{"id":4840,"date":"2022-07-18T15:58:35","date_gmt":"2022-07-18T07:58:35","guid":{"rendered":"http:\/\/139.9.1.231\/?p=4840"},"modified":"2022-09-30T16:03:00","modified_gmt":"2022-09-30T08:03:00","slug":"fine-tuning","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2022\/07\/18\/fine-tuning\/","title":{"rendered":"MAE–transformer\u6a21\u578b\u9884\u8bad\u7ec3"},"content":{"rendered":"\n

\u5047\u8bbe\u6211\u4eec\u60f3\u4ece\u56fe\u50cf\u4e2d\u8bc6\u522b\u51fa\u4e0d\u540c\u79cd\u7c7b\u7684\u6905\u5b50\uff0c\u7136\u540e\u5c06\u8d2d\u4e70\u94fe\u63a5\u63a8\u8350\u7ed9\u7528\u6237\u3002\u4e00\u79cd\u53ef\u80fd\u7684\u65b9\u6cd5\u662f\u5148\u627e\u51fa100\u79cd\u5e38\u89c1\u7684\u6905\u5b50\uff0c\u4e3a\u6bcf\u79cd\u6905\u5b50\u62cd\u64441,000\u5f20\u4e0d\u540c\u89d2\u5ea6\u7684\u56fe\u50cf\uff0c\u7136\u540e\u5728\u6536\u96c6\u5230\u7684\u56fe\u50cf\u6570\u636e\u96c6\u4e0a\u8bad\u7ec3\u4e00\u4e2a\u5206\u7c7b\u6a21\u578b\u3002\u8fd9\u4e2a\u6905\u5b50\u6570\u636e\u96c6\u867d\u7136\u53ef\u80fd\u6bd4Fashion-MNIST\u6570\u636e\u96c6\u8981\u5e9e\u5927\uff0c\u4f46\u6837\u672c\u6570\u4ecd\u7136\u4e0d\u53caImageNet\u6570\u636e\u96c6\u4e2d\u6837\u672c\u6570\u7684\u5341\u5206\u4e4b\u4e00\u3002\u8fd9\u53ef\u80fd\u4f1a\u5bfc\u81f4\u9002\u7528\u4e8eImageNet\u6570\u636e\u96c6\u7684\u590d\u6742\u6a21\u578b\u5728\u8fd9\u4e2a\u6905\u5b50\u6570\u636e\u96c6\u4e0a\u8fc7\u62df\u5408\u3002\u540c\u65f6\uff0c\u56e0\u4e3a\u6570\u636e\u91cf\u6709\u9650\uff0c\u6700\u7ec8\u8bad\u7ec3\u5f97\u5230\u7684\u6a21\u578b\u7684\u7cbe\u5ea6\u4e5f\u53ef\u80fd\u8fbe\u4e0d\u5230\u5b9e\u7528\u7684\u8981\u6c42\u3002<\/p>\n\n\n\n

\u4e3a\u4e86\u5e94\u5bf9\u4e0a\u8ff0\u95ee\u9898\uff0c\u4e00\u4e2a\u663e\u800c\u6613\u89c1\u7684\u89e3\u51b3\u529e\u6cd5\u662f\u6536\u96c6\u66f4\u591a\u7684\u6570\u636e\u3002\u7136\u800c\uff0c\u6536\u96c6\u548c\u6807\u6ce8\u6570\u636e\u4f1a\u82b1\u8d39\u5927\u91cf\u7684\u65f6\u95f4\u548c\u8d44\u91d1\u3002\u4f8b\u5982\uff0c\u4e3a\u4e86\u6536\u96c6ImageNet\u6570\u636e\u96c6\uff0c\u7814\u7a76\u4eba\u5458\u82b1\u8d39\u4e86\u6570\u767e\u4e07\u7f8e\u5143\u7684\u7814\u7a76\u7ecf\u8d39\u3002\u867d\u7136\u76ee\u524d\u7684\u6570\u636e\u91c7\u96c6\u6210\u672c\u5df2\u964d\u4f4e\u4e86\u4e0d\u5c11\uff0c\u4f46\u5176\u6210\u672c\u4ecd\u7136\u4e0d\u53ef\u5ffd\u7565\u3002<\/p>\n\n\n\n

\u53e6\u5916\u4e00\u79cd\u89e3\u51b3\u529e\u6cd5\u662f\u5e94\u7528\u8fc1\u79fb\u5b66\u4e60\uff08transfer learning\uff09\uff0c\u5c06\u4ece\u6e90\u6570\u636e\u96c6\u5b66\u5230\u7684\u77e5\u8bc6\u8fc1\u79fb\u5230\u76ee\u6807\u6570\u636e\u96c6\u4e0a\u3002\u4f8b\u5982\uff0c\u867d\u7136ImageNet\u6570\u636e\u96c6\u7684\u56fe\u50cf\u5927\u591a\u8ddf\u6905\u5b50\u65e0\u5173\uff0c\u4f46\u5728\u8be5\u6570\u636e\u96c6\u4e0a\u8bad\u7ec3\u7684\u6a21\u578b\u53ef\u4ee5\u62bd\u53d6\u8f83\u901a\u7528\u7684\u56fe\u50cf\u7279\u5f81\uff0c\u4ece\u800c\u80fd\u591f\u5e2e\u52a9\u8bc6\u522b\u8fb9\u7f18\u3001\u7eb9\u7406\u3001\u5f62\u72b6\u548c\u7269\u4f53\u7ec4\u6210\u7b49\u3002\u8fd9\u4e9b\u7c7b\u4f3c\u7684\u7279\u5f81\u5bf9\u4e8e\u8bc6\u522b\u6905\u5b50\u4e5f\u53ef\u80fd\u540c\u6837\u6709\u6548\u3002<\/p>\n\n\n\n

\u672c\u8282\u6211\u4eec\u4ecb\u7ecd\u8fc1\u79fb\u5b66\u4e60\u4e2d\u7684\u4e00\u79cd\u5e38\u7528\u6280\u672f\uff1a\u5fae\u8c03\uff08fine tuning\uff09\u3002\u5982\u56fe9.1\u6240\u793a\uff0c\u5fae\u8c03\u7531\u4ee5\u4e0b4\u6b65\u6784\u6210\u3002<\/p>\n\n\n\n

  1. \u5728\u6e90\u6570\u636e\u96c6\uff08\u5982ImageNet\u6570\u636e\u96c6\uff09\u4e0a\u9884\u8bad\u7ec3\u4e00\u4e2a\u795e\u7ecf\u7f51\u7edc\u6a21\u578b\uff0c\u5373\u6e90\u6a21\u578b\u3002<\/li>
  2. \u521b\u5efa\u4e00\u4e2a\u65b0\u7684\u795e\u7ecf\u7f51\u7edc\u6a21\u578b\uff0c\u5373\u76ee\u6807\u6a21\u578b\u3002\u5b83\u590d\u5236\u4e86\u6e90\u6a21\u578b\u4e0a\u9664\u4e86\u8f93\u51fa\u5c42\u5916\u7684\u6240\u6709\u6a21\u578b\u8bbe\u8ba1\u53ca\u5176\u53c2\u6570\u3002\u6211\u4eec\u5047\u8bbe\u8fd9\u4e9b\u6a21\u578b\u53c2\u6570\u5305\u542b\u4e86\u6e90\u6570\u636e\u96c6\u4e0a\u5b66\u4e60\u5230\u7684\u77e5\u8bc6\uff0c\u4e14\u8fd9\u4e9b\u77e5\u8bc6\u540c\u6837\u9002\u7528\u4e8e\u76ee\u6807\u6570\u636e\u96c6\u3002\u6211\u4eec\u8fd8\u5047\u8bbe\u6e90\u6a21\u578b\u7684\u8f93\u51fa\u5c42\u8ddf\u6e90\u6570\u636e\u96c6\u7684\u6807\u7b7e\u7d27\u5bc6\u76f8\u5173\uff0c\u56e0\u6b64\u5728\u76ee\u6807\u6a21\u578b\u4e2d\u4e0d\u4e88\u91c7\u7528\u3002<\/li>
  3. \u4e3a\u76ee\u6807\u6a21\u578b\u6dfb\u52a0\u4e00\u4e2a\u8f93\u51fa\u5927\u5c0f\u4e3a\u76ee\u6807\u6570\u636e\u96c6\u7c7b\u522b\u4e2a\u6570\u7684\u8f93\u51fa\u5c42\uff0c\u5e76\u968f\u673a\u521d\u59cb\u5316\u8be5\u5c42\u7684\u6a21\u578b\u53c2\u6570\u3002<\/li>
  4. \u5728\u76ee\u6807\u6570\u636e\u96c6\uff08\u5982\u6905\u5b50\u6570\u636e\u96c6\uff09\u4e0a\u8bad\u7ec3\u76ee\u6807\u6a21\u578b\u3002\u6211\u4eec\u5c06\u4ece\u5934\u8bad\u7ec3\u8f93\u51fa\u5c42\uff0c\u800c\u5176\u4f59\u5c42\u7684\u53c2\u6570\u90fd\u662f\u57fa\u4e8e\u6e90\u6a21\u578b\u7684\u53c2\u6570\u5fae\u8c03\u5f97\u5230\u7684\u3002<\/li><\/ol>\n\n\n\n
    \"\"\/<\/figure>\n\n\n\n

    \u5f53\u76ee\u6807\u6570\u636e\u96c6\u8fdc\u5c0f\u4e8e\u6e90\u6570\u636e\u96c6\u65f6\uff0c\u5fae\u8c03\u6709\u52a9\u4e8e\u63d0\u5347\u6a21\u578b\u7684\u6cdb\u5316\u80fd\u529b\u3002<\/p>\n\n\n\n

    \u4ee3\u7801\u5b9e\u73b0\u5fae\u8c03\uff1a<\/p>\n\n\n\n

    pretrained_net = models.resnet18(pretrained=True)\npretrained_net.load_state_dict(torch.load('\/home\/kesci\/input\/resnet185352\/resnet18-5c106cde.pth'))<\/code><\/pre>\n\n\n\n
    \u4e0b\u9762\u6253\u5370\u6e90\u6a21\u578b\u7684\u6210\u5458\u53d8\u91cffc<\/code>\u3002\u4f5c\u4e3a\u4e00\u4e2a\u5168\u8fde\u63a5\u5c42\uff0c\u5b83\u5c06ResNet\u6700\u7ec8\u7684\u5168\u5c40\u5e73\u5747\u6c60\u5316\u5c42\u8f93\u51fa\u53d8\u6362\u6210ImageNet\u6570\u636e\u96c6\u4e0a1000\u7c7b\u7684\u8f93\u51fa\u3002<\/p>\n\n\n\n
    print(pretrained_net.fc)<\/code><\/pre>\n\n\n\n
    \u8f93\u51fa\uff1aLinear(in_features=512, out_features=1000, bias=True)<\/p>\n\n\n\n
    \u53ef\u89c1\u6b64\u65f6pretrained_net<\/code>\u6700\u540e\u7684\u8f93\u51fa\u4e2a\u6570\u7b49\u4e8e\u76ee\u6807\u6570\u636e\u96c6\u7684\u7c7b\u522b\u65701000\u3002\u6240\u4ee5\u6211\u4eec\u5e94\u8be5\u5c06\u6700\u540e\u7684fc<\/code>\u6210\u4fee\u6539\u6211\u4eec\u9700\u8981\u7684\u8f93\u51fa\u7c7b\u522b\u6570:<\/p>\n\n\n\n
    pretrained_net.fc = nn.Linear(512, 2)\nprint(pretrained_net.fc)<\/code><\/pre>\n\n\n\n
    \u6b64\u65f6\uff0cpretrained_net<\/code>\u7684fc<\/code>\u5c42\u5c31\u88ab\u968f\u673a\u521d\u59cb\u5316\u4e86\uff0c\u4f46\u662f\u5176\u4ed6\u5c42\u4f9d\u7136\u4fdd\u5b58\u7740\u9884\u8bad\u7ec3\u5f97\u5230\u7684\u53c2\u6570\u3002\u7531\u4e8e\u662f\u5728\u5f88\u5927\u7684ImageNet\u6570\u636e\u96c6\u4e0a\u9884\u8bad\u7ec3\u7684\uff0c\u6240\u4ee5\u53c2\u6570\u5df2\u7ecf\u8db3\u591f\u597d\uff0c\u56e0\u6b64\u4e00\u822c\u53ea\u9700\u4f7f\u7528\u8f83\u5c0f\u7684\u5b66\u4e60\u7387\u6765\u5fae\u8c03\u8fd9\u4e9b\u53c2\u6570\uff0c\u800cfc<\/code>\u4e2d\u7684\u968f\u673a\u521d\u59cb\u5316\u53c2\u6570\u4e00\u822c\u9700\u8981\u66f4\u5927\u7684\u5b66\u4e60\u7387\u4ece\u5934\u8bad\u7ec3\u3002PyTorch\u53ef\u4ee5\u65b9\u4fbf\u7684\u5bf9\u6a21\u578b\u7684\u4e0d\u540c\u90e8\u5206\u8bbe\u7f6e\u4e0d\u540c\u7684\u5b66\u4e60\u53c2\u6570\uff0c\u6211\u4eec\u5728\u4e0b\u9762\u4ee3\u7801\u4e2d\u5c06fc<\/code>\u7684\u5b66\u4e60\u7387\u8bbe\u4e3a\u5df2\u7ecf\u9884\u8bad\u7ec3\u8fc7\u7684\u90e8\u5206\u768410\u500d\u3002<\/p>\n\n\n\n
    output_params = list(map(id, pretrained_net.fc.parameters()))\nfeature_params = filter(lambda p: id(p) not in output_params, pretrained_net.parameters())\n\nlr = 0.01\noptimizer = optim.SGD([{'params': feature_params},\n                       {'params': pretrained_net.fc.parameters(), 'lr': lr * 10}],\n                       lr=lr, weight_decay=0.001)<\/code><\/pre>\n\n\n\n
    \u8bb0\u5f55\uff1a \u5728MAE\u7684\u5fae\u8c03\u8bad\u7ec3\u4e2d\uff0c\u63d0\u4f9b\u4e86\u4e24\u79cd \u5fae\u8c03<\/strong>\uff1a<\/p>\n\n\n\n
    1. Linear probing:<\/strong> \u9501\u6b7btransformer\u7684\u53c2\u6570\uff0c\u53ea\u8bad\u7ec3CIFAR10\u7684\u90a3\u4e2aLinear\u5c42\u3002<\/li>
    2. Fine-tuning:<\/strong> \u63a5\u7740\u8bad\u7ec3transformer\u7684\u53c2\u6570\uff0c\u540c\u65f6\u4e5f\u8bad\u7ec3CIFAR10\u7684\u90a3\u4e2aLinear\u3002<\/li><\/ol>\n\n\n\n
      \u8bba\u6587\u505a\u4e86MAE\u5404\u4e2a\u90e8\u5206\u7684\u4e0d\u540c\u8bbe\u7f6e\u5bf9\u6bd4\u5b9e\u9a8c\uff0c\u8fd9\u4e9b\u5b9e\u9a8c\u80fd\u591f\u63ed\u793aMAE\u66f4\u591a\u7684\u7279\u6027\u3002\u9996\u5148\u662fmasking ratio<\/strong>\uff0c\u4ece\u4e0b\u56fe\u53ef\u4ee5\u770b\u5230\uff0c\u6700\u4f18\u7684\u8bbe\u7f6e\u662f75%\u7684masking ratio\uff0c\u6b64\u65f6linear probing\u548cfinetune\u6548\u679c\u6700\u597d\uff0c\u8fd9\u6bd4\u4e4b\u524d\u7684\u7814\u7a76\u8981\u9ad8\u5f88\u591a\uff0c\u6bd4\u5982BEiT\u7684masking ratio\u662f40%\u3002\u53e6\u5916\u4e5f\u53ef\u4ee5\u770b\u5230linear probing\u548cfinetune\u7684\u8868\u73b0\u4e0d\u4e00\u6837\uff0clinear probing\u6548\u679c\u968f\u7740masking ratio\u7684\u589e\u52a0\u9010\u6e10\u63d0\u9ad8\u76f4\u81f3\u4e00\u4e2a\u5cf0\u503c\u540e\u51fa\u73b0\u4e0b\u964d\uff0c\u800cfinetune\u6548\u679c\u5728\u4e0d\u540cmaking ratio\u4e0b\u5dee\u5f02\u5c0f\uff0cmasking ratio\u572840%~80%\u8303\u56f4\u5185\u5747\u80fd\u8868\u73b0\u8f83\u597d\u3002 \u200b<\/p>\n\n\n\n
      \"preview\"\/<\/figure>\n","protected":false},"excerpt":{"rendered":"
      \u5047\u8bbe\u6211\u4eec\u60f3\u4ece\u56fe\u50cf\u4e2d\u8bc6\u522b\u51fa\u4e0d\u540c\u79cd\u7c7b\u7684\u6905\u5b50\uff0c\u7136\u540e\u5c06\u8d2d\u4e70\u94fe\u63a5\u63a8\u8350\u7ed9\u7528\u6237\u3002\u4e00\u79cd\u53ef\u80fd\u7684\u65b9\u6cd5\u662f\u5148\u627e\u51fa100\u79cd\u5e38\u89c1\u7684\u6905\u5b50\uff0c\u4e3a … \u7ee7\u7eed\u9605\u8bfbMAE–transformer\u6a21\u578b\u9884\u8bad\u7ec3<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[11,4,9],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/4840"}],"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=4840"}],"version-history":[{"count":11,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/4840\/revisions"}],"predecessor-version":[{"id":8701,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/4840\/revisions\/8701"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=4840"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=4840"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=4840"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}