{"id":6101,"date":"2022-08-23T19:24:14","date_gmt":"2022-08-23T11:24:14","guid":{"rendered":"http:\/\/139.9.1.231\/?p=6101"},"modified":"2022-11-11T11:12:59","modified_gmt":"2022-11-11T03:12:59","slug":"model-evalith-torch-no_grad","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2022\/08\/23\/model-evalith-torch-no_grad\/","title":{"rendered":"Pytorch \u4e2d model.eval() model.train() \u548c with torch.no_grad() \u7684\u533a\u522b"},"content":{"rendered":"\n<h2>1\u3001model.eval() model.train()\u533a\u522b<\/h2>\n\n\n\n<p class=\"has-light-pink-background-color has-background\">       <strong>model.train()\u548cmodel.eval()\u7684\u533a\u522b\u4e3b\u8981\u5728\u4e8eBatch Normalization\u548cDropout\u4e24\u5c42\u3002<\/strong><\/p>\n\n\n\n<p class=\"has-normal-font-size\"><a rel=\"noreferrer noopener\" href=\"https:\/\/pytorch.org\/docs\/stable\/generated\/torch.nn.Module.html#torch.nn.Module\" target=\"_blank\"><strong>\u5b98\u65b9\u6587\u6863<\/strong><\/a>  <strong>model.train()<\/strong> \uff1a<br><strong>\u542f\u7528 Batch Normalization \u548c Dropout\u3002<\/strong><br>\u5982\u679c\u6a21\u578b\u4e2d\u6709BN\u5c42(Batch Normalization\uff09\u548c Dropout\uff0c\u9700\u8981\u5728\u8bad\u7ec3\u65f6\u6dfb\u52a0<strong><code>model.train()<\/code>\u3002<code>model.train()<\/code>\u662f\u4fdd\u8bc1BN\u5c42\u80fd\u591f\u7528\u5230\u6bcf\u4e00\u6279\u6570\u636e\u7684\u5747\u503c\u548c\u65b9\u5dee\u3002\u5bf9\u4e8eDropout\uff0c<code>model.train()<\/code>\u662f\u968f\u673a\u53d6\u4e00\u90e8\u5206\u7f51\u7edc\u8fde\u63a5\u6765\u8bad\u7ec3\u66f4\u65b0\u53c2\u6570\u3002<\/strong><\/p>\n\n\n\n<p class=\"has-light-pink-background-color has-background\"><a rel=\"noreferrer noopener\" href=\"https:\/\/pytorch.org\/docs\/stable\/generated\/torch.nn.Module.html#torch.nn.Module\" target=\"_blank\">\u5b98\u65b9\u6587\u6863<\/a> <strong>model.eval()<\/strong> <br><strong>\u4e0d\u542f\u7528 Batch Normalization \u548c Dropout\u3002<\/strong><br>\u5982\u679c\u6a21\u578b\u4e2d\u6709BN\u5c42(Batch Normalization\uff09\u548cDropout\uff0c\u5728\u6d4b\u8bd5\u65f6\u6dfb\u52a0<code>model.eval()<\/code>\u3002<code>model.eval<\/code>()\u662f\u4fdd\u8bc1BN\u5c42\u80fd\u591f\u7528\u5168\u90e8\u8bad\u7ec3\u6570\u636e\u7684\u5747\u503c\u548c\u65b9\u5dee\uff0c\u5373\u6d4b\u8bd5\u8fc7\u7a0b\u4e2d\u8981\u4fdd\u8bc1BN\u5c42\u7684\u5747\u503c\u548c\u65b9\u5dee\u4e0d\u53d8\u3002\u5bf9\u4e8eDropout\uff0cmodel.eval()\u662f\u5229\u7528\u5230\u4e86\u6240\u6709\u7f51\u7edc\u8fde\u63a5\uff0c\u5373\u4e0d\u8fdb\u884c\u968f\u673a\u820d\u5f03\u795e\u7ecf\u5143\u3002<\/p>\n\n\n\n<p class=\"has-light-pink-background-color has-background\">\u8bad\u7ec3\u5b8ctrain\u6837\u672c\u540e\uff0c\u751f\u6210\u7684\u6a21\u578bmodel\u8981\u7528\u6765\u6d4b\u8bd5\u6837\u672c\u3002\u5728model(test)\u4e4b\u524d\uff0c\u9700\u8981\u52a0\u4e0a<code>model.eval()<\/code>\uff0c\u5426\u5219\u7684\u8bdd\uff0c\u6709\u8f93\u5165\u6570\u636e\uff0c\u5373\u4f7f\u4e0d\u8bad\u7ec3\uff0c\u5b83\u4e5f\u4f1a\u6539\u53d8\u6743\u503c\u3002\u8fd9\u662fmodel\u4e2d\u542b\u6709BN\u5c42\u548cDropout\u6240\u5e26\u6765\u7684\u7684\u6027\u8d28\u3002<\/p>\n\n\n\n<p class=\"has-light-pink-background-color has-background\">\u5728\u505aone classification\u7684\u65f6\u5019\uff0c\u8bad\u7ec3\u96c6\u548c\u6d4b\u8bd5\u96c6\u7684\u6837\u672c\u5206\u5e03\u662f\u4e0d\u4e00\u6837\u7684\uff0c\u5c24\u5176\u9700\u8981\u6ce8\u610f\u8fd9\u4e00\u70b9\u3002<\/p>\n\n\n\n<h2>2 . model.eval()\u548cwith torch.no_grad()\u7684\u533a\u522b\uff1a<\/h2>\n\n\n\n<p><br>\u5728PyTorch\u4e2d\u8fdb\u884cvalidation\u65f6\uff0c\u4f1a\u4f7f\u7528model.eval()\u5207\u6362\u5230\u6d4b\u8bd5\u6a21\u5f0f\uff0c\u5728\u8be5\u6a21\u5f0f\u4e0b\uff0c<\/p>\n\n\n\n<p>\u4e3b\u8981\u7528\u4e8e\u901a\u77e5dropout\u5c42\u548cbatchnorm\u5c42\u5728train\u548cval\u6a21\u5f0f\u95f4\u5207\u6362<br>\u5728train\u6a21\u5f0f\u4e0b\uff0cdropout\u7f51\u7edc\u5c42\u4f1a\u6309\u7167\u8bbe\u5b9a\u7684\u53c2\u6570p\u8bbe\u7f6e\u4fdd\u7559\u6fc0\u6d3b\u5355\u5143\u7684\u6982\u7387\uff08\u4fdd\u7559\u6982\u7387=p); batchnorm\u5c42\u4f1a\u7ee7\u7eed\u8ba1\u7b97\u6570\u636e\u7684mean\u548cvar\u7b49\u53c2\u6570\u5e76\u66f4\u65b0\u3002<br>\u5728val\u6a21\u5f0f\u4e0b\uff0cdropout\u5c42\u4f1a\u8ba9\u6240\u6709\u7684\u6fc0\u6d3b\u5355\u5143\u90fd\u901a\u8fc7\uff0c\u800cbatchnorm\u5c42\u4f1a\u505c\u6b62\u8ba1\u7b97\u548c\u66f4\u65b0mean\u548cvar\uff0c\u76f4\u63a5\u4f7f\u7528\u5728\u8bad\u7ec3\u9636\u6bb5\u5df2\u7ecf\u5b66\u51fa\u7684mean\u548cvar\u503c\u3002<br>\u8be5\u6a21\u5f0f\u4e0d\u4f1a\u5f71\u54cd\u5404\u5c42\u7684gradient\u8ba1\u7b97\u884c\u4e3a\uff0c\u5373gradient\u8ba1\u7b97\u548c\u5b58\u50a8\u4e0etraining\u6a21\u5f0f\u4e00\u6837\uff0c\u53ea\u662f\u4e0d\u8fdb\u884c\u53cd\u4f20\uff08backprobagation\uff09<\/p>\n\n\n\n<p><br>with torch.no_grad()\u5219\u4e3b\u8981\u662f\u7528\u4e8e\u505c\u6b62autograd\u6a21\u5757\u7684\u5de5\u4f5c\uff0c\u4ee5\u8d77\u5230\u52a0\u901f\u548c\u8282\u7701\u663e\u5b58\u7684\u4f5c\u7528\uff0c\u5177\u4f53\u884c\u4e3a\u5c31\u662f\u505c\u6b62gradient\u8ba1\u7b97\uff0c\u4ece\u800c\u8282\u7701\u4e86GPU\u7b97\u529b\u548c\u663e\u5b58\uff0c\u4f46\u662f\u5e76\u4e0d\u4f1a\u5f71\u54cddropout\u548cbatchnorm\u5c42\u7684\u884c\u4e3a\u3002<\/p>\n\n\n\n<p><br>\u4f7f\u7528\u573a\u666f\uff1a<br>\u5982\u679c\u4e0d\u5728\u610f\u663e\u5b58\u5927\u5c0f\u548c\u8ba1\u7b97\u65f6\u95f4\u7684\u8bdd\uff0c\u4ec5\u4ec5\u4f7f\u7528model.eval()\u5df2\u8db3\u591f\u5f97\u5230\u6b63\u786e\u7684validation\u7684\u7ed3\u679c\uff1b\u800cwith torch.zero_grad()\u5219\u662f\u66f4\u8fdb\u4e00\u6b65\u52a0\u901f\u548c\u8282\u7701gpu\u7a7a\u95f4\uff08\u56e0\u4e3a\u4e0d\u7528\u8ba1\u7b97\u548c\u5b58\u50a8gradient\uff09\uff0c\u4ece\u800c\u53ef\u4ee5\u66f4\u5feb\u8ba1\u7b97\uff0c\u4e5f\u53ef\u4ee5\u8dd1\u66f4\u5927\u7684batch\u6765\u6d4b\u8bd5\u3002<\/p>\n\n\n\n<p><\/p>\n","protected":false},"excerpt":{"rendered":"<p>1\u3001model.eval() model.train()\u533a\u522b model.train()\u548cmodel.eval &hellip; <a href=\"http:\/\/139.9.1.231\/index.php\/2022\/08\/23\/model-evalith-torch-no_grad\/\" class=\"more-link\">\u7ee7\u7eed\u9605\u8bfb<span class=\"screen-reader-text\">Pytorch \u4e2d model.eval() model.train() \u548c with torch.no_grad() \u7684\u533a\u522b<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[11],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/6101"}],"collection":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/comments?post=6101"}],"version-history":[{"count":19,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/6101\/revisions"}],"predecessor-version":[{"id":10259,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/6101\/revisions\/10259"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=6101"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=6101"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=6101"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}