{"id":7683,"date":"2022-10-30T17:24:43","date_gmt":"2022-10-30T09:24:43","guid":{"rendered":"http:\/\/139.9.1.231\/?p=7683"},"modified":"2022-10-30T19:24:25","modified_gmt":"2022-10-30T11:24:25","slug":"clip-contrastive-language-image-pre-training","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2022\/10\/30\/clip-contrastive-language-image-pre-training\/","title":{"rendered":"CLIP: Contrastive Language-Image Pre-Training"},"content":{"rendered":"\n
\"\"<\/figure>\n\n\n\n

CLIP\u8bba\u6587\u8bb2\u89e3<\/strong><\/a><\/p>\n\n\n\n

\u80cc\u666f<\/h2>\n\n\n\n

\u8bba\u6587\u6765\u81ea Open AI 2021 \u5e74\u63d0\u51fa\u7684\u4e00\u4e2a\u6210\u679c\uff0c\u76f8\u5173\u53ef\u53c2\u8003\u4fe1\u606f\uff1a github<\/a>\u3001 paper<\/a>\u3001\u4e3b\u9875<\/a> \u3002 \u4e4b\u524d\u5176\u5b9e\u5e76\u4e0d\u592a\u4e86\u89e3\u591a\u6a21\u6001\u9884\u8bad\u7ec3\u9886\u57df\u7684\u6210\u679c\uff0c\u6700\u8fd1\u770b\u5230\u4e86\u8fd9\u7bc7\u8d28\u91cf\u5f88\u9ad8\u7684\u6210\u679c\u3002<\/p>\n\n\n\n

Hugging Face<\/strong><\/a>\uff1a https:\/\/huggingface.co\/openai\/clip-vit-base-patch32<\/a>\uff08\u9884\u8bad\u7ec3\u6a21\u578b\u5e93\uff09<\/p>\n\n\n\n

\u6548\u679c<\/h2>\n\n\n\n

\u6211\u4eec\u53ef\u4ee5\u8fd0\u884c\u8fd9colab<\/a>\uff0c\u8be5\u4f5c\u8005\u5c06 Unsplash \u7684\u6240\u6709\u7d20\u6750\u8ba1\u7b97\u4e86 clip image embedding \uff0c\u7136\u540e\u4f7f\u7528 clip word embedding \u8fdb\u884c\u914d\u56fe\u3002<\/p>\n\n\n\n

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

\u6548\u679c\u770b\u8d77\u6765\u4f3c\u4e4e\u4e0d\u9519\uff0c\u51e0\u4e4e\u5b9e\u73b0\u4e86\u901a\u8fc7\u4e00\u53e5\u8bdd\u5c31\u627e\u5230\u5408\u9002\u7684\u56fe\u7247\u3002\u4e0d\u53ef\u5426\u5b9a\uff0c\u4f1a\u5b58\u5728\u5927\u91cf\u7684badcase\u3002\u4f46\u662f\u5728\u4e0d\u9700\u8981fine-tune\/\u4e0b\u6e38\u4efb\u52a1\uff0c\u76f4\u63a5zero-shot\u5f97\u5230\u7684embedding\u53ef\u4ee5\u5b9e\u73b0\u8fd9\u6837\u7684\u6548\u679c\u5df2\u7ecf\u5f88\u5389\u5bb3\u4e86\u3002<\/p>\n\n\n\n

\u4f5c\u8005\u56e2\u961f\u6765\u81ea OPEN AI<\/p>\n\n\n\n

CLIP\u5de5\u4f5c\uff1a<\/p>\n\n\n\n

1 \u65b9\u6cd5\u7b80\u5355\uff0c\u6548\u679c\u597d<\/p>\n\n\n\n

2 \u8fc1\u79fb\u5b66\u4e60\u80fd\u529b\u5f3a\uff08\u5df2\u8bad\u7ec3\u597d\u7684\u6a21\u578b\uff0c\u53ef\u4ee5\u5728\u4efb\u610f\u6570\u636e\u96c6\u4e0a\u53d6\u5f97\u597d\u6548\u679c\uff09<\/p>\n\n\n\n

Q1 CLIP \u662f\u4ec0\u4e48\uff1fhow \u505azero-shot\uff08\u4e00\u79cd\u5206\u7c7b\u65b9\u5f0f\uff09\uff1f<\/p>\n\n\n\n

Q2 CLIP How \u9884\u8bad\u7ec3\uff1f<\/p>\n\n\n\n

\u5229\u7528\u4fe1\u53f7\uff08\u6765\u81ea\u81ea\u7136\u8bed\u8a00\u5904\u7406\uff09\u8bad\u7ec3\u4e00\u4e2a\u6a21\u578b\uff08\u8fc1\u79fb\u6548\u679c\u597d\uff09<\/p>\n\n\n\n

Q3 \u7ecf\u8fc7\u9884\u8bad\u7ec3\u80fd\u5f97\u5230\u4ec0\u4e48\uff1f<\/p>\n\n\n\n

A3 \u4ec5\u5f97\u5230\u56fe\u7247\u6216\u6587\u672c\u7684\u7279\u5f81\uff0c\u6ca1\u6709\u5728\u5206\u7c7b\u4efb\u52a1\u4e0a\u7ee7\u7eed\u505a\u8bad\u7ec3\u6216\u5fae\u8c03\u3002\u5373CLIP\u6ca1\u6709\u5206\u7c7b\u5934\u3002<\/p>\n\n\n\n

Q4 \u6ca1\u6709\u5206\u7c7b\u5934\uff0chow \u505a\u63a8\u7406\uff1f<\/p>\n\n\n\n

A4 \u5229\u7528\u81ea\u7136\u8bed\u8a00\u7684\u65b9\u6cd5—prompt template;<\/p>\n\n\n\n

\u5c061000\u4e2a\u7c7b\uff0c\u751f\u6210\u4e00\u4e2a1000\u4e2a\u53e5\u5b50\uff08object\uff09,<\/p>\n\n\n\n

\u4f8b\u5b50\uff1aplane \u53d8\u6210 object<\/p>\n\n\n\n

1000\u4e2a\u53e5\u5b50\u901a\u8fc7\u6587\u672c\u7f16\u7801\u5668\uff08text-Encode\uff09\u751f\u62101000\u4e2a\u7279\u5f81\u3002<\/p>\n\n\n\n

Q5 \u76f4\u63a5\u4ece1000\u4e2a\u7c7b\u91cc\u9762\u62bd\u53d6\u7279\u5f81\u4e5f\u53ef\u4ee5\uff0c why \u8fd8\u8981\u8fdb\u884cPrompt- template \uff1f<\/p>\n\n\n\n

A5 \u5728\u9884\u8bad\u7ec3\u65f6\uff0cmodel \u770b\u5230\u7684\u662fsample-pair\uff0c\u82e5\u5728\u63a8\u7406\u65f6\uff0c\u628a\u6240\u6709\u6587\u672c\u53d8\u6210\u4e00\u4e2a\u5355\u8bcd\uff08word)\uff0c\u5bfc\u81f4model\u770b\u5230\u7684\u4e1c\u897f\u548c\u9884\u8bad\u7ec3\u65f6\u4e0d\u4e00\u6837\uff0c\u5bfc\u81f4\u8bc6\u522b\u6548\u679c\u7a0d\u4e0b\u964d\u3002<\/p>\n\n\n\n

Q6 \u5982\u4f55\u5c061000\u5206\u7c7b\u53d8\u6210 1000\u4e2a\u53e5\u5b50\uff08object\uff09\uff1fwhy \u8fd9\u6837\u505a\uff1f<\/p>\n\n\n\n

A6 2\u4e2a\u65b9\u6cd5\uff1aprompt engineering \u548c prompot ensambol\uff1b<\/p>\n\n\n\n

\u63d0\u9ad8\u6a21\u578b\u51c6\u786e\u7387\uff0c\u4e14\u4e0d\u9700\u8981\u91cd\u65b0\u8bad\u7ec3\u6a21\u578b<\/p>\n\n\n\n

Q7 prompt template \u64cd\u4f5c\u4e4b\u540e\u8981\u5e72\u561b\uff1f<\/p>\n\n\n\n

A7 input \u56fe\u7247\uff0c\u7ecf\u8fc7image_Encode \u5f97\u5230\u56fe\u7247\u7279\u5f81\uff0c\u5229\u7528image_feature \u548c text-feature \u8ba1\u7b97\u76f8\u4f3c\u6027\uff0c\u6311\u51fa\u503c\uff08\u6700\u76f8\u4f3c\uff09\uff0c\u8fdb\u800c\u5b8c\u6210\u5206\u7c7b\u4efb\u52a1\u3002<\/p>\n\n\n\n

Q8 how \u7406\u89e3\u5206\u7c7b\u4efb\u52a1\uff1f<\/p>\n\n\n\n

A8 judge image \u4e2d\u6709\u54ea\u4e9b\u7269\u4f53<\/p>\n\n\n\n

Q9 text and image \u53ef\u4ee5\u6539\u5417\uff1f<\/p>\n\n\n\n

A9 yes \uff08all of anything\uff09<\/p>\n\n\n\n

Q10 \u82e5\u7528imageNet\u505a\u8bad\u7ec3\uff0cinput\u4e09\u8f6e\u8f66\uff08image\uff09\uff0cwhy \u5f97\u5230\u8f66\uff0c\u800c\u4e0d\u662f\uff0c\u4e09\u8f6e\u8f66\uff1f<\/p>\n\n\n\n

A10 \u56e0\u4e3a\uff0cimageNet \u65e0\u6cd5\u5b9e\u65f6\u66f4\u65b0\u5df2\u6709\u7c7b\u522b\u3002<\/p>\n\n\n\n

\u4f46CLIP\u53ef\u4ee5\u5b9e\u65f6\u66f4\u65b0\uff0c\u6545 imput = output\u3002<\/p>\n\n\n\n

\u8fd9\u4e5f\u662fCLIP\u7684\u5f3a\u5927\u4e4b\u5904\uff0c\u5f7b\u5e95\u6446\u8131\u4e86categoricel label \u9650\u5236\u3002<\/p>\n\n\n\n

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

\u4e3a\u4e86\u63d0\u9ad8model\u6cdb\u5316\u6027\uff0c\u4f5c\u8005\u63d0\u51fa\u65b0\u529e\u6cd5\uff0c\u4ece text \u4e2d\u63d0\u53d6\u76d1\u7763\u4fe1\u53f7\u3002\uff08\u6b63\u662f\u6709\u4e86\u76d1\u7763\u4fe1\u53f7\uff08\u8986\u76d6\u8303\u56f4\u5e7f\uff09\u7684\u5b58\u5728\uff0cmodel \u7684\u6cdb\u5316\u80fd\u529b\u5f97\u5230\u63d0\u9ad8\uff09\u4f5c\u8005\u5229\u75284\u4ebf text-image-pair-dataset \uff0c\u9009\u62e9\u81ea\u76d1\u7763\u8bad\u7ec3\u65b9\u5f0f\uff0c\u8fdb\u800c\u8bad\u7ec3\u6a21\u578b\u3002<\/p>\n\n\n\n

CLIP \u5229\u7528\u591a\u6a21\u6001\u5bf9\u6bd4\u5b66\u4e60\u5b8c\u6210\u8bad\u7ec3\uff0c\u5e76\u53ef\u4ee5\u505a\u7269\u4f53\u5206\u7c7b\uff08\u5373prompt\uff09\uff0c\u8fd9\u79cd\u5206\u7c7b\u4e0d\u9650\u4e8e\u5df2\u6709\u7c7b\u522b\uff0c\u53ef\u6269\u5c55\u5230\u65b0\u7c7b\u522b\u3002\uff08\u5373\u5f53\u524d\u5b66\u5230\u7684model\uff0c\u53ef\u4ee5\u76f4\u63a5\u5728 downstream tasks\uff08\u4e0b\u6e38\u4efb\u52a1\uff09\u4e0a\u505a\u63a8\u7406\u3002<\/p>\n\n\n\n

2017\u5e74\u6709\u4eba\u7814\u7a76\uff0c\u4f46\u662f\u5f71\u54cd\u529b\u5c0f\uff0c\u6548\u679c\u5dee\uff1a<\/p>\n\n\n\n

\u4e3b\u8981\u67093\u4e2a\u5de5\u4f5c(\u5747\u57fa\u4e8etransformer\uff09\u548cCLIP\u50cf\uff0c\u4f46\u6709\u533a\u522b\uff1aVIrTex\uff1a\u7528\u81ea\u56de\u5f52\u9884\u6d4b\u65b9\u5f0f\uff0c\u505amodel\u9884\u8bad\u7ec3\uff0cICMLM \uff1a\u7528\u5b8c\u578b\u586b\u7a7a\u65b9\u5f0f\uff0c\u505amodel\u9884\u8bad\u7ec3\uff0cConVIRT\uff1a\u548cCLIP\u7c7b\u4f3c\uff0c\u4f46\u4ec5\u5728\u533b\u7597\u56fe\u50cf\u4e0a\u505a\u5b9e\u9a8c\u3002\u4f46\u7531\u4e8edata \u548c model\u89c4\u6a21\u5c0f\uff0c\u6240\u4ee5\u6548\u679c\u4e0d\u597d\u3002<\/p>\n\n\n\n

\u5229\u7528\u81ea\u7136\u8bed\u8a00\u7684\u76d1\u7763\u4fe1\u53f7\uff0c\u6765\u8bad\u7ec3\u597d\u7684\u89c6\u89c9\u6a21\u578b\u3002\u5728\u81ea\u76d1\u7763\u5b66\u4e60\uff08\u5b8c\u578b\u586b\u7a7a\uff09\u7684\u8303\u5f0f\u4e0b\uff0cNLP\u53ef\u4ee5\u5229\u7528\uff08\u53d6\u4e0d\u5c3d\uff09\u7684\u6587\u672c\u76d1\u7763\u4fe1\u53f7\u3002\u7528\u6b64\u65b9\u6cd5\u8bad\u7ec3\u51fa\u7684\u6a21\u578b\uff0c\u7b80\u5355\uff0c\u6cdb\u5316\u529b\u5f3a\uff0c\u4e3a\u591a\u6a21\u6001\u8bad\u7ec3\u94fa\u8def\u3002<\/p>\n\n\n\n

why \u7528\u81ea\u7136\u8bed\u8a00\u76d1\u7763\u4fe1\u53f7\u8bad\u7ec3\u89c6\u89c9\u6a21\u578b\uff1f<\/p>\n\n\n\n

1 \u65e0\u9700\u6807\u6ce8\u8fd9\u4e9b\u6570\u636e\uff08\u6570\u636e\u89c4\u6a21\u53d8\u5927\uff09<\/p>\n\n\n\n

2 \u6b64\u65f6\u76d1\u7763\u4fe1\u53f7\u662f\u6587\u672c\uff08\u4e0d\u662fn\u90091 \u7684\u6807\u7b7e\uff09\uff0c\u610f\u5473\u7740input\uff0coutput \u81ea\u7531\u5ea6\u5927\u4e86\u5f88\u591a.<\/p>\n\n\n\n

3 \u56e0\u4e3aimage-text-pair\u6570\u636e\uff0cmodel\u6240\u5b66\u7279\u5f81\u4e0d\u5355\u662f\u89c6\u89c9\u7279\u5f81\uff0c\u800c\u662f\u591a\u6a21\u6001\u7279\u5f81\u3002\u5f53image\u548c\u8bed\u8a00\u8054\u7cfb\u5728\u4e00\u8d77\uff0c\u4fbf\u5bb9\u6613\u505azero-shot\u8fc1\u79fb\u5b66\u4e60\u3002<\/p>\n\n\n\n

\u82e5\u4ec5\u505a\u5355\u6a21\u6001\u81ea\u76d1\u7763\u5b66\u4e60\uff0c\u65e0\u8bba\u662f\u5355\u6a21\u6001\u5bf9\u6bd4\u5b66\u4e60\uff08MOCO\uff09\uff0c\u8fd8\u662f\u5355\u6a21\u6001\u63a9\u7801\u5b66\u4e60\uff08MAE\uff09\uff0cmodel\u4ec5\u5b66\u5230\u89c6\u89c9\u7279\u5f81\uff0c\u65e0\u6cd5\u548c\u81ea\u7136\u8bed\u8a00\u8054\u7cfb\u5728\u4e00\u8d77\uff0c\u4f9d\u65e7\u5f88\u96be\u505azero-shot\u7684\u8fc1\u79fb\u3002\u9700\u8981\u5927\u91cf\u7684image-text-pair\uff084\u4ebf\u4e2aImage-text-pair\uff09<\/p>\n\n\n\n

\u603b\u7ed3\uff1a\u7528 \u6587\u672c\u76d1\u7763\u4fe1\u53f7\u6765\u8bad\u7ec3\u89c6\u89c9model \u8fd9\u79cd\u505a\u6cd5\u5f88\u6709\u6f5c\u529b\u3002<\/p>\n\n\n\n

\u6574\u4e2a\u8bad\u7ec3\u8fc7\u7a0b\uff1a<\/p>\n\n\n\n

\u7ed9\u5b9a\u4e00\u5f20image\uff0c\u6765\u9884\u6d4b\u6587\u672c\uff0c\u4f1a\u4ea7\u751f\u8f83\u5927\u6b67\u4e49\uff08\u5373\u53ef\u80fd\u6027\u592a\u591a\uff09\uff1b\u82e5\u9010\u5b57\u53e5\u9884\u6d4b\u6587\u672c\uff0c\u592a\u96be\u4e86\u3002\u4f1a\u5bfc\u81f4\u6a21\u578b\u8bad\u7ec3\u6162\u56e0\u6b64\u91c7\u7528\u5bf9\u6bd4\u5b66\u4e60\uff0c\u8ba9 model \u5224\u65ad\uff0cimage \u548c text \u662f\u5426\u914d\u5bf9\u3002\u628a \u201d \u8bad\u7ec3\u4efb\u52a1 \u201c \u6362\u6210 \u201d \u5bf9\u6bd4\u4efb\u52a1 \u201c \uff0c\u8bad\u7ec3\u6548\u7387\u63d0\u9ad84\u500d<\/p>\n\n\n\n

2\u4e2a\u8f93\u5165\uff1aimage \u548c text\u5f52\u4e00\u5316\uff0c\u6295\u5c04\u5c42\uff1a\u5c06\u5355\u6a21\u6001\u53d8\u6210\u591a\u6a21\u6001\uff0c\u83b7\u5f97 n \u4e2a\u56fe\u50cf\u7684\u7279\u5f81\uff0cn \u4e2a\u6587\u672c\u7684\u7279\u5f81\u3002\u8ba1\u7b97 image-feature \u548c text-feature\u76f8\u4f3c\u5ea6\u3002\u5229\u7528\u76f8\u4f3c\u5ea6\u505a\u5206\u7c7b\u3002\u5229\u7528\u4ea4\u53c9\u71b5\u76ee\u6807\u51fd\u6570\u8ba1\u7b97loss<\/p>\n\n\n\n

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

\u7ec6\u8282\uff1a<\/p>\n\n\n\n

1 \u7531\u4e8e\u6536\u96c6\u6570\u636e\u5927\uff0cmodel \u4e0d\u5b58\u5728 overfitting<\/p>\n\n\n\n

\u7b80\u5316\u4e86\u5de5\u4f5c\uff1a\u5f53\u8bad\u7ec3CLIP – model\u65f6\uff0c\u5bf9\u5e94\u7684 image-Encode \u548c text-Encode \u65e0\u9700\u8fdb\u884c\u9884\u8bad\u7ec3<\/p>\n\n\n\n

2 \u5728\u591a\u6a21\u6001\u8bad\u7ec3\u4e2d\uff0c\u6295\u5c04\u65f6\uff0c\u7528\u7ebf\u6027\u6295\u5c04\u5c42\uff0c<\/p>\n\n\n\n

\u975e\u7ebf\u6027\u6295\u5c04\u5c42\uff08\u4f5c\u8005\u63a8\u6d4b\uff0c\u9002\u914d\u7eaf Image \u5355\u6a21\u6001\u5b66\u4e60\uff09\uff0c\u5e26\u676510\u4e2a\u70b9\u7684\u6027\u80fd\u63d0\u5347<\/p>\n\n\n\n

3 \u4f7f\u7528 \u201d \u968f\u5373\u88c1\u51cf \u201c \u8fdb\u884c\u6570\u636e\u589e\u5f3a<\/p>\n\n\n\n

4 \u6570\u636e\u96c6\u548c model \u592a\u5927\uff0c\u4e0d\u597d\u8c03\u53c2<\/p>\n\n\n\n

5 temperature parm\uff08\u6781\u91cd\u8981\u8d85\u53c2\u6570\uff09\u7a0d\u8c03\uff0cmodel \u6027\u80fd\u4f1a\u63d0\u9ad8\u5f88\u591a \u4f46\u4f5c\u8005\uff0c\u5c06\u5176\u8bbe\u7f6e\u6210\uff0c\u53ef\u5b66\u4e60\u7684\u6807\u91cf<\/p>\n\n\n\n

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

\u6a21\u578b\u9009\u62e9\u548c\u53c2\u6570\u8bbe\u7f6e\uff1a<\/p>\n\n\n\n

\u89c6\u89c9\u65b9\u9762\uff1a<\/p>\n\n\n\n

\u8bad\u7ec38\u4e2amodel\uff1b<\/p>\n\n\n\n

ResNet = 5\u4e2a\uff0cVIT = 3\u4e2a<\/p>\n\n\n\n

\u6b8b\u5dee\u7f51\u7edc\u53d8\u4f53ResNet50*4:*16:*64\uff1a\u7528 efficientNet-style \u65b9\u6cd5 \u5c06input-image \u5927\u5c0f\uff0cchannel\u5bbd\u5ea6\uff0cmodel-depth \u505a\u5fae\u8c03<\/p>\n\n\n\n

\u9488\u5bf9 transformer-model\uff0c\u4f5c\u8005\u9009\u62e9\u6570\u636e\u96c6 VIT-B\/32:\/16:\/64(\u963f\u62c9\u4f2f\u6570\u5b57\u8868\u793apatch\u5927\u5c0f\uff09<\/p>\n\n\n\n

\u6587\u672c\u65b9\u9762\uff1atransformer<\/p>\n\n\n\n

all-model \u8bad\u7ec3\u4e86epoch = 32\uff1bAdam optimizer\u4f18\u5316\u5668\uff1b\u624b\u52a8\u8c03\u6574\u8d85\u53c2\u6570\uff0c\u7528ResNet-50\u4f5c\u4e3a\u8d85\u53c2\u641c\u7d22\uff0c\u4e3a\u4e86\u5feb\u901f\u8c03\u53c2\uff08\u8bad\u7ec3epoch = 1\uff09\u8bad\u7ec3\u65f6\uff1a\u9009\u7528 batch-size = 32768\uff08\u5f88\u5927\uff09\uff08\u6b64model\u5728\u5f88\u591a\u673a\u5668\u4e0a\u505a\u5206\u5e03\u5f0f\u8bad\u7ec3\uff09<\/p>\n\n\n\n

CLIP \u6587\u7ae0\u7684\u6838\u5fc3 = Zero-shot Transfer<\/strong><\/p>\n\n\n\n

\u4f5c\u8005\u7814\u7a76\u8fc1\u79fb\u5b66\u4e60\u7684\u52a8\u673a\uff1a\u4e4b\u524d\u81ea\u76d1\u7763or\u65e0\u76d1\u7763\u7684\u65b9\u6cd5\uff0c\u4e3b\u8981\u7814\u7a76 frature \u5b66\u4e60\u7684\u80fd\u529b\uff0cmodel\u7684\u76ee\u6807\u662f\u5b66\u4e60\u6cdb\u5316\u6027\u80fd\u597d\u7684\u7279\u5f81\uff0c\u867d\u7136\u5b66\u4e60\u5230good-feature\uff0c\u4f46down-work\u4e2d\uff0c\u8fd8\u662f\u9700\u8981\u6709\u6807\u7b7e\u6570\u636e\u505a\u5fae\u8c03\u3002\u4f5c\u8005\u60f3\u4ec5\u8bad\u7ec3\u4e00\u4e2amodel\uff0c\u5728down-work\u4e2d\u4e0d\u518d\u5fae\u8c03\u3002<\/p>\n\n\n\n

\u8861\u91cfmodel \u5b66\u5230\u7684feature \u597d\u4e0d\u597d\u7684\u65b9\u6cd5\u6709\u4e3b\u8981\u67092\u79cd\uff1a\u7b2c\u4e00\u79cd\uff1alinear\uff1a\u51bb\u7ed3\u8bad\u7ec3\u597d\u7684model\uff0c\u518d\u8bad\u7ec3\u4e00\u4e2a\u5206\u7c7b\u5934\u3002\u7b2c\u4e8c\u79cd\uff1a\u5fae\u8c03\uff1a\u628a\u6574\u4e2a\u7f51\u7edc\u653e\u5f00\uff0c\u505aend-to-end\u7684\u5b66\u4e60\u3002\u5fae\u8c03\u7684\u4f18\u70b9\uff1a \u7075\u6d3b\u3001\u5f53down-work\u6570\u636e\u96c6\u5927\uff0c\u5fae\u8c03\u6548\u679c\u597d\u3002\u4f46\u8fd9\u91cc\u4f5c\u8005\u4f7f\u7528\u53ea\u8bad\u7ec3liner\u5206\u7c7b\u5934\uff1a CLIP\u672c\u5c31\u7528\u6765\u7814\u7a76\u66f4\u6570\u636e\u96c6\u65e0\u5173\u7684\u8bad\u7ec3\u65b9\u5f0f\uff0c\u82e5\u7528 \u201c \u5fae\u8c03 \u201c \u65b9\u6cd5\uff0c\u65e0\u6cd5\u5224\u65ad\u9884\u8bad\u7ec3model\u6548\u679c\u5982\u4f55\u3002\uff08\u56e0\u4e3a\uff0c\u5982\u679c\u9884\u8bad\u7ec3model\u6548\u679c\u4e0d\u597d\uff0c\u7ecf\u8fc7\u5728down-work\u4e0a\u505a\u5fae\u8c03\uff0c\u4f1a\u5bfc\u81f4\u6700\u7ec8\u7ed3\u679c\u597d\u3002\uff09<\/p>\n\n\n\n

CLIP\u8fd9\u4e48\u5f3a\u5927\uff0c\u5b83\u6709\u4ec0\u4e48\u7f3a\u70b9\uff1f<\/p>\n\n\n\n

\u5e73\u5747\u6765\u770b\uff0cCLIIP\u53ef\u4ee5\u548c\u673a\u68b0\u6a21\u578b\uff08ResNet-50\uff08\u5728ImageNet\u4e0a\u8bad\u7ec3\uff09\uff09\u6301\u5e73<\/p>\n\n\n\n

\u82e5\u7ee7\u7eed\u589e\u52a0\u6570\u636e\u96c6\u548cmodel\u89c4\u6a21\uff0cCLIP\u6027\u80fd\u53ef\u4ee5\u7ee7\u7eed\u63d0\u9ad8\uff0c\u4f46\u662f\u4ee3\u4ef7\u5f88\u5927\uff08\u9700\u63d0\u9ad8\u8ba1\u7b97\u548c\u6570\u636e\u7684\u9ad8\u6548\u6027\uff09<\/p>\n\n\n\n

zreo-shot\u7ed3\u679c\u5e76\u4e0d\u597d<\/p>\n\n\n\n

1 \u5728\u7ec6\u5206\u7c7b\u6570\u636e\u96c6\u4e0a\uff0cCLIP\u6548\u679c\u4f4e\u4e8e\uff08\u6709\u76d1\u7763\u8bad\u7ec3\uff09ResNet-50\uff08\u6781\u9650\u7f51\u7edc\uff09<\/p>\n\n\n\n

2 CLIP\u65e0\u6cd5\u5904\u7406\u62bd\u8c61\u6982\u5ff5\u539f\u56e0\uff1aCLUP\u65e0\u6cd5\u533a\u5206 what is \u5f02\u5e38\uff1fwhat is \u5b89\u5168\uff1f\u4f8b\u5982\uff1a\u6570\u4e00\u6570\u56fe\u7247\u4e2d\u7684\u7269\u4f53\u4e2a\u6570\uff1b\u5728\u89c6\u9891\u4e2d\uff0c\u533a\u5206\u8fd9\u4e00\u5e27\u662f\u5f02\u5e38\u8fd8\u662f\u975e\u5f02\u5e38\uff1b\u4f5c\u8005\u63d0\u51fa\uff1a\u5728\u5f88\u591a\u9886\u57df\uff0cCLIP\u6027\u80fd\u548c\u778e\u731c\u5dee\u4e0d\u591a<\/p>\n\n\n\n

3 \u82e5\u6570\u636e\u96c6\u4e2d\u7684data \u5df2\u7ecf out-of-distribution,\u90a3\u4e48CLIP-model\u6cdb\u5316\u7167\u6837\u5dee\uff1b\u4f8b\u5b50\uff1a\u5728MNIST\u6570\u636e\u96c6\u4e0a\uff0cCLIP\u51c6\u786e\u7387\u4ec5\u670988% \u3002\u63a8\u6d4b\u539f\u56e0\uff1a\u4f5c\u8005\u6536\u96c6\u7684\u6570\u636e\u96c6\u67094\u4ebf\u4e2a\u6837\u672c\uff0c\u4f46\u6ca1\u6709\u548cMINIS\u957f\u5f97\u50cf\u7684\uff0c\u6240\u4ee5MINIS\u6570\u636e\u96c6\u5bf9\u4e8eCLIP\u6765\u8bf4\u5c31\u662fout-of-distribution\u6570\u636e\u96c6<\/p>\n\n\n\n

\u8bc4\u4ef7<\/p>\n\n\n\n

\u521b\u65b0\u5ea6\u9ad81 \u6253\u7834\u56fa\u5b9a\u7c7b\u522b\u6807\u7b7e\u505a\u6cd52 \u653e\u98de\u89c6\u89c9model\u8bad\u7ec3\u8fc7\u7a0b3 \u5f15\u53d1\u540e\u7eed\u5927\u91cf\u5de5\u4f5c<\/p>\n\n\n\n

\u6709\u6548\u6027\u9ad81 \u5927\u6570\u636e\u96c6\uff0c\u6548\u679c\u597d2 \u6cdb\u5316\u6027\u80fd\u597d3 zero-shot\u6027\u80fd\u8d85\u8fc7\u4eba\u7c7b<\/p>\n","protected":false},"excerpt":{"rendered":"

CLIP\u8bba\u6587\u8bb2\u89e3 \u80cc\u666f \u8bba\u6587\u6765\u81ea Open AI 2021 \u5e74\u63d0\u51fa\u7684\u4e00\u4e2a\u6210\u679c\uff0c\u76f8\u5173\u53ef\u53c2\u8003\u4fe1\u606f\uff1a gi … \u7ee7\u7eed\u9605\u8bfbCLIP: Contrastive Language-Image Pre-Training<\/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\/7683"}],"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=7683"}],"version-history":[{"count":60,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/7683\/revisions"}],"predecessor-version":[{"id":9783,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/7683\/revisions\/9783"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=7683"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=7683"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=7683"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}