{"id":10703,"date":"2022-11-28T17:10:10","date_gmt":"2022-11-28T09:10:10","guid":{"rendered":"http:\/\/139.9.1.231\/?p=10703"},"modified":"2022-12-27T21:37:42","modified_gmt":"2022-12-27T13:37:42","slug":"visual-prompt-tuning","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2022\/11\/28\/visual-prompt-tuning\/","title":{"rendered":"Visual Prompt–\u89c6\u89c9\u6a21\u677f"},"content":{"rendered":"\n

Visual Prompt Tuning<\/a><\/p>https:\/\/github.com\/KMnP\/vpt<\/cite><\/blockquote><\/figure>\n\n\n\n

Exploring Visual Prompts for Adapting Large-Scale Models<\/a><\/p>https:\/\/github.com\/hjbahng\/visual_prompting<\/cite><\/blockquote><\/figure>\n\n\n\n\n\n

Visual Prompt Tuning<\/h2>\n\n\n\n
\"\"<\/figure><\/div>\n\n\n\n

\u6700\u8fd1NLP\u9886\u57df\u63d0\u51fa\u4e86Prompt\u8303\u65b0\u5f0f\uff0c\u4f01\u56fe\u9769\u65b0\u539f\u5148\u7684Fine-tuning\u65b9\u6cd5\uff0c\u800c\u5728CV\u9886\u57df\u4e2d\uff0cPrompt\u5176\u5b9e\u53ef\u4ee5\u7406\u89e3\u4e3a\u56fe\u50cflabel\u7684\u8bbe\u8ba1\uff0c\u4ece\u8fd9\u4e2a\u89d2\u5ea6\u770b\uff0cPrompt(\u9884\u6d4b\u6587\u672c\u4e2dmask\u7684\u5b57\u7b26\uff0c\u7c7b\u4f3c\u5b8c\u5f62\u586b\u7a7a)\u5176\u5b9e\u662f\u4ecb\u4e8eImage caption(\u8fed\u4ee3\u9884\u6d4b\u51fa\u6bcf\u4e00\u4e2a\u5b57\u7b26)\u548cone-hot label(one-hot\u53ef\u4ee5\u8ba4\u4e3a\u662fprompt\u7684\u7279\u4f8b\uff0c\u5355\u5b57\u7b26\u901a\u8fc7text encoder\u6210one-hot)\u4e4b\u95f4\u7684\u4efb\u52a1\u3002<\/strong>\u6700\u8fd1\u5728Visual-Language Model(\u7f29\u5199VLM)\u4efb\u52a1\u4e2d\uff0cprompt\u5f00\u59cb\u5c55\u73b0\u51fa\u5f3a\u5927\u7684\u80fd\u529b.<\/p>\n\n\n\n

Fine-tuning\u4e2d\uff1a\u662f\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u201c\u8fc1\u5c31\u201c\u5404\u79cd\u4e0b\u6e38\u4efb\u52a1\u3002<\/strong>\u5177\u4f53\u4f53\u73b0\u5c31\u662f\u4e0a\u9762\u63d0\u5230\u7684\u901a\u8fc7\u5f15\u5165\u5404\u79cd\u8f85\u52a9\u4efb\u52a1loss\uff0c\u5c06\u5176\u6dfb\u52a0\u5230\u9884\u8bad\u7ec3\u6a21\u578b\u4e2d\uff0c\u7136\u540e\u7ee7\u7eedpre-training\uff0c\u4ee5\u4fbf\u8ba9\u5176\u66f4\u52a0\u9002\u914d\u4e0b\u6e38\u4efb\u52a1\u3002\u603b\u4e4b\uff0c\u8fd9\u4e2a\u8fc7\u7a0b\u4e2d\uff0c\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u505a\u51fa\u4e86\u66f4\u591a\u7684\u727a\u7272\u3002<\/p>\n\n\n\n

Prompting\u4e2d\uff0c\u662f\u5404\u79cd\u4e0b\u6e38\u4efb\u52a1\u201c\u8fc1\u5c31\u201c\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u3002<\/strong>\u5177\u4f53\u4f53\u73b0\u4e5f\u662f\u4e0a\u9762\u4ecb\u7ecd\u7684\uff0c\u6211\u4eec\u9700\u8981\u5bf9\u4e0d\u540c\u4efb\u52a1\u8fdb\u884c\u91cd\u6784\uff0c\u4f7f\u5f97\u5b83\u8fbe\u5230\u9002\u914d\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u7684\u6548\u679c\u3002\u603b\u4e4b\uff0c\u8fd9\u4e2a\u8fc7\u7a0b\u4e2d\uff0c\u662f\u4e0b\u6e38\u4efb\u52a1\u505a\u51fa\u4e86\u66f4\u591a\u7684\u727a\u7272\u3002<\/p>\n\n\n\n

Abstract<\/h3>\n\n\n\n

\u76ee\u524d\u8c03\u6574\u9884\u8bad\u7ec3\u6a21\u578b\u7684\u65b9\u6cd5\u662ffull fine-tuning\uff0c\u5373\u5b8c\u5168\u5fae\u8c03\u3002\u672c\u6587\u4ecb\u7ecdVisual Prompt Tuning\uff08VPT\uff09\u4f5c\u4e3a\u4e00\u79cd\u6709\u6548\u7684\u7528\u4e8e\u5927\u89c4\u6a21Transformer\u7684\u89c6\u89c9\u5fae\u8c03\u3002\u5b83\u53ea\u9700\u8981\u5728\u8f93\u5165\u7a7a\u95f4\u5f15\u5165\u5c11\u91cf\uff08\u4e0d\u52301%\u7684\u6a21\u578b\u53c2\u6570\uff09\u7684\u53ef\u8bad\u7ec3\u53c2\u6570\uff0c\u540c\u65f6\u51bb\u7ed3backbone\u3002\u4f1a\u53d1\u73b0\u5728\u5f88\u591a\u60c5\u51b5\u4e0b\uff0c\u4f18\u4e8e\u5b8c\u5168\u5fae\u8c03\u3002<\/p>\n\n\n\n

Introduction<\/h3>\n\n\n\n

\u5bf9\u4e8e\u5927\u89c4\u6a21\u6a21\u578b\u9002\u5e94\u4e0b\u6e38\u4efb\u52a1\u65f6\uff0c\u901a\u5e38\u7684\u7b56\u7565\u662f\u8fdb\u884c\u7aef\u5230\u7aef\u7684\u5168\u9762\u5fae\u8c03\uff0c\u7136\u800c\u8fd9\u79cd\u7b56\u7565\u9700\u8981\u4e3a\u6bcf\u4e2a\u4eba\u7269\u5b58\u50a8\u90e8\u7f72\u5355\u72ec\u7684\u4e3b\u5e72\u53c2\u6570\uff0c\u4ee3\u4ef7\u6bd4\u8f83\u9ad8\uff0c\u6bd5\u7adf\u73b0\u5728\u7684Transformer\u4f53\u7cfb\u7ed3\u6784\u6bd4\u8f83\u5927\u3002<\/p>\n\n\n\n

\u4e00\u79cd\u7b80\u5355\u7684\u65b9\u6cd5\u662f\u4f7f\u7528\u5df2\u7ecf\u5b8c\u5584\u7684\u5176\u4ed6\u7b56\u7565\uff0c\u5982\u4e0b\u56fe\uff08a\uff09\uff1a\u4ec5\u5fae\u8c03\u53c2\u6570\u7684\u5b50\u96c6\uff0c\u5982\u5206\u7c7b\u5668\u5934\u90e8\u6216\u8005\u504f\u5dee\u9879\u3002\u4e4b\u524d\u7684\u7814\u7a76\u8fd8\u4f1a\u8bd5\u7740\u5411\u4e3b\u5e72\u6dfb\u52a0\u989d\u5916\u7684\u6b8b\u5dee\u7ed3\u6784\u6216\u8005adapter\uff0c\u53ef\u4ee5\u5bf9Transformer\u5b9e\u65bd\u7c7b\u4f3c\u7684\u7b56\u7565\u3002\u7136\u800c\uff0c\u8fd9\u4e9b\u7b56\u7565\u4f1a\u5728\u51c6\u786e\u5ea6\u4e0a\u6267\u884c\u5b8c\u5168\u5fae\u8c03\u3002<\/p>\n\n\n\n

\u4f5c\u8005\u8bd5\u56fe\u63a2\u7d22\u4e00\u79cd\u4e0d\u540c\u7684\u65b9\u6cd5\uff1a\u5e76\u4e0d\u901a\u8fc7\u6539\u53d8\u6216\u8005\u5fae\u8c03\u9884\u8bad\u7ec3\u597d\u7684Transformer\u672c\u8eab\uff0c\u800c\u662f\u4fee\u6539\u5176\u8f93\u5165\u3002\u5982\u4e0b\u56fe\uff08b\uff09\u6240\u793a\uff1a\u5c06\u5c11\u91cf\u7279\u5b9a\u4efb\u52a1\u7684\u53ef\u5b66\u4e60\u53c2\u6570\u5f15\u5165\u8f93\u5165\u7a7a\u95f4\uff0c\u540c\u65f6\u5728\u4e0b\u6e38\u8bad\u7ec3\u671f\u95f4\u51bb\u7ed3backbone\u3002\u5b9e\u8df5\u4e2d\uff0c\u8fd9\u4e9b\u9644\u52a0\u53c2\u6570\u53ea\u662f\u9884\u5148\u52a0\u5165\u5230Transformer\u6bcf\u5c42\u8f93\u5165\u5e8f\u5217\u4e2d\uff0c\u5e76\u5728\u5fae\u8c03\u65f6\u548c\u7ebf\u6027\u5934\u4e00\u8d77\u5b66\u4e60\u3002<\/p>\n\n\n\n

\u5728\u9884\u8bad\u7ec3\u4e3b\u5e72\u7528ViT\u768424\u4e2a\u8de8\u57df\u7684\u4e0b\u6e38\u4efb\u52a1\u4e2d\uff0cVPT\u4f18\u4e8e\u4e86\u5176\u4ed6\u8fc1\u79fb\u5b66\u4e60\u7684baseline\uff0c\u670920\u4e2a\u8d85\u8fc7\u4e86\u5b8c\u5168\u5fae\u8c03\uff0c\u540c\u65f6\u4fdd\u6301\u4e86\u4e3a\u6bcf\u4e2a\u5355\u72ec\u4efb\u52a1\u50a8\u5b58\u8f83\u5c11\u53c2\u6570\u7684\u4f18\u52bf\u3002\uff08NLP\u4efb\u52a1\u4e2d\uff0cprompt tuning\u65e8\u5728\u67d0\u4e9b\u60c5\u51b5\u4e0b\u624d\u5339\u914d\u5b8c\u5168\u5fae\u8c03\u7684\u6027\u80fd\uff09\u3002\u5982\u4e0b\u56fe\uff08c\uff09\u6240\u793a\uff0cVPT\u5728\u5730\u6570\u636e\u533a\u5c24\u5176\u6709\u6548\uff0c\u7ed3\u679c\u4e5f\u8fdb\u4e00\u6b65\u8868\u660e\uff0cVPT\u662f\u9002\u5e94\u4e0d\u65ad\u589e\u957f\u7684\u89c6\u89c9\u4e3b\u5e72\u7684\u6700\u6709\u6548\u65b9\u6cd5\u4e4b\u4e00\u3002<\/p>\n\n\n\n

\"\"
Visual-Prompt Tuning (VPT) vs . other transfer learning methods. (a) Current
transfer learning protocols are grouped based on the tuning scope: Full fine-tuning,
Head-oriented, and Backbone-oriented approaches. (b) VPT instead adds extra pa-
rameters in the input space. (c) Performance of different methods on a wide range
of downstream classification tasks adapting a pre-trained ViT-B backbone, with mean
and standard deviation annotated. VPT outperforms Full fine-tuning 20 out of 24 cases
while using less than 1% of all model parameters<\/figcaption><\/figure>\n\n\n\n

Related Work<\/h3>\n\n\n\n

\u8fc1\u79fb\u5b66\u4e60\u4e24\u79cd\u4ee3\u8868\u6027\u65b9\u6cd5\uff1aAdapter\u5728\u6bcf\u4e2aTransformer\u5c42\u540e\u63d2\u5165\u4e00\u4e2a\u989d\u5916\u7684\u5c0f\u6a21\u5757\u3002\u901a\u5e38\u5305\u542b\u4e00\u4e2a\u7ebf\u6027\u5411\u4e0b\u5934\u50cf\u3001\u7ebf\u6027\u5411\u4e0a\u6295\u5f71\u53ca\u4e00\u4e2a\u6b8b\u5dee\u8fde\u63a5\u3002BitFit\u662f\u5728\u5fae\u8c03\u7f51\u7edc\u65f6\u66f4\u65b0\u504f\u7f6e\u9879\u5e76\u51bb\u7ed3\u5176\u4f59backbone\u53c2\u6570\u3002\u8fd9\u4e9b\u65b9\u6cd5\u5728NLP\u5df2\u7ecf\u6210\u719f\u8fd0\u7528\u3002\u4f5c\u8005\u5219\u8fdb\u4e00\u6b65\u5b9e\u9a8c\u8868\u660e\u4e86VPT\u5728\u89c6\u89c9\u4efb\u52a1\u7684Transformer\u7684\u6a21\u578b\u8c03\u6574\u4e0a\u6027\u80fd\u66f4\u52a0\u826f\u597d\u3002<\/p>\n\n\n\n

Prompt\u6700\u521d\u6307\u7684\u662f\u5728\u8f93\u5165\u6587\u672c\u4e2d\u9884\u7f16\u8bed\u8a00\u6c41\u4e86\uff0c\u4ee5\u4fbf\u9884\u8bad\u7ec3\u597d\u7684LM(Language Model)\u80fd\u591f\u201c\u7406\u89e3\u201d\u4efb\u52a1\u3002\u6700\u8fd1\u7684\u5de5\u4f5c\u5219\u662f\u5c06prompt\u89c6\u4e3a\u4efb\u52a1\u7279\u5b9a\u7684\u8fde\u7eed\u5411\u91cf\uff0c\u5e76\u5728\u5fae\u8c03\u8fc7\u7a0b\u4e2d\u901a\u8fc7\u68af\u5ea6\u76f4\u63a5\u4f18\u5316\uff0c\u5373Prompt Tuning\u3002prompting\u4f9d\u7136\u5c40\u9650\u4e8e\u6587\u672c\u7f16\u7801\u5668\u7684\u8f93\u5165\u3002\u4f5c\u8005\u662f\u7b2c\u4e00\u4e2a\u89e3\u51b3\uff08\u540c\u6837\u7684\u65b9\u6cd5\u80fd\u6210\u529f\u7684\u5e94\u7528\u5230\u89c6\u89c9\u4e3b\u5e72\uff09\u5e76\u7814\u7a76\u89c6\u89c9prompt\u7684\u666e\u904d\u6027\u548c\u53ef\u884c\u6027\u7684\u5de5\u4f5c\u3002<\/p>\n\n\n\n

Approach<\/h3>\n\n\n\n

\u6574\u4f53\u6846\u67b6\u56fe\u5982\u4e0b\u56fe\u6240\u793a\uff1a<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

Visual-Prompt Tuning\uff1a<\/p>\n\n\n\n

\u7ed9\u5b9a\u4e00\u4e2a\u9884\u5148\u8bad\u7ec3\u597d\u7684Transformer\uff0c\u5728Embed\u5c42\u540e\u7684\u8f93\u5165\u7a7a\u95f4\u5f15\u5165\u4e00\u7ec4d\u7ef4\u7684p\u8fde\u7eedembedding\u3002\u5728\u5fae\u8c03\u8fc7\u7a0b\u4e2d\uff0c\u53ea\u6709prompt\u4f1a\u88ab\u66f4\u65b0\uff0c\u4e3b\u5e72\u5c06\u4f1a\u51bb\u7ed3\uff0c\u6839\u636e\u52a0\u5165prompt\u7684\u5c42\u6570\u91cf\u5206\u4e3a\u6d45VPT\u548c\u6df1VPT\u3002<\/p>\n\n\n\n

\u6d45VPT :<\/p>\n\n\n\n

Prompt\u4ec5\u63d2\u5165\u7b2c\u4e00\u5c42\u3002\u6bcf\u4e00\u4e2aprompt token\u90fd\u662f\u4e00\u4e2a\u53ef\u5b66\u4e60\u7684d\u7ef4\u53c2\u6570\u3002\u96c6\u5408\u548c\u6d45VPT\u8868\u793a\u5982\u4e0b\uff1a<\/p>\n\n\n\n

\"\"\/<\/figure><\/div>\n\n\n\n
\"\"<\/figure>\n\n\n\n

VPT-Deep: Prompt\u88ab\u63d2\u5165\u6bcf\u4e00\u5c42\u7684\u8f93\u5165\u63a7\u4ef6\u3002\u96c6\u5408\u548c\u6df1VPT\u8868\u793a\u5982\u4e0b\uff1a<\/p>\n\n\n\n

\"\"\/<\/figure>\n\n\n\n
\"\"\/<\/figure>\n\n\n\n

VPT\u5bf9\u4e8e\u591a\u4e2a\u4e0b\u6e38\u4efb\u52a1\u90fd\u662f\u6709\u5e2e\u52a9\u7684\uff0c\u53ea\u9700\u8981\u4e3a\u6bcf\u4e2a\u4efb\u52a1\u5b58\u50a8\u5b66\u4e60\u5230\u7684prompt\u548c\u5206\u7c7b\u5934\uff0c\u91cd\u65b0\u4f7f\u7528\u9884\u8bad\u7ec3\u7684Transformer\uff0c\u4ece\u800c\u663e\u7740\u964d\u4f4e\u5b58\u50a8\u6210\u672c\u3002<\/p>\n\n\n\n

Experiments:<\/h3>\n\n\n\n
\"\"<\/figure>\n\n\n\n

\u4e0a\u56fe\u662f\u5173\u4e8eprompt\u7684\u4f4d\u7f6e\uff0c\u672c\u6587\u63d0\u51fa\u7684\u662fprepend\uff0c\u4e0e\u76f4\u63a5\u5728embedding\u4e0a\u6dfb\u52a0\u5bf9\u6bd4\u6548\u679c\u66f4\u597d\u3002\u9664\u6b64\u4e4b\u5916\uff0c\u4f5c\u4e3a\u524d\u7f6e\u50cf\u7d20\u6216\u8005concat\u901a\u9053\u7684\u6548\u679c\u4e5f\u90fd\u5728\u4e0b\u964d\u3002<\/p>\n\n\n\n

\u4e0b\u56fe\u5219\u662f\u5bf9prompt\u957f\u5ea6\u3001\u6df1\u5ea6\u7684\u6d88\u878d\u5b9e\u9a8c\uff1a<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

\u6700\u4f73\u63d0\u793a\u957f\u5ea6\u56e0\u4efb\u52a1\u800c\u5f02\uff0c\u5373\u4f7f\u53ea\u6709\u4e00\u4e2aprompt\uff0c\u6df1VPT\u7684\u6548\u679c\u4ecd\u663e\u7740\u4f18\u4e8e\u53e6\u5916\u4e24\u79cd\u65b9\u6cd5\u3002<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

\u4e0b\u56fe\u4e3a\u6700\u7ec8\u8f93\u51fa\u7684\u6d88\u878d\u5b9e\u9a8c\uff1a<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

\u8865\u5145\uff1a<\/h3>\n\n\n\n

CoOp<\/strong>\uff1a<\/h4>\n\n\n\n

CoOp\u660e\u663e\u662f\u53d7\u5230\u4e86AutoPrompt\u7684\u542f\u53d1\uff0c\u5e76\u4e14CoOp\u53d1\u73b0CLIP\u5b9e\u9645\u4e0a\u5c31\u662fprompt\u5728visual-language model\u4e2d\u7684\u4e00\u4e2a\u5e94\u7528\uff0c\u4e8e\u662fCoOp\u5728CLIP\u7684\u57fa\u7840\u4e0a\u8fdb\u4e00\u6b65\u8fdb\u884c\u6539\u8fdb\u3002<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

CoOp\u5148\u5728\u56db\u4e2a\u6570\u636e\u96c6\u4e0a\u505a\u5b9e\u9a8c\uff0c\u53d1\u73b0\u66f4\u5408\u7406\u7684prompt\u80fd\u591f\u5927\u5e45\u5ea6\u7684\u63d0\u5347\u5206\u7c7b\u7cbe\u5ea6\u5c24\u5176\u662f\u4f7f\u7528\u4e86\u672c\u6587\u63d0\u51fa\u7684CoOp\u4e4b\u540e\uff0c\u6700\u7ec8\u7684\u5206\u7c7b\u7cbe\u5ea6\u8fdc\u8d85CLIP\u4eba\u4e3a\u8bbe\u8ba1\u7684prompt\u3002<\/p>\n\n\n\n

\"\"\/<\/figure>\n\n\n\n

\u548cCLIP\u7684\u4e3b\u8981\u4e0d\u540c\u4e4b\u5904\u5728\u4e8e\uff0cCoOp\u5728CLIP\u7684\u7b2c\u4e8c\u4e2a\u9636\u6bb5\u4e2d\u5f15\u5165\u4e86context optimization\u3002\u5177\u4f53\u7684\uff0cCoOp\u5c06prompt\u8bbe\u8ba1\u4e3a\uff1a<\/p>\n\n\n\n

t=[V]1[ V]2\u2026[V]M[CLASS]<\/p>\n\n\n\n

\u5176\u4e2d\u6bcf\u4e2a[V]M\u5411\u91cf\u8ddfword embedding\u7684\u7ef4\u5ea6\u76f8\u540c\uff0c\u53ef\u4ee5\u7406\u89e3\u4e3a\u53ef\u5b66\u4e60\u7684context\uff0c\u5e76\u4e14\u6240\u6709\u7c7b\u522b\u5bf9\u5e94\u7684context\u5171\u4eab\u53c2\u6570\u3002<\/strong><\/p>\n\n\n\n

Exploring Visual Prompts for Adapting Large-Scale Models<\/a><\/p><\/h2>\n\n\n\n

\u8fd9\u7bc7\u6587\u7ae0\u53c2\u8003\u4e86Ian Goodfellow\u7b49\u4eba\uff082018\uff09\u5bf9\u6297\u6837\u672c\u4e2d\u7684\u5bf9\u6297\u91cd\u7f16\u7a0b\u601d\u60f3<\/a>\uff0c \u5bf9\u6297\u91cd\u7f16\u7a0b\u7684\u76ee\u6807\u662f\u6211\u4f7f\u7528\u4e00\u4e2a\u4efb\u52a1A\u7684\u7f51\u7edc\uff08\u65e0\u9700\u91cd\u65b0\u8bad\u7ec3\u8be5\u7f51\u7edc\uff09\u6765\u505a\u4efb\u52a1B\u3002\u6211\u4eec\u8bbe\u4e00\u4e2a\u7ecf\u8fc7\u8bad\u7ec3\u7684\u6a21\u578b\uff0c\u5176\u539f\u672c\u7684\u4efb\u52a1\u662f\uff0c\u7ed9\u5b9a\u8f93\u5165x\uff0c\u4f1a\u5f97\u5230\u8f93\u51faf(x)\u73b0\u5728\u6709\u4e00\u4e2a\u653b\u51fb\u8005\uff0c\u5176\u5bf9\u6297\u4efb\u52a1\u662f\u5bf9\u4e8e\u7ed9\u5b9a\u7684\u8f93\u5165x~\uff0c\u4f1a\u5f97\u5230\u8f93\u51fag(x~),\u8fd9\u91ccx~\u548cx\u4e0d\u4e00\u5b9a\u9700\u8981\u662f\u540c\u57df\u7684\u3002\u8fd9\u770b\u8d77\u6765\u662f\u4e0d\u53ef\u884c\u7684\uff0c\u4f46\u662f\u653b\u51fb\u8005\u901a\u8fc7\u5b66\u4e60\u5bf9\u6297\u91cd\u7f16\u7a0b\u51fd\u6570hf(.;\u03b8)\u548chg(.;\u03b8)\u6765\u5b9e\u73b0\u8fd9\u4e24\u4e2a\u4efb\u52a1\u4e4b\u95f4\u7684\u6620\u5c04\u3002hf(.;\u03b8)\u7528\u4e8e\u5c06\u8f93\u5165\u4ecex~\u6240\u5728\u7684\u57df\u8f6c\u6362\u5230x\u6240\u5728\u7684\u57df\uff0c\u4e5f\u5c31\u662f\u8bf4\uff0c\u7ecf\u8fc7hf\u7684\u5904\u7406\u540e\u5f97\u5230\u7684hf(x~;\u03b8)\u5bf9\u4e8ef\u800c\u8a00\u662f\u6709\u6548\u7684\u8f93\u5165\uff1b\u800chg\u5219\u5c06f\u7684\u8f93\u51faf(h(x~;\u03b8))\u6620\u5c04\u4f1ag(x~)\u7684\u8f93\u51fa\u3002<\/p>\n\n\n\n

\"\"<\/figure><\/div>\n\n\n\n
\"\"
\u56fe 1\uff1a\u5bf9\u6297\u91cd\u7f16\u7a0b\u56fe\u793a\u3002\uff08a\uff09\u5c06 ImageNet \u6807\u7b7e\u6620\u5c04\u81f3\u5bf9\u6297\u4efb\u52a1\u6807\u7b7e\uff08\u56fe\u50cf\u4e2d\u7684\u65b9\u5757\uff09\u3002\uff08b\uff09\u6765\u81ea\u5bf9\u6297\u4efb\u52a1\u7684\u56fe\u50cf\uff08\u5de6\uff09\u88ab\u5d4c\u5165\u5bf9\u6297\u7a0b\u5e8f\u7684\u4e2d\u5fc3\uff08\u4e2d\uff09\uff0c\u5f97\u5230\u5bf9\u6297\u56fe\u50cf\uff08\u53f3\uff09\u3002\u8be5\u5bf9\u6297\u7a0b\u5e8f\u4f7f Inception V3 \u7f51\u7edc\u6267\u884c\u8ba1\u7b97\u56fe\u50cf\u4e2d\u65b9\u5757\u6570\u91cf\u7684\u4efb\u52a1\u3002\uff08c\uff09\u4f7f\u7528\u5bf9\u6297\u56fe\u50cf\u8fdb\u884c\u63a8\u65ad\u7684\u56fe\u793a\u3002\u628a\u5bf9\u6297\u56fe\u50cf\u8f93\u5165\u8be5\u7f51\u7edc\u65f6\uff0c\u7f51\u7edc\u9884\u6d4b\u6620\u5c04\u81f3\u5bf9\u6297\u4efb\u52a1\u7684 ImageNet \u6807\u7b7e\u3002<\/figcaption><\/figure>\n\n\n\n
\"\"<\/figure>\n\n\n\n

\u56fe2\u63d0\u4f9b\u4e86\u9002\u5e94\u9884\u8bad\u7ec3\u6a21\u578b\u7684\u4e0d\u540c\u65b9\u6cd5\u7684\u6458\u8981\u3002\u5fae\u8c03\u548c\u7ebf\u6027\u63a2\u6d4b\u7684\u7528\u6cd5\u975e\u5e38\u7075\u6d3b\uff1a\u5b83\u4eec\u53ef\u7528\u4e8e\u4f7f\u6a21\u578b\u9002\u5e94\u65b0\u7684\u8f93\u5165\u57df\u6216\u5177\u6709\u4e0d\u540c\u8f93\u51fa\u8bed\u4e49\u7684\u65b0\u4efb\u52a1\u3002\u4f46\u662f\uff0c\u5728\u7ebf\u6027\u63a2\u9488\u7684\u60c5\u51b5\u4e0b\uff0c\u5728\u5fae\u8c03\u548c\u6a21\u578b\u8f93\u51fa\uff08\u901a\u5e38\u5728\u5012\u6570\u7b2c\u4e8c\u5c42\u7684\u6fc0\u6d3b\uff09\u7684\u60c5\u51b5\u4e0b\uff0c\u4ed6\u4eec\u8fd8\u9700\u8981\u4e00\u5b9a\u7a0b\u5ea6\u7684\u8bbf\u95ee\u6a21\u578b\uff1a\u53c2\u6570\u3002\u57df\u7684\u9002\u5e94\u6027\u662f\u6a21\u578b\u9002\u5e94\u7684\u6709\u8da3\u66ff\u4ee3\u65b9\u6cd5\uff0c\u56e0\u4e3a\u5b83\u4ec5\u4f7f\u7528\u56fe\u50cf\u5230\u56fe\u50cf\u7ffb\u8bd1\u7b49\u6280\u672f\u6765\u4fee\u6539\u6a21\u578b\u7684\u8f93\u5165[50\uff0c19]\u3002\u50cf\u57df\u7684\u9002\u5e94\u6027\u4e00\u6837\uff0c\u89c6\u89c9\u63d0\u793a\u4e5f\u5c06\u8f93\u5165\u4fee\u6539\u4e3a\u6a21\u578b\u3002\u56e0\u6b64\uff0c\u4e00\u65e6\u6700\u7ec8\u7528\u6237\u627e\u5230\u4e86\u89c6\u89c9\u63d0\u793a\uff0c\u5c31\u4e0d\u9700\u8981\u5728\u6d4b\u8bd5\u65f6\u95f4\u63a7\u5236\u6a21\u578b\u672c\u8eab\u3002\u8fd9\u6253\u5f00\u4e86\u72ec\u7279\u7684\u5e94\u7528\u7a0b\u5e8f\uff1b\u4f8b\u5982\uff0c\u7528\u6237\u53ef\u4ee5\u5c06\u9002\u5e94\u57df\u7684\u56fe\u50cf\u9988\u9001\u5230\u53ea\u80fd\u901a\u8fc7\u8f93\u5165\u6765\u64cd\u7eb5\u7684\u5728\u7ebfAPI\u3002\u57df\u7684\u9002\u5e94\u6027\u91cd\u70b9\u662f\u8c03\u6574\u6e90\u57df\u4ee5\u770b\u8d77\u6765\u50cf\u76ee\u6807\u57df\uff0c\u540c\u65f6\u9700\u8981\u6e90\u548c\u76ee\u6807\u6570\u636e\u96c6\u3002\u53e6\u4e00\u65b9\u9762\uff0c\u6211\u4eec\u8ba4\u4e3a\u89c6\u89c9\u63d0\u793a\u53ef\u4ee5\u4ee5\u66f4\u4efb\u610f\u7684\u65b9\u5f0f\u5f15\u5bfc\u6a21\u578b\u3002\u4f8b\u5982\uff0c\u53ef\u4ee5\u901a\u8fc7\u6270\u52a8\u8f93\u5165\u50cf\u7d20\u6765\u4f7f\u7528\u65b0\u7684\u8f93\u51fa\u8bed\u4e49\u8fdb\u884c\u4e00\u4e2a\u5b8c\u5168\u4e0d\u540c\u7684\u5206\u7c7b\u4efb\u52a1\u6765\u6267\u884c\u5b8c\u5168\u4e0d\u540c\u7684\u5206\u7c7b\u4efb\u52a1\u3002\u540c\u6837\uff0c\u5c3d\u7ba1\u57df\u9002\u5e94\u65b9\u6cd5\u901a\u5e38\u662f\u8f93\u5165\u6761\u4ef6\uff0c\u4f46\u6211\u4eec\u5728\u672c\u6587\u4e2d\u63a2\u7d22\u7684\u89c6\u89c9\u63d0\u793a\u662f\u56fa\u5b9a\u7684\uff08\u5373\u8f93\u5165 – agnostic\uff09\uff0c\u5982NLP\uff0c\u4f8b\u5982\u5728NLP\u4e2d\uff0c\u5c06\u76f8\u540c\u7684\u81ea\u7136\u8bed\u8a00\u63d0\u793a\u6dfb\u52a0\u5230\u6240\u6709\u6a21\u578b\u67e5\u8be2\u4e2d<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

<\/p>\n","protected":false},"excerpt":{"rendered":"

Visual Prompt Tuning https:\/\/github.com\/KMnP\/vpt Explor … \u7ee7\u7eed\u9605\u8bfbVisual Prompt–\u89c6\u89c9\u6a21\u677f<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[21],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/10703"}],"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=10703"}],"version-history":[{"count":39,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/10703\/revisions"}],"predecessor-version":[{"id":11264,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/10703\/revisions\/11264"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=10703"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=10703"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=10703"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}