u4e00\u5316 -&gt; \u589e\u52a0 Batch \u7ef4\u5ea6 (1, 1, 28, 28)\n    img_tensor = transform(img).unsqueeze(0).to(device)\n\n    # 3. \u6267\u884c\u9884\u6d4b\n    with torch.no_grad():\n        output = model(img_tensor)\n        prediction = output.argmax(dim=1).item()\n\n    print(f\"\u8fd9\u5f20\u56fe\u7247\u7684\u8bc6\u522b\u7ed3\u679c\u662f: {prediction}\")\n    return prediction\n\npredict_single_image(r'D:\\Code\\class_all\\mnist_sample.png')<\/code><\/pre>\n","protected":false},"excerpt":{"rendered":"<p>0 \u56fe\u7247\u5206\u7c7b\u95ee\u9898 \u56fe\u7247\u5206\u7c7b\u95ee\u9898\u5c31\u662f\u8fa8\u8ba4\u8f93\u5165\u7684\u56fe\u7247\u7c7b\u522b\u7684\u95ee\u9898\uff0c\u4e14\u56fe\u7247\u7684\u7c7b\u522b\u5c5e\u4e8e\u4e8b\u5148\u7ed9\u5b9a\u7684\u4e00\u4e2a\u7c7b\u522b\u7ec4\u4e2d\u3002\u5c3d\u7ba1\u8fd9\u770b\u8d77 [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-405","post","type-post","status-publish","format-standard","hentry","category-uncategorized"],"_links":{"self":[{"href":"https:\/\/icalkzhangzihao.com\/index.php?rest_route=\/wp\/v2\/posts\/405","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/icalkzhangzihao.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/icalkzhangzihao.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/icalkzhangzihao.com\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/icalkzhangzihao.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=405"}],"version-history":[{"count":29,"href":"https:\/\/icalkzhangzihao.com\/index.php?rest_route=\/wp\/v2\/posts\/405\/revisions"}],"predecessor-version":[{"id":599,"href":"https:\/\/icalkzhangzihao.com\/index.php?rest_route=\/wp\/v2\/posts\/405\/revisions\/599"}],"wp:attachment":[{"href":"https:\/\/icalkzhangzihao.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=405"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/icalkzhangzihao.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=405"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/icalkzhangzihao.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=405"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}