{"version":2,"relationshipURLBase":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html","relationshipResolution":"Resolve relative relationship URLs against relationshipURLBase; complete records and absolute URLs are in blocks.","title":"CLS 10 · CLIP：配对学习与可复用表示","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html","markdown":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.md","pages":[{"id":"reading-opening","title":"CLS 10 · CLIP：配对学习与可复用表示","unit":"已知图文配对怎样变成可用的文字候选接口？","unitId":"clip-pairing-to-candidates","chapter":"III","number":1},{"id":"reading-block-ed31be25c84b4c4eb511e2b846038ac3","title":"两个编码器，一个配对目标","unit":"已知图文配对怎样变成可用的文字候选接口？","unitId":"clip-pairing-to-candidates","chapter":"III","number":2},{"id":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","title":"真实猫图结果：保留模型、照片与历史身份","unit":"已知图文配对怎样变成可用的文字候选接口？","unitId":"clip-pairing-to-candidates","chapter":"III","number":3}],"aliases":[{"oldId":"b-2133e06b58b81a6298c57b3f","targetId":"b-2133e06b58b81a6298c57b3f","page":"reading-opening","operation":"move","beforeSHA256":"271272473744b0e1471941a678bc961da666faaf6e5f5117d03c24ec7bfb1a92","afterSHA256":"271272473744b0e1471941a678bc961da666faaf6e5f5117d03c24ec7bfb1a92","reason":"Same source body in authored semantic page"},{"oldId":"b-7272c979a6f6fc7b76ed5483","targetId":"b-7272c979a6f6fc7b76ed5483","page":"reading-opening","operation":"move","beforeSHA256":"e91f24e4e4ca40e2099a0d81d30db6568f4bf2074b499eeee6e3959617bd97ac","afterSHA256":"e91f24e4e4ca40e2099a0d81d30db6568f4bf2074b499eeee6e3959617bd97ac","reason":"Same source body in authored semantic page"},{"oldId":"b-6341975a56cc5ebe45a12980","targetId":"b-6341975a56cc5ebe45a12980","page":"reading-opening","operation":"move","beforeSHA256":"8b4a38fbe2d21e8c471d155a892cffc7decd7adfc064650236630fb3edf7a7d0","afterSHA256":"8b4a38fbe2d21e8c471d155a892cffc7decd7adfc064650236630fb3edf7a7d0","reason":"Same source body in authored semantic page"},{"oldId":"b-32ae2b6149fd1928b5ba0b0b","targetId":"b-32ae2b6149fd1928b5ba0b0b","page":"reading-opening","operation":"move","beforeSHA256":"f22152fd55dcdb0090b5653d37145694fd19388d31e10226549464fe8b483578","afterSHA256":"f22152fd55dcdb0090b5653d37145694fd19388d31e10226549464fe8b483578","reason":"Same source body in authored semantic page"},{"oldId":"b-c6b7904edcdfc53b19c45331","targetId":"b-c6b7904edcdfc53b19c45331","page":"reading-block-ed31be25c84b4c4eb511e2b846038ac3","operation":"move","beforeSHA256":"bab4bd804571aa5afbaf197e9c1539510fc766578c9825aa46cadbb18444cc32","afterSHA256":"bab4bd804571aa5afbaf197e9c1539510fc766578c9825aa46cadbb18444cc32","reason":"Same source body in authored semantic page"},{"oldId":"b-8c4e14f696186a7250637b50","targetId":"b-8c4e14f696186a7250637b50","page":"reading-block-ed31be25c84b4c4eb511e2b846038ac3","operation":"move","beforeSHA256":"50590b614b01aacf740b6f59f0a25be53dce523d2843825124cc6e1195f0f09c","afterSHA256":"50590b614b01aacf740b6f59f0a25be53dce523d2843825124cc6e1195f0f09c","reason":"Same source body in authored semantic page"},{"oldId":"b-fc0851c6deeffcdc567be124","targetId":"b-fc0851c6deeffcdc567be124","page":"reading-block-ed31be25c84b4c4eb511e2b846038ac3","operation":"move","beforeSHA256":"2e7f9ca1ab75355dcb74d4a3f8cd95d8ec9fc0ddc39e8919ee0ca1d8fa2e2052","afterSHA256":"2e7f9ca1ab75355dcb74d4a3f8cd95d8ec9fc0ddc39e8919ee0ca1d8fa2e2052","reason":"Same source body in authored semantic page"},{"oldId":"b-447dcc8027bd54aac453b630","targetId":"b-447dcc8027bd54aac453b630","page":"reading-block-ed31be25c84b4c4eb511e2b846038ac3","operation":"move","beforeSHA256":"2b31d6a9ce476da9fbc54ff1cd84e4d23c0a209eef12ce6ab7054f0dfa6f5c76","afterSHA256":"2b31d6a9ce476da9fbc54ff1cd84e4d23c0a209eef12ce6ab7054f0dfa6f5c76","reason":"Same source body in authored semantic page"},{"oldId":"b-95667271fd87b3983c5ccfca","targetId":"b-95667271fd87b3983c5ccfca","page":"reading-block-ed31be25c84b4c4eb511e2b846038ac3","operation":"move","beforeSHA256":"bfed64de92e5326d9f2ea08a5fbf216a9101f67d42e71710906c6ab6d2692320","afterSHA256":"bfed64de92e5326d9f2ea08a5fbf216a9101f67d42e71710906c6ab6d2692320","reason":"Same source body in authored semantic page"},{"oldId":"b-135d9ab06069fb9299dbbfeb","targetId":"b-135d9ab06069fb9299dbbfeb","page":"reading-block-ed31be25c84b4c4eb511e2b846038ac3","operation":"move","beforeSHA256":"063457994938d4e5dfe61567141bd163cb7d7cce6959928c1c0eb5510469ab6b","afterSHA256":"063457994938d4e5dfe61567141bd163cb7d7cce6959928c1c0eb5510469ab6b","reason":"Same source body in authored semantic page"},{"oldId":"b-7bb2ed61498b0b916e947855","targetId":"b-7bb2ed61498b0b916e947855","page":"reading-block-ed31be25c84b4c4eb511e2b846038ac3","operation":"move","beforeSHA256":"421bb86160826d098ebd40ff162672a976ef2d422bd0fce72fa666c4eaad7d3e","afterSHA256":"421bb86160826d098ebd40ff162672a976ef2d422bd0fce72fa666c4eaad7d3e","reason":"Same source body in authored semantic page"},{"oldId":"b-930d6b51abdffcd925a659ac","targetId":"b-930d6b51abdffcd925a659ac","page":"reading-block-ed31be25c84b4c4eb511e2b846038ac3","operation":"move","beforeSHA256":"eba1b58a7fce19f59185999c38f5f654b59946a33ae131684e1abbc607f712e6","afterSHA256":"eba1b58a7fce19f59185999c38f5f654b59946a33ae131684e1abbc607f712e6","reason":"Same source body in authored semantic page"},{"oldId":"b-7bb9e12f4dfb95950657ed76","targetId":"b-7bb9e12f4dfb95950657ed76","page":"reading-block-ed31be25c84b4c4eb511e2b846038ac3","operation":"move","beforeSHA256":"6b5dab1b128812a45c0237b9943cd89a557634629a8386260132bba6eb7d7a14","afterSHA256":"6b5dab1b128812a45c0237b9943cd89a557634629a8386260132bba6eb7d7a14","reason":"Same source body in authored semantic page"},{"oldId":"b-9bd19c6a16a5443435bab304","targetId":"b-9bd19c6a16a5443435bab304","page":"reading-block-ed31be25c84b4c4eb511e2b846038ac3","operation":"move","beforeSHA256":"b82eb99f0b75c3c6d07cae11fdb8041ed3871643065e2b41ebf2f1d956f99de4","afterSHA256":"b82eb99f0b75c3c6d07cae11fdb8041ed3871643065e2b41ebf2f1d956f99de4","reason":"Same source body in authored semantic page"},{"oldId":"b-9d936330364cc5aa6e283082","targetId":"b-9d936330364cc5aa6e283082","page":"reading-block-ed31be25c84b4c4eb511e2b846038ac3","operation":"move","beforeSHA256":"eeb43459eb9f075a95c7403ff26d52e85c14cf632e9eef7d81193d82eb7eae86","afterSHA256":"eeb43459eb9f075a95c7403ff26d52e85c14cf632e9eef7d81193d82eb7eae86","reason":"Same source body in authored semantic page"},{"oldId":"b-8c31ad492335867d64c520a2","targetId":"b-8c31ad492335867d64c520a2","page":"reading-block-ed31be25c84b4c4eb511e2b846038ac3","operation":"move","beforeSHA256":"e2fd6ed94ff5812ee027b0c1d2a3d46127eb13403d0e3f6c267ac4ef37f49552","afterSHA256":"e2fd6ed94ff5812ee027b0c1d2a3d46127eb13403d0e3f6c267ac4ef37f49552","reason":"Same source body in authored semantic page"},{"oldId":"b-76475b87f615c444a5940365","targetId":"b-76475b87f615c444a5940365","page":"reading-block-ed31be25c84b4c4eb511e2b846038ac3","operation":"move","beforeSHA256":"38c3c7217ff456579df20fe9c5703f47cfc3e63d9dcbb6bd611964eef7aa9ad0","afterSHA256":"38c3c7217ff456579df20fe9c5703f47cfc3e63d9dcbb6bd611964eef7aa9ad0","reason":"Same source body in authored semantic page"},{"oldId":"b-d3d06900a49a6a0a90867f1e","targetId":"b-d3d06900a49a6a0a90867f1e","page":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","operation":"move","beforeSHA256":"b81b865ed7c96a04916ece8e45b5f513d6d48a0b95cc2731a46f2c60802de664","afterSHA256":"b81b865ed7c96a04916ece8e45b5f513d6d48a0b95cc2731a46f2c60802de664","reason":"Same source body in authored semantic page"},{"oldId":"b-5857325209e2365e5dae8218","targetId":"b-5857325209e2365e5dae8218","page":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","operation":"move","beforeSHA256":"3a74a62e574be11b30f988dbb15451bb02c361669a9eefdaa6127d8a74bad6f5","afterSHA256":"3a74a62e574be11b30f988dbb15451bb02c361669a9eefdaa6127d8a74bad6f5","reason":"Same source body in authored semantic page"},{"oldId":"b-9d65fc5ff6b28cfbf499b2e1","targetId":"b-9d65fc5ff6b28cfbf499b2e1","page":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","operation":"move","beforeSHA256":"c7061a8ad0794772fc5fd4b5f0bdb6e66e098c0aaa2732e4c041ff8bb739810e","afterSHA256":"c7061a8ad0794772fc5fd4b5f0bdb6e66e098c0aaa2732e4c041ff8bb739810e","reason":"Same source body in authored semantic page"},{"oldId":"b-adc2df8499b31659b5a5bec2","targetId":"b-adc2df8499b31659b5a5bec2","page":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","operation":"move","beforeSHA256":"1fb9bd54f6a6d811bd1b68919726fd1177e6a5264ddb1afa54180876084827ba","afterSHA256":"1fb9bd54f6a6d811bd1b68919726fd1177e6a5264ddb1afa54180876084827ba","reason":"Same source body in authored semantic page"},{"oldId":"b-d249a93f276ebd6170824bdf","targetId":"b-d249a93f276ebd6170824bdf","page":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","operation":"move","beforeSHA256":"24e101bfca64bca25487af0223ddf64733850ac2f198d1326f78447406f724ff","afterSHA256":"24e101bfca64bca25487af0223ddf64733850ac2f198d1326f78447406f724ff","reason":"Same source body in authored semantic page"},{"oldId":"b-e30f6b010a6a48ccdaf011be","targetId":"b-e30f6b010a6a48ccdaf011be","page":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","operation":"move","beforeSHA256":"fa47d13f4375702edee587849800946fb46b0cb96bc5dfacf00c80ad1dd47c44","afterSHA256":"fa47d13f4375702edee587849800946fb46b0cb96bc5dfacf00c80ad1dd47c44","reason":"Same source body in authored semantic page"},{"oldId":"b-70aca0628f9012a0af754378","targetId":"b-70aca0628f9012a0af754378","page":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","operation":"move","beforeSHA256":"0d2f7a9737f9a95c5d7150d1ce3364baa7d3b82559a2be501d1690dcd74647f4","afterSHA256":"0d2f7a9737f9a95c5d7150d1ce3364baa7d3b82559a2be501d1690dcd74647f4","reason":"Same source body in authored semantic page"},{"oldId":"b-02e48d56c1938c354a142f24","targetId":"b-02e48d56c1938c354a142f24","page":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","operation":"move","beforeSHA256":"0d52d45b91d64683943b3870efe42eb2211cb249baaa5d700568f83f16ac824f","afterSHA256":"0d52d45b91d64683943b3870efe42eb2211cb249baaa5d700568f83f16ac824f","reason":"Same source body in authored semantic page"},{"oldId":"b-fa4fdd5898ca2d21852cf336","targetId":"b-fa4fdd5898ca2d21852cf336","page":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","operation":"move","beforeSHA256":"548903116868e99139fe6cb659aa140a85b965dcc46a1ad6a79904bd982e3beb","afterSHA256":"548903116868e99139fe6cb659aa140a85b965dcc46a1ad6a79904bd982e3beb","reason":"Same source body in authored semantic page"},{"oldId":"b-c941b94f48d71cc3fbf154f1","targetId":"b-c941b94f48d71cc3fbf154f1","page":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","operation":"move","beforeSHA256":"a0b9c4c8d9eebe422e20a0d9595195dc58efc43923716e264bd41dc0bfb60c5d","afterSHA256":"a0b9c4c8d9eebe422e20a0d9595195dc58efc43923716e264bd41dc0bfb60c5d","reason":"Same source body in authored semantic page"},{"oldId":"b-2602b161f592612b6f37f89b","targetId":"b-2602b161f592612b6f37f89b","page":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","operation":"move","beforeSHA256":"5a2ad5d4eb4c55adef619a9d7099e3489a0df2782017a87a2a4593130fcb2eaf","afterSHA256":"5a2ad5d4eb4c55adef619a9d7099e3489a0df2782017a87a2a4593130fcb2eaf","reason":"Same source body in authored semantic page"},{"oldId":"b-6f5b6cb742b0ec0e2a0effc9","targetId":"b-6f5b6cb742b0ec0e2a0effc9","page":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","operation":"move","beforeSHA256":"e4d7a3e0afa1d5e068c5eb162d5fae7ff4cdb68424305df478e479d6de6d3cbc","afterSHA256":"e4d7a3e0afa1d5e068c5eb162d5fae7ff4cdb68424305df478e479d6de6d3cbc","reason":"Same source body in authored semantic page"},{"oldId":"b-2e17c00ccaa0efdcdfcc7553","targetId":"b-2e17c00ccaa0efdcdfcc7553","page":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","operation":"move","beforeSHA256":"c54f1c8eedd49ebd3230440feaf9249ecf82351a2c3ac32e70e43f150b45a031","afterSHA256":"c54f1c8eedd49ebd3230440feaf9249ecf82351a2c3ac32e70e43f150b45a031","reason":"Same source body in authored semantic page"},{"oldId":"b-7a411282657df1acb7cb8644","targetId":"b-7a411282657df1acb7cb8644","page":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","operation":"move","beforeSHA256":"7d1e522bf167385f95073a108d6651706c7ab2f25d134521e17b40fe2992bc59","afterSHA256":"7d1e522bf167385f95073a108d6651706c7ab2f25d134521e17b40fe2992bc59","reason":"Same source body in authored semantic page"},{"oldId":"b-39ced42e12636a5a74bc42ea","targetId":"b-39ced42e12636a5a74bc42ea","page":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","operation":"move","beforeSHA256":"e60979fc29d0931ab0fa6565feb357a2b9d204702fac523f27809082dbf318a7","afterSHA256":"e60979fc29d0931ab0fa6565feb357a2b9d204702fac523f27809082dbf318a7","reason":"Same source body in authored semantic page"},{"oldId":"b-913e50be755c439d698f1327","targetId":"b-913e50be755c439d698f1327","page":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","operation":"move","beforeSHA256":"44fa60a15468578fc2e7b4d15edcd2f5936efdbc31c698e97d1d17c94fdf3a93","afterSHA256":"44fa60a15468578fc2e7b4d15edcd2f5936efdbc31c698e97d1d17c94fdf3a93","reason":"Same source body in authored semantic page"},{"oldId":"b-5a01031cf9b76362130c983c","targetId":"b-5a01031cf9b76362130c983c","page":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","operation":"move","beforeSHA256":"c1afda030628985fb13e6d5c22f9cfe7d7b0fe2cab59e7e3f191abb1dcef9feb","afterSHA256":"c1afda030628985fb13e6d5c22f9cfe7d7b0fe2cab59e7e3f191abb1dcef9feb","reason":"Same source body in authored semantic page"},{"oldId":"b-fbf2fd28eec78b34ef0b44b5","targetId":"b-fbf2fd28eec78b34ef0b44b5","page":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","operation":"move","beforeSHA256":"053e231943f3ce6c919967bff3b098a25f5d7c5ed258867bf88b6fda5464e1f2","afterSHA256":"053e231943f3ce6c919967bff3b098a25f5d7c5ed258867bf88b6fda5464e1f2","reason":"Same source body in authored semantic page"},{"oldId":"b-badbf96a9016a6dfa98db812","targetId":"b-badbf96a9016a6dfa98db812","page":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","operation":"move","beforeSHA256":"fedaaf52a99a89fdf8983619b82d1d3f90c7f2425a6c4494e17d99a8fd97c48f","afterSHA256":"fedaaf52a99a89fdf8983619b82d1d3f90c7f2425a6c4494e17d99a8fd97c48f","reason":"Same source body in authored semantic page"},{"oldId":"b-34ab49a4fd1511ac5bdc564c","targetId":"b-34ab49a4fd1511ac5bdc564c","page":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","operation":"move","beforeSHA256":"ab2e895629a39a47c6fc0f16ba70892a54bee20eb9e4b1ddf18d8f3334fb81fe","afterSHA256":"ab2e895629a39a47c6fc0f16ba70892a54bee20eb9e4b1ddf18d8f3334fb81fe","reason":"Same source body in authored semantic page"},{"oldId":"b-a401b86ad1bcff4eb7980e3a","targetId":"b-a401b86ad1bcff4eb7980e3a","page":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","operation":"move","beforeSHA256":"dfb749be9ea401298591c13840469fdf0010697e73c244978111ebf419a560a3","afterSHA256":"dfb749be9ea401298591c13840469fdf0010697e73c244978111ebf419a560a3","reason":"Same source body in authored semantic page"},{"oldId":"b-2827633ff499e0da8d457338","targetId":"b-2827633ff499e0da8d457338","page":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","operation":"move","beforeSHA256":"6424f8550f5b5c6628f1965c5c1b4aafa939764425130e58f1a53d555fcca12a","afterSHA256":"6424f8550f5b5c6628f1965c5c1b4aafa939764425130e58f1a53d555fcca12a","reason":"Same source body in authored semantic page"},{"oldId":"b-ffe973d19777952ad08767e0","targetId":"b-ffe973d19777952ad08767e0","page":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","operation":"move","beforeSHA256":"75e182f642c1113669edafb04efc844003f32942bf07649c400ab44b5772c86d","afterSHA256":"75e182f642c1113669edafb04efc844003f32942bf07649c400ab44b5772c86d","reason":"Same source body in authored semantic page"},{"oldId":"b-682a7ed23acdeb685cd33332","targetId":"b-682a7ed23acdeb685cd33332","page":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","operation":"move","beforeSHA256":"79b23bf9a1f450a215537903032b76943a29701344a0725834d43be47d02bdb8","afterSHA256":"79b23bf9a1f450a215537903032b76943a29701344a0725834d43be47d02bdb8","reason":"Same source body in authored semantic page"},{"oldId":"b-132c629cab172fe46126e461","targetId":"b-132c629cab172fe46126e461","page":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","operation":"move","beforeSHA256":"7a070e4079949de63d3f40004e06bc2b44eabd82cfe803d7183c8a4e3aa60118","afterSHA256":"7a070e4079949de63d3f40004e06bc2b44eabd82cfe803d7183c8a4e3aa60118","reason":"Same source body in authored semantic page"},{"oldId":"b-e539cf39a5808921df8c79e5","targetId":"b-e539cf39a5808921df8c79e5","page":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","operation":"move","beforeSHA256":"e58f3bfde63e21eff0229440f4df21ebc4312e7d153980a079fc9492e8714ce5","afterSHA256":"e58f3bfde63e21eff0229440f4df21ebc4312e7d153980a079fc9492e8714ce5","reason":"Same source body in authored semantic page"},{"oldId":"b-76aee3b094abdc7169424ba0","targetId":"b-76aee3b094abdc7169424ba0","page":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","operation":"move","beforeSHA256":"a03704d3c05d5ec2a67c2178b051ac4fdd8b69d54e6f999b8cffb44f26c2335f","afterSHA256":"a03704d3c05d5ec2a67c2178b051ac4fdd8b69d54e6f999b8cffb44f26c2335f","reason":"Same source body in authored semantic page"},{"oldId":"b-b7d3739e5ef0b2b3a27df848","targetId":"b-b7d3739e5ef0b2b3a27df848","page":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","operation":"move","beforeSHA256":"5bc03f71d0e73034dde1ae62b299fce1769bd31220bdeb1b8cbc64c7b96cfa93","afterSHA256":"5bc03f71d0e73034dde1ae62b299fce1769bd31220bdeb1b8cbc64c7b96cfa93","reason":"Same source body in authored semantic page"},{"oldId":"b-810f813400087ff1c2d70a37","targetId":"b-810f813400087ff1c2d70a37","page":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","operation":"move","beforeSHA256":"8017cdceef3018595cd3113e4fc86bca4d857d85b80fd4e531a4badfb1d7c20b","afterSHA256":"8017cdceef3018595cd3113e4fc86bca4d857d85b80fd4e531a4badfb1d7c20b","reason":"Same source body in authored semantic page"},{"oldId":"b-b4a8aeee378e54afae9f4905","targetId":"b-b4a8aeee378e54afae9f4905","page":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","operation":"move","beforeSHA256":"950dd0014bf5ab61cb0bc697072261d2c0637d1caec586a17dcaee2049d756ca","afterSHA256":"950dd0014bf5ab61cb0bc697072261d2c0637d1caec586a17dcaee2049d756ca","reason":"Same source body in authored semantic page"},{"oldId":"block-ed31be25c84b4c4eb511e2b846038ac3--两个编码器一个配对目标","targetId":"b-135d9ab06069fb9299dbbfeb","operation":"retain-source-identity"},{"oldId":"block-ed31be25c84b4c4eb511e2b846038ac3--为什么这是可用的新接口","targetId":"b-930d6b51abdffcd925a659ac","operation":"retain-source-identity"},{"oldId":"block-ed31be25c84b4c4eb511e2b846038ac3--三个干预三个不同问题","targetId":"b-9bd19c6a16a5443435bab304","operation":"retain-source-identity"},{"oldId":"block-b7ab2768ecf04c3aa9e1c6307a4de22b--真实猫图结果保留模型照片与历史身份","targetId":"b-d3d06900a49a6a0a90867f1e","operation":"retain-source-identity"},{"oldId":"block-b7ab2768ecf04c3aa9e1c6307a4de22b--检索为什么能缓存每张图编码一次查询再比较","targetId":"b-70aca0628f9012a0af754378","operation":"retain-source-identity"},{"oldId":"block-b7ab2768ecf04c3aa9e1c6307a4de22b--hitrecall-与-exact-重合率各自在问什么","targetId":"b-2e17c00ccaa0efdcdfcc7553","operation":"retain-source-identity"},{"oldId":"block-b7ab2768ecf04c3aa9e1c6307a4de22b--第八讲预备阅读近似搜索与查询交互","targetId":"b-badbf96a9016a6dfa98db812","operation":"retain-source-identity"},{"oldId":"block-b7ab2768ecf04c3aa9e1c6307a4de22b--真实-clip-关系反例与规则四格应分开读","targetId":"b-132c629cab172fe46126e461","operation":"retain-source-identity"}],"blocks":[{"id":"b-2133e06b58b81a6298c57b3f","type":"h1","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-2133e06b58b81a6298c57b3f","content":"# CLS 10 · CLIP：配对学习与可复用表示","sha256":"271272473744b0e1471941a678bc961da666faaf6e5f5117d03c24ec7bfb1a92","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-opening","title":"CLS 10 · CLIP：配对学习与可复用表示","number":1},"sectionTitle":"CLS 10 · CLIP：配对学习与可复用表示","currentQuestion":"CLS 10 · CLIP：配对学习与可复用表示","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-7272c979a6f6fc7b76ed5483","prerequisites":[{"id":"b-7272c979a6f6fc7b76ed5483","url":"#b-7272c979a6f6fc7b76ed5483","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-7272c979a6f6fc7b76ed5483","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-7272c979a6f6fc7b76ed5483","content":"按问题继续","sha256":"e91f24e4e4ca40e2099a0d81d30db6568f4bf2074b499eeee6e3959617bd97ac","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-opening","title":"CLS 10 · CLIP：配对学习与可复用表示","number":1},"sectionTitle":"CLS 10 · CLIP：配对学习与可复用表示","currentQuestion":"CLS 10 · CLIP：配对学习与可复用表示","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-2133e06b58b81a6298c57b3f","nextBlock":"b-6341975a56cc5ebe45a12980","prerequisites":[{"id":"b-2133e06b58b81a6298c57b3f","url":"#b-2133e06b58b81a6298c57b3f","role":"section-title"}],"relatedBlocks":[]},{"id":"b-6341975a56cc5ebe45a12980","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-6341975a56cc5ebe45a12980","content":"CLIP对称目标、候选集合、缓存、IVF及真实关系分数","sha256":"8b4a38fbe2d21e8c471d155a892cffc7decd7adfc064650236630fb3edf7a7d0","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-opening","title":"CLS 10 · CLIP：配对学习与可复用表示","number":1},"sectionTitle":"CLS 10 · CLIP：配对学习与可复用表示","currentQuestion":"CLS 10 · CLIP：配对学习与可复用表示","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-7272c979a6f6fc7b76ed5483","nextBlock":"b-32ae2b6149fd1928b5ba0b0b","prerequisites":[{"id":"b-2133e06b58b81a6298c57b3f","url":"#b-2133e06b58b81a6298c57b3f","role":"section-title"},{"id":"b-7272c979a6f6fc7b76ed5483","url":"#b-7272c979a6f6fc7b76ed5483","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-32ae2b6149fd1928b5ba0b0b","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-32ae2b6149fd1928b5ba0b0b","content":"可先阅读已有结果，再决定是否运行。重算时请核对各节记录的数据、参数、环境与日期。","sha256":"f22152fd55dcdb0090b5653d37145694fd19388d31e10226549464fe8b483578","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-opening","title":"CLS 10 · CLIP：配对学习与可复用表示","number":1},"sectionTitle":"CLS 10 · CLIP：配对学习与可复用表示","currentQuestion":"CLS 10 · CLIP：配对学习与可复用表示","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-6341975a56cc5ebe45a12980","nextBlock":null,"prerequisites":[{"id":"b-2133e06b58b81a6298c57b3f","url":"#b-2133e06b58b81a6298c57b3f","role":"section-title"},{"id":"b-7272c979a6f6fc7b76ed5483","url":"#b-7272c979a6f6fc7b76ed5483","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-c6b7904edcdfc53b19c45331","type":"p","scope":"block-ed31be25c84b4c4eb511e2b846038ac3","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-c6b7904edcdfc53b19c45331","content":"**学习路径：** [文字候选与图文检索](/course/study/clip-retrieval.html)","sha256":"bab4bd804571aa5afbaf197e9c1539510fc766578c9825aa46cadbb18444cc32","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-ed31be25c84b4c4eb511e2b846038ac3","title":"两个编码器，一个配对目标","number":2},"sectionTitle":"两个编码器，一个配对目标","currentQuestion":"两个编码器，一个配对目标","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-8c4e14f696186a7250637b50","prerequisites":[],"relatedBlocks":[]},{"id":"b-8c4e14f696186a7250637b50","type":"p","scope":"block-ed31be25c84b4c4eb511e2b846038ac3","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-8c4e14f696186a7250637b50","content":"从[P57 配对矩阵](/course/pages/l3-r2-p57.html)进入，沿着[P58 文字候选](/course/pages/l3-r2-p58.html)到[P59 图库缓存](/course/pages/l3-r2-p59.html)：同一图文表示怎样从训练目标变成可用接口？","sha256":"50590b614b01aacf740b6f59f0a25be53dce523d2843825124cc6e1195f0f09c","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-ed31be25c84b4c4eb511e2b846038ac3","title":"两个编码器，一个配对目标","number":2},"sectionTitle":"两个编码器，一个配对目标","currentQuestion":"两个编码器，一个配对目标","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-c6b7904edcdfc53b19c45331","nextBlock":"b-fc0851c6deeffcdc567be124","prerequisites":[{"id":"b-c6b7904edcdfc53b19c45331","url":"#b-c6b7904edcdfc53b19c45331","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-fc0851c6deeffcdc567be124","type":"p","scope":"block-ed31be25c84b4c4eb511e2b846038ac3","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-fc0851c6deeffcdc567be124","content":"这里的zero-shot是**不使用当前下游数据集再训练模型，直接用已有图像和文字编码器比较候选描述**。不保证相关概念未进入预训练；没有专用分类头也不足以单独证明这个协议成立。","sha256":"2e7f9ca1ab75355dcb74d4a3f8cd95d8ec9fc0ddc39e8919ee0ca1d8fa2e2052","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-ed31be25c84b4c4eb511e2b846038ac3","title":"两个编码器，一个配对目标","number":2},"sectionTitle":"两个编码器，一个配对目标","currentQuestion":"两个编码器，一个配对目标","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-8c4e14f696186a7250637b50","nextBlock":"b-447dcc8027bd54aac453b630","prerequisites":[{"id":"b-c6b7904edcdfc53b19c45331","url":"#b-c6b7904edcdfc53b19c45331","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-447dcc8027bd54aac453b630","type":"p","scope":"block-ed31be25c84b4c4eb511e2b846038ac3","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-447dcc8027bd54aac453b630","content":"先看配对损失与候选分数，再到[LAB 04](https://deepnote.com/project/38755e4e-23f6-4f5f-8c68-8a2e4114b1c1/notebook/0ba963931ddf4866b507b210d93a43bf)核对真实照片的cosine/softmax。关心检索时继续[LAB 12](https://deepnote.com/project/38755e4e-23f6-4f5f-8c68-8a2e4114b1c1/notebook/921eeaf784284ce19568c6ca7619b55a)的缓存、命中ID和评价；近似索引及查询交互留第八讲选读。","sha256":"2b31d6a9ce476da9fbc54ff1cd84e4d23c0a209eef12ce6ab7054f0dfa6f5c76","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-ed31be25c84b4c4eb511e2b846038ac3","title":"两个编码器，一个配对目标","number":2},"sectionTitle":"两个编码器，一个配对目标","currentQuestion":"两个编码器，一个配对目标","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-fc0851c6deeffcdc567be124","nextBlock":"b-95667271fd87b3983c5ccfca","prerequisites":[{"id":"b-c6b7904edcdfc53b19c45331","url":"#b-c6b7904edcdfc53b19c45331","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-95667271fd87b3983c5ccfca","type":"p","scope":"block-ed31be25c84b4c4eb511e2b846038ac3","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-95667271fd87b3983c5ccfca","content":"读完应能区分表示质量、候选定义和索引误差：一个返回结果错了，是模型没排对，还是近似检索漏掉了本来排前的项？","sha256":"bfed64de92e5326d9f2ea08a5fbf216a9101f67d42e71710906c6ab6d2692320","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-ed31be25c84b4c4eb511e2b846038ac3","title":"两个编码器，一个配对目标","number":2},"sectionTitle":"两个编码器，一个配对目标","currentQuestion":"两个编码器，一个配对目标","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-447dcc8027bd54aac453b630","nextBlock":"b-135d9ab06069fb9299dbbfeb","prerequisites":[{"id":"b-c6b7904edcdfc53b19c45331","url":"#b-c6b7904edcdfc53b19c45331","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-135d9ab06069fb9299dbbfeb","type":"h2","scope":"block-ed31be25c84b4c4eb511e2b846038ac3","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-135d9ab06069fb9299dbbfeb","content":"## 两个编码器，一个配对目标","sha256":"063457994938d4e5dfe61567141bd163cb7d7cce6959928c1c0eb5510469ab6b","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-ed31be25c84b4c4eb511e2b846038ac3","title":"两个编码器，一个配对目标","number":2},"sectionTitle":"两个编码器，一个配对目标","currentQuestion":"两个编码器，一个配对目标","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-95667271fd87b3983c5ccfca","nextBlock":"b-7bb2ed61498b0b916e947855","prerequisites":[{"id":"b-c6b7904edcdfc53b19c45331","url":"#b-c6b7904edcdfc53b19c45331","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-7bb2ed61498b0b916e947855","type":"p","scope":"block-ed31be25c84b4c4eb511e2b846038ac3","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-7bb2ed61498b0b916e947855","content":"图像与文字分别归一化为 $v_i,u_j$，矩阵 $S_{ij}=v_i^Tu_j$ 同时支持图搜文与文搜图。双向交叉熵让原有配对相对其他候选更容易被选出。课堂三对图文的完整数值见 [三对图文的双向损失：从单位向量到 0.09567](/course/blocks/clip-matrix-worked.html)：温度 .2，正确概率约 .90876，双向损失约 .09567。CLS03 热身采用另一组向量与温度，不能混用两套数值。","sha256":"421bb86160826d098ebd40ff162672a976ef2d422bd0fce72fa666c4eaad7d3e","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-ed31be25c84b4c4eb511e2b846038ac3","title":"两个编码器，一个配对目标","number":2},"sectionTitle":"两个编码器，一个配对目标","currentQuestion":"两个编码器，一个配对目标","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"v_i,u_j","definitionContext":"两个编码器，一个配对目标"},{"latex":"S_{ij}=v_i^Tu_j","definitionContext":"两个编码器，一个配对目标"}],"previousBlock":"b-135d9ab06069fb9299dbbfeb","nextBlock":"b-930d6b51abdffcd925a659ac","prerequisites":[{"id":"b-135d9ab06069fb9299dbbfeb","url":"#b-135d9ab06069fb9299dbbfeb","role":"section-title"},{"id":"b-c6b7904edcdfc53b19c45331","url":"#b-c6b7904edcdfc53b19c45331","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-930d6b51abdffcd925a659ac","type":"h2","scope":"block-ed31be25c84b4c4eb511e2b846038ac3","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-930d6b51abdffcd925a659ac","content":"## 为什么这是可用的新接口","sha256":"eba1b58a7fce19f59185999c38f5f654b59946a33ae131684e1abbc607f712e6","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-ed31be25c84b4c4eb511e2b846038ac3","title":"两个编码器，一个配对目标","number":2},"sectionTitle":"两个编码器，一个配对目标","currentQuestion":"两个编码器，一个配对目标","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-7bb2ed61498b0b916e947855","nextBlock":"b-7bb9e12f4dfb95950657ed76","prerequisites":[{"id":"b-135d9ab06069fb9299dbbfeb","url":"#b-135d9ab06069fb9299dbbfeb","role":"section-title"},{"id":"b-c6b7904edcdfc53b19c45331","url":"#b-c6b7904edcdfc53b19c45331","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-7bb9e12f4dfb95950657ed76","type":"p","scope":"block-ed31be25c84b4c4eb511e2b846038ac3","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-7bb9e12f4dfb95950657ed76","content":"原始 CLIP 研究展示自然语言监督能迁移到多种视觉任务，并用文字定义当前分类候选；无需为每次候选集合重训一个固定类别头。[论文中的正面迁移证据](/course/blocks/clip-positive-transfer-evidence.html)。它留下的是图文比较接口，描述生成与定位仍需相应输出。","sha256":"6b5dab1b128812a45c0237b9943cd89a557634629a8386260132bba6eb7d7a14","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-ed31be25c84b4c4eb511e2b846038ac3","title":"两个编码器，一个配对目标","number":2},"sectionTitle":"两个编码器，一个配对目标","currentQuestion":"两个编码器，一个配对目标","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-930d6b51abdffcd925a659ac","nextBlock":"b-9bd19c6a16a5443435bab304","prerequisites":[{"id":"b-930d6b51abdffcd925a659ac","url":"#b-930d6b51abdffcd925a659ac","role":"section-title"},{"id":"b-c6b7904edcdfc53b19c45331","url":"#b-c6b7904edcdfc53b19c45331","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-9bd19c6a16a5443435bab304","type":"h2","scope":"block-ed31be25c84b4c4eb511e2b846038ac3","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-9bd19c6a16a5443435bab304","content":"## 三个干预，三个不同问题","sha256":"b82eb99f0b75c3c6d07cae11fdb8041ed3871643065e2b41ebf2f1d956f99de4","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-ed31be25c84b4c4eb511e2b846038ac3","title":"两个编码器，一个配对目标","number":2},"sectionTitle":"两个编码器，一个配对目标","currentQuestion":"两个编码器，一个配对目标","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-7bb9e12f4dfb95950657ed76","nextBlock":"b-9d936330364cc5aa6e283082","prerequisites":[{"id":"b-930d6b51abdffcd925a659ac","url":"#b-930d6b51abdffcd925a659ac","role":"section-title"},{"id":"b-c6b7904edcdfc53b19c45331","url":"#b-c6b7904edcdfc53b19c45331","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-9d936330364cc5aa6e283082","type":"p","scope":"block-ed31be25c84b4c4eb511e2b846038ac3","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-9d936330364cc5aa6e283082","content":"固定图像和候选文字，只新增选项：原余弦可保持，但 softmax 分母改变。改模板：文本向量改变，余弦也可能改变。改图片：视觉向量改变，应检查分数是否跟随所需视觉内容。","sha256":"eeb43459eb9f075a95c7403ff26d52e85c14cf632e9eef7d81193d82eb7eae86","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-ed31be25c84b4c4eb511e2b846038ac3","title":"两个编码器，一个配对目标","number":2},"sectionTitle":"两个编码器，一个配对目标","currentQuestion":"两个编码器，一个配对目标","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-9bd19c6a16a5443435bab304","nextBlock":"b-8c31ad492335867d64c520a2","prerequisites":[{"id":"b-9bd19c6a16a5443435bab304","url":"#b-9bd19c6a16a5443435bab304","role":"section-title"},{"id":"b-c6b7904edcdfc53b19c45331","url":"#b-c6b7904edcdfc53b19c45331","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-8c31ad492335867d64c520a2","type":"p","scope":"block-ed31be25c84b4c4eb511e2b846038ac3","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-8c31ad492335867d64c520a2","content":"若猫/狗 logits 为 2/1，猫概率为 .7311；添加 logit 为2的“幼猫”，原猫项变为 .4223。两次图像与原猫文字并没有变。候选内概率不能单独解读为开放世界真实性。[完整算例](/course/blocks/clip-candidate-probability.html)。","sha256":"e2fd6ed94ff5812ee027b0c1d2a3d46127eb13403d0e3f6c267ac4ef37f49552","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-ed31be25c84b4c4eb511e2b846038ac3","title":"两个编码器，一个配对目标","number":2},"sectionTitle":"两个编码器，一个配对目标","currentQuestion":"两个编码器，一个配对目标","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-9d936330364cc5aa6e283082","nextBlock":"b-76475b87f615c444a5940365","prerequisites":[{"id":"b-9bd19c6a16a5443435bab304","url":"#b-9bd19c6a16a5443435bab304","role":"section-title"},{"id":"b-c6b7904edcdfc53b19c45331","url":"#b-c6b7904edcdfc53b19c45331","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-76475b87f615c444a5940365","type":"p","scope":"block-ed31be25c84b4c4eb511e2b846038ac3","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-76475b87f615c444a5940365","content":"LAB04 的真实猫图已保存上述接口对照。LAB12 使用完整 512 维编码，不是把候选分数拼成向量。两本共享模型身份、各答不同问题：前者解释模型与候选，后者解释缓存和搜索。","sha256":"38c3c7217ff456579df20fe9c5703f47cfc3e63d9dcbb6bd611964eef7aa9ad0","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-ed31be25c84b4c4eb511e2b846038ac3","title":"两个编码器，一个配对目标","number":2},"sectionTitle":"两个编码器，一个配对目标","currentQuestion":"两个编码器，一个配对目标","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-8c31ad492335867d64c520a2","nextBlock":null,"prerequisites":[{"id":"b-9bd19c6a16a5443435bab304","url":"#b-9bd19c6a16a5443435bab304","role":"section-title"},{"id":"b-c6b7904edcdfc53b19c45331","url":"#b-c6b7904edcdfc53b19c45331","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-d3d06900a49a6a0a90867f1e","type":"h2","scope":"block-b7ab2768ecf04c3aa9e1c6307a4de22b","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-d3d06900a49a6a0a90867f1e","content":"## 真实猫图结果：保留模型、照片与历史身份","sha256":"b81b865ed7c96a04916ece8e45b5f513d6d48a0b95cc2731a46f2c60802de664","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","title":"真实猫图结果：保留模型、照片与历史身份","number":3},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"真实猫图结果：保留模型、照片与历史身份","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-5857325209e2365e5dae8218","prerequisites":[{"id":"b-5857325209e2365e5dae8218","url":"#b-5857325209e2365e5dae8218","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-5857325209e2365e5dae8218","type":"p","scope":"block-b7ab2768ecf04c3aa9e1c6307a4de22b","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-5857325209e2365e5dae8218","content":"当前 LAB 04 保存真实 Chelsea 猫照片与 OpenAI CLIP ViT-B/32 实验。2026-09-22 的本地复核使用固定 checkpoint revision `3d74acf9a28c67741b2f4f2ea7635f0aaf6f0268`，没有重新训练 CLIP；当前整理也没有重跑该推理。图片先转 RGB、缩放并中心裁剪到 $224\\times224$，再按模型配置归一化。","sha256":"3a74a62e574be11b30f988dbb15451bb02c361669a9eefdaa6127d8a74bad6f5","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","title":"真实猫图结果：保留模型、照片与历史身份","number":3},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"真实猫图结果：保留模型、照片与历史身份","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"224\\times224","definitionContext":"真实猫图结果：保留模型、照片与历史身份"}],"previousBlock":"b-d3d06900a49a6a0a90867f1e","nextBlock":"b-9d65fc5ff6b28cfbf499b2e1","prerequisites":[{"id":"b-d3d06900a49a6a0a90867f1e","url":"#b-d3d06900a49a6a0a90867f1e","role":"section-title"}],"relatedBlocks":[]},{"id":"b-9d65fc5ff6b28cfbf499b2e1","type":"p","scope":"block-b7ab2768ecf04c3aa9e1c6307a4de22b","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-9d65fc5ff6b28cfbf499b2e1","content":"基础 cat/dog/horse 候选下，cat 余弦约 0.28691、候选 softmax 约 0.99577；添加 tiger 与 fox 后，共享 cat 的余弦保持不变，softmax 降至约 0.98195；再删除 cat，tiger 成为剩余候选的最高项。新旧猫图余弦最大差约 $1.04\\times10^{-7}$，把后来的复核与历史材料连接起来。","sha256":"c7061a8ad0794772fc5fd4b5f0bdb6e66e098c0aaa2732e4c041ff8bb739810e","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","title":"真实猫图结果：保留模型、照片与历史身份","number":3},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"真实猫图结果：保留模型、照片与历史身份","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"1.04\\times10^{-7}","definitionContext":"真实猫图结果：保留模型、照片与历史身份"}],"previousBlock":"b-5857325209e2365e5dae8218","nextBlock":"b-adc2df8499b31659b5a5bec2","prerequisites":[{"id":"b-d3d06900a49a6a0a90867f1e","url":"#b-d3d06900a49a6a0a90867f1e","role":"section-title"},{"id":"b-5857325209e2365e5dae8218","url":"#b-5857325209e2365e5dae8218","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-adc2df8499b31659b5a5bec2","type":"p","scope":"block-b7ab2768ecf04c3aa9e1c6307a4de22b","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-adc2df8499b31659b5a5bec2","content":"正面结果是模型可以在这张真实照片上把图像与合理文字对应，并且其候选接口的变化符合预期。数值 0.99577 是给定候选集合内的相对分配，不是“现实中有猫的校准概率”。一张猫图也不能替代数据集级泛化评测。","sha256":"1fb9bd54f6a6d811bd1b68919726fd1177e6a5264ddb1afa54180876084827ba","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","title":"真实猫图结果：保留模型、照片与历史身份","number":3},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"真实猫图结果：保留模型、照片与历史身份","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-9d65fc5ff6b28cfbf499b2e1","nextBlock":"b-d249a93f276ebd6170824bdf","prerequisites":[{"id":"b-d3d06900a49a6a0a90867f1e","url":"#b-d3d06900a49a6a0a90867f1e","role":"section-title"},{"id":"b-5857325209e2365e5dae8218","url":"#b-5857325209e2365e5dae8218","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-d249a93f276ebd6170824bdf","type":"p","scope":"block-b7ab2768ecf04c3aa9e1c6307a4de22b","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-d249a93f276ebd6170824bdf","content":"LAB 04 负责模型输入、候选、模板和关系检查；LAB 12 复用同一模型产生的完整 512 维向量，研究图库缓存与搜索。不要把四个候选分数拼成所谓图像表示；真正可复用的是编码器输出的向量。先在 LAB 04 理解分数身份，再在 LAB 12 看系统组织，就无需重复训练或建立两份同义实验。","sha256":"24e101bfca64bca25487af0223ddf64733850ac2f198d1326f78447406f724ff","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","title":"真实猫图结果：保留模型、照片与历史身份","number":3},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"真实猫图结果：保留模型、照片与历史身份","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-adc2df8499b31659b5a5bec2","nextBlock":"b-e30f6b010a6a48ccdaf011be","prerequisites":[{"id":"b-d3d06900a49a6a0a90867f1e","url":"#b-d3d06900a49a6a0a90867f1e","role":"section-title"},{"id":"b-5857325209e2365e5dae8218","url":"#b-5857325209e2365e5dae8218","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-e30f6b010a6a48ccdaf011be","type":"p","scope":"block-b7ab2768ecf04c3aa9e1c6307a4de22b","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-e30f6b010a6a48ccdaf011be","content":"[本节固定网页 URL](/course/blocks/clip-real-checkpoint-evidence.html)","sha256":"fa47d13f4375702edee587849800946fb46b0cb96bc5dfacf00c80ad1dd47c44","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","title":"真实猫图结果：保留模型、照片与历史身份","number":3},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"真实猫图结果：保留模型、照片与历史身份","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-d249a93f276ebd6170824bdf","nextBlock":"b-70aca0628f9012a0af754378","prerequisites":[{"id":"b-d3d06900a49a6a0a90867f1e","url":"#b-d3d06900a49a6a0a90867f1e","role":"section-title"},{"id":"b-5857325209e2365e5dae8218","url":"#b-5857325209e2365e5dae8218","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-70aca0628f9012a0af754378","type":"h2","scope":"block-b7ab2768ecf04c3aa9e1c6307a4de22b","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-70aca0628f9012a0af754378","content":"## 检索为什么能缓存：每张图编码一次，查询再比较","sha256":"0d2f7a9737f9a95c5d7150d1ce3364baa7d3b82559a2be501d1690dcd74647f4","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","title":"真实猫图结果：保留模型、照片与历史身份","number":3},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"真实猫图结果：保留模型、照片与历史身份","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-e30f6b010a6a48ccdaf011be","nextBlock":"b-02e48d56c1938c354a142f24","prerequisites":[{"id":"b-d3d06900a49a6a0a90867f1e","url":"#b-d3d06900a49a6a0a90867f1e","role":"section-title"},{"id":"b-5857325209e2365e5dae8218","url":"#b-5857325209e2365e5dae8218","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-02e48d56c1938c354a142f24","type":"p","scope":"block-b7ab2768ecf04c3aa9e1c6307a4de22b","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-02e48d56c1938c354a142f24","content":"设图库有 $N$ 张图，收到 $Q$ 条文字查询。双编码器的图像输出不依赖当前查询，所以可以先计算图库矩阵 $V\\in\\mathbb R^{N\\times d}$，查询到来时只计算文字向量 $u$，再用 $Vu$ 得分。这把“理解输入的深度编码”和“在现有表示中比较”分成两段。","sha256":"0d52d45b91d64683943b3870efe42eb2211cb249baaa5d700568f83f16ac824f","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","title":"真实猫图结果：保留模型、照片与历史身份","number":3},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"真实猫图结果：保留模型、照片与历史身份","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"N","definitionContext":"真实猫图结果：保留模型、照片与历史身份"},{"latex":"Q","definitionContext":"真实猫图结果：保留模型、照片与历史身份"},{"latex":"V\\in\\mathbb R^{N\\times d}","definitionContext":"真实猫图结果：保留模型、照片与历史身份"},{"latex":"u","definitionContext":"真实猫图结果：保留模型、照片与历史身份"},{"latex":"Vu","definitionContext":"真实猫图结果：保留模型、照片与历史身份"}],"previousBlock":"b-70aca0628f9012a0af754378","nextBlock":"b-fa4fdd5898ca2d21852cf336","prerequisites":[{"id":"b-70aca0628f9012a0af754378","url":"#b-70aca0628f9012a0af754378","role":"section-title"},{"id":"b-5857325209e2365e5dae8218","url":"#b-5857325209e2365e5dae8218","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-fa4fdd5898ca2d21852cf336","type":"p","scope":"block-b7ab2768ecf04c3aa9e1c6307a4de22b","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-fa4fdd5898ca2d21852cf336","content":"若每个图文对都重新分别编码，需 $NQ$ 次图片样本编码和 $NQ$ 次文字样本编码；有缓存则是 $N+Q$。本课程 $N=24,Q=12$，两者为 576 与 36 次样本编码。它是解析工作量，不是秒数；一个批次调用可以同时编码多个样本。","sha256":"548903116868e99139fe6cb659aa140a85b965dcc46a1ad6a79904bd982e3beb","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","title":"真实猫图结果：保留模型、照片与历史身份","number":3},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"真实猫图结果：保留模型、照片与历史身份","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"NQ","definitionContext":"真实猫图结果：保留模型、照片与历史身份"},{"latex":"NQ","definitionContext":"真实猫图结果：保留模型、照片与历史身份"},{"latex":"N+Q","definitionContext":"真实猫图结果：保留模型、照片与历史身份"},{"latex":"N=24,Q=12","definitionContext":"真实猫图结果：保留模型、照片与历史身份"}],"previousBlock":"b-02e48d56c1938c354a142f24","nextBlock":"b-c941b94f48d71cc3fbf154f1","prerequisites":[{"id":"b-70aca0628f9012a0af754378","url":"#b-70aca0628f9012a0af754378","role":"section-title"},{"id":"b-5857325209e2365e5dae8218","url":"#b-5857325209e2365e5dae8218","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-c941b94f48d71cc3fbf154f1","type":"p","scope":"block-b7ab2768ecf04c3aa9e1c6307a4de22b","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-c941b94f48d71cc3fbf154f1","content":"缓存并没有消除搜索本身。exact 查询仍要与全部 $N$ 向量做点积；float32 图库矩阵占 $4Nd$ 字节。本例 $24\\times512\\times4=49,152$ 字节，不含图片、模型、ID 与索引。改变 checkpoint、预处理或图片内容需要重建对应缓存；只新增文字候选不改变旧图片向量。","sha256":"a0b9c4c8d9eebe422e20a0d9595195dc58efc43923716e264bd41dc0bfb60c5d","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","title":"真实猫图结果：保留模型、照片与历史身份","number":3},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"真实猫图结果：保留模型、照片与历史身份","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"N","definitionContext":"真实猫图结果：保留模型、照片与历史身份"},{"latex":"4Nd","definitionContext":"真实猫图结果：保留模型、照片与历史身份"},{"latex":"24\\times512\\times4=49,152","definitionContext":"真实猫图结果：保留模型、照片与历史身份"}],"previousBlock":"b-fa4fdd5898ca2d21852cf336","nextBlock":"b-2602b161f592612b6f37f89b","prerequisites":[{"id":"b-70aca0628f9012a0af754378","url":"#b-70aca0628f9012a0af754378","role":"section-title"},{"id":"b-5857325209e2365e5dae8218","url":"#b-5857325209e2365e5dae8218","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-2602b161f592612b6f37f89b","type":"p","scope":"block-b7ab2768ecf04c3aa9e1c6307a4de22b","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-2602b161f592612b6f37f89b","content":"P59 使用已有 24 项缓存显示猫查询的前三项：`chelsea-center90`、`chelsea-full`、`chelsea-right90`。24 项其实来自六张照片的四种相关裁剪，不能当成 24 张独立语义样本。它足以观察排名与缓存复用；更大的索引工程属于第八讲和 LAB 12 的选读部分。","sha256":"5a2ad5d4eb4c55adef619a9d7099e3489a0df2782017a87a2a4593130fcb2eaf","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","title":"真实猫图结果：保留模型、照片与历史身份","number":3},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"真实猫图结果：保留模型、照片与历史身份","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-c941b94f48d71cc3fbf154f1","nextBlock":"b-6f5b6cb742b0ec0e2a0effc9","prerequisites":[{"id":"b-70aca0628f9012a0af754378","url":"#b-70aca0628f9012a0af754378","role":"section-title"},{"id":"b-5857325209e2365e5dae8218","url":"#b-5857325209e2365e5dae8218","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-6f5b6cb742b0ec0e2a0effc9","type":"p","scope":"block-b7ab2768ecf04c3aa9e1c6307a4de22b","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-6f5b6cb742b0ec0e2a0effc9","content":"[本节固定网页 URL](/course/blocks/retrieval-cache-interface.html)","sha256":"e4d7a3e0afa1d5e068c5eb162d5fae7ff4cdb68424305df478e479d6de6d3cbc","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","title":"真实猫图结果：保留模型、照片与历史身份","number":3},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"真实猫图结果：保留模型、照片与历史身份","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-2602b161f592612b6f37f89b","nextBlock":"b-2e17c00ccaa0efdcdfcc7553","prerequisites":[{"id":"b-70aca0628f9012a0af754378","url":"#b-70aca0628f9012a0af754378","role":"section-title"},{"id":"b-5857325209e2365e5dae8218","url":"#b-5857325209e2365e5dae8218","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-2e17c00ccaa0efdcdfcc7553","type":"h2","scope":"block-b7ab2768ecf04c3aa9e1c6307a4de22b","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-2e17c00ccaa0efdcdfcc7553","content":"## Hit、Recall 与 exact 重合率，各自在问什么","sha256":"c54f1c8eedd49ebd3230440feaf9249ecf82351a2c3ac32e70e43f150b45a031","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","title":"真实猫图结果：保留模型、照片与历史身份","number":3},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"真实猫图结果：保留模型、照片与历史身份","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-6f5b6cb742b0ec0e2a0effc9","nextBlock":"b-7a411282657df1acb7cb8644","prerequisites":[{"id":"b-70aca0628f9012a0af754378","url":"#b-70aca0628f9012a0af754378","role":"section-title"},{"id":"b-5857325209e2365e5dae8218","url":"#b-5857325209e2365e5dae8218","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-7a411282657df1acb7cb8644","type":"p","scope":"block-b7ab2768ecf04c3aa9e1c6307a4de22b","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-7a411282657df1acb7cb8644","content":"固定一个查询，相关集合为 $R(q)$，检索前 $K$ 个结果为 $A_K(q)$。三个常见指标不是同义词：Hit@K 问是否至少命中一个相关结果；Recall@K 为 $|A_K\\cap R|/|R|$；ANN overlap 问近似搜索是否复现精确搜索，计算 $|A_K\\cap E_K|/K$，其中 $E_K$ 为 exact 前 K。","sha256":"7d1e522bf167385f95073a108d6651706c7ab2f25d134521e17b40fe2992bc59","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","title":"真实猫图结果：保留模型、照片与历史身份","number":3},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"真实猫图结果：保留模型、照片与历史身份","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"R(q)","definitionContext":"真实猫图结果：保留模型、照片与历史身份"},{"latex":"K","definitionContext":"真实猫图结果：保留模型、照片与历史身份"},{"latex":"A_K(q)","definitionContext":"真实猫图结果：保留模型、照片与历史身份"},{"latex":"|A_K\\cap R|/|R|","definitionContext":"真实猫图结果：保留模型、照片与历史身份"},{"latex":"|A_K\\cap E_K|/K","definitionContext":"真实猫图结果：保留模型、照片与历史身份"},{"latex":"E_K","definitionContext":"真实猫图结果：保留模型、照片与历史身份"}],"previousBlock":"b-2e17c00ccaa0efdcdfcc7553","nextBlock":"b-39ced42e12636a5a74bc42ea","prerequisites":[{"id":"b-2e17c00ccaa0efdcdfcc7553","url":"#b-2e17c00ccaa0efdcdfcc7553","role":"section-title"},{"id":"b-5857325209e2365e5dae8218","url":"#b-5857325209e2365e5dae8218","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-39ced42e12636a5a74bc42ea","type":"p","scope":"block-b7ab2768ecf04c3aa9e1c6307a4de22b","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-39ced42e12636a5a74bc42ea","content":"假设某猫查询有四个猫裁剪算相关，前三名命中两个。Hit@3 是 1，Recall@3 是 $2/4=0.5$。若 exact 的前三名是 a,b,c，ANN 找到 a,b,d，二者 overlap 是 $2/3$；d 即使仍然是猫，语义命中也可能不变。模型相关性与索引逼近的是两个不同问题。","sha256":"e60979fc29d0931ab0fa6565feb357a2b9d204702fac523f27809082dbf318a7","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","title":"真实猫图结果：保留模型、照片与历史身份","number":3},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"真实猫图结果：保留模型、照片与历史身份","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"2/4=0.5","definitionContext":"真实猫图结果：保留模型、照片与历史身份"},{"latex":"2/3","definitionContext":"真实猫图结果：保留模型、照片与历史身份"}],"previousBlock":"b-7a411282657df1acb7cb8644","nextBlock":"b-913e50be755c439d698f1327","prerequisites":[{"id":"b-2e17c00ccaa0efdcdfcc7553","url":"#b-2e17c00ccaa0efdcdfcc7553","role":"section-title"},{"id":"b-5857325209e2365e5dae8218","url":"#b-5857325209e2365e5dae8218","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-913e50be755c439d698f1327","type":"p","scope":"block-b7ab2768ecf04c3aa9e1c6307a4de22b","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-913e50be755c439d698f1327","content":"LAB 12 的历史结果正体现这个区别：exact 文搜图 Hit@3=1；IVF 只查一个簇时 exact top-3 重合率为 0.7778，语义 Hit@3 仍为 1。它漏掉部分同源裁剪，却仍保留某个相关裁剪。这不是指标互相矛盾，而是各自回答了自己的问题。","sha256":"44fa60a15468578fc2e7b4d15edcd2f5936efdbc31c698e97d1d17c94fdf3a93","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","title":"真实猫图结果：保留模型、照片与历史身份","number":3},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"真实猫图结果：保留模型、照片与历史身份","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-39ced42e12636a5a74bc42ea","nextBlock":"b-5a01031cf9b76362130c983c","prerequisites":[{"id":"b-2e17c00ccaa0efdcdfcc7553","url":"#b-2e17c00ccaa0efdcdfcc7553","role":"section-title"},{"id":"b-5857325209e2365e5dae8218","url":"#b-5857325209e2365e5dae8218","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-5a01031cf9b76362130c983c","type":"p","scope":"block-b7ab2768ecf04c3aa9e1c6307a4de22b","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-5a01031cf9b76362130c983c","content":"阅读结果先找查询集合、相关集合和 K，再看成绩。若一个查询只有一个正例，Hit@K 与单正例 Recall@K 恰好同值；换成多正例后必须重新写清口径。保存查询和命中 ID 比只保存平均数字更便于 AI 或同学追溯解释。","sha256":"c1afda030628985fb13e6d5c22f9cfe7d7b0fe2cab59e7e3f191abb1dcef9feb","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","title":"真实猫图结果：保留模型、照片与历史身份","number":3},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"真实猫图结果：保留模型、照片与历史身份","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-913e50be755c439d698f1327","nextBlock":"b-fbf2fd28eec78b34ef0b44b5","prerequisites":[{"id":"b-2e17c00ccaa0efdcdfcc7553","url":"#b-2e17c00ccaa0efdcdfcc7553","role":"section-title"},{"id":"b-5857325209e2365e5dae8218","url":"#b-5857325209e2365e5dae8218","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-fbf2fd28eec78b34ef0b44b5","type":"p","scope":"block-b7ab2768ecf04c3aa9e1c6307a4de22b","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-fbf2fd28eec78b34ef0b44b5","content":"[本节固定网页 URL](/course/blocks/retrieval-metric-contract.html)","sha256":"053e231943f3ce6c919967bff3b098a25f5d7c5ed258867bf88b6fda5464e1f2","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","title":"真实猫图结果：保留模型、照片与历史身份","number":3},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"真实猫图结果：保留模型、照片与历史身份","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-5a01031cf9b76362130c983c","nextBlock":"b-badbf96a9016a6dfa98db812","prerequisites":[{"id":"b-2e17c00ccaa0efdcdfcc7553","url":"#b-2e17c00ccaa0efdcdfcc7553","role":"section-title"},{"id":"b-5857325209e2365e5dae8218","url":"#b-5857325209e2365e5dae8218","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-badbf96a9016a6dfa98db812","type":"h2","scope":"block-b7ab2768ecf04c3aa9e1c6307a4de22b","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-badbf96a9016a6dfa98db812","content":"## 第八讲预备阅读：近似搜索与查询交互","sha256":"fedaaf52a99a89fdf8983619b82d1d3f90c7f2425a6c4494e17d99a8fd97c48f","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","title":"真实猫图结果：保留模型、照片与历史身份","number":3},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"真实猫图结果：保留模型、照片与历史身份","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-fbf2fd28eec78b34ef0b44b5","nextBlock":"b-34ab49a4fd1511ac5bdc564c","prerequisites":[{"id":"b-2e17c00ccaa0efdcdfcc7553","url":"#b-2e17c00ccaa0efdcdfcc7553","role":"section-title"},{"id":"b-5857325209e2365e5dae8218","url":"#b-5857325209e2365e5dae8218","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-34ab49a4fd1511ac5bdc564c","type":"p","scope":"block-b7ab2768ecf04c3aa9e1c6307a4de22b","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-34ab49a4fd1511ac5bdc564c","content":"这部分服务于第八讲的检索预备阅读，不是 P59 必须讲完的技术章。LAB 12 的 IVF-Flat 先把图库分成 8 个簇，查询只打开 `nprobe` 个最近簇，再对桶内原向量精确打分。近似出现在没有打开的桶；把所有 8 个桶打开应恢复 exact 排名。","sha256":"ab2e895629a39a47c6fc0f16ba70892a54bee20eb9e4b1ddf18d8f3334fb81fe","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","title":"真实猫图结果：保留模型、照片与历史身份","number":3},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"真实猫图结果：保留模型、照片与历史身份","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-badbf96a9016a6dfa98db812","nextBlock":"b-a401b86ad1bcff4eb7980e3a","prerequisites":[{"id":"b-badbf96a9016a6dfa98db812","url":"#b-badbf96a9016a6dfa98db812","role":"section-title"},{"id":"b-5857325209e2365e5dae8218","url":"#b-5857325209e2365e5dae8218","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-a401b86ad1bcff4eb7980e3a","type":"p","scope":"block-b7ab2768ecf04c3aa9e1c6307a4de22b","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-a401b86ad1bcff4eb7980e3a","content":"2026-09-22 保存的本地结果中，nprobe=1 平均扫描 4.17 个图像向量，exact 扫描 24 个；但整批 12 查询中位耗时为约 0.1016 ms 与 0.0456 ms。小库矩阵乘法已经很便宜，Python 路由和排序成本反而更大。扫描更少不等于实测更快。计时排除模型编码、磁盘读取和建索引。","sha256":"dfb749be9ea401298591c13840469fdf0010697e73c244978111ebf419a560a3","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","title":"真实猫图结果：保留模型、照片与历史身份","number":3},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"真实猫图结果：保留模型、照片与历史身份","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-34ab49a4fd1511ac5bdc564c","nextBlock":"b-2827633ff499e0da8d457338","prerequisites":[{"id":"b-badbf96a9016a6dfa98db812","url":"#b-badbf96a9016a6dfa98db812","role":"section-title"},{"id":"b-5857325209e2365e5dae8218","url":"#b-5857325209e2365e5dae8218","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-2827633ff499e0da8d457338","type":"p","scope":"block-b7ab2768ecf04c3aa9e1c6307a4de22b","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-2827633ff499e0da8d457338","content":"另一个方向是让查询参与局部交互。缓存左右区域向量 $r_j$，对查询 $q$ 计算 $u_j=q^Tr_j$、$w_j=\\mathrm{softmax}(u_j/.05)$，再汇总 $\\sum_j w_ju_j$。猫查询可提高猫区域权重，宇航员查询可提高宇航员区域权重；但这个聚合对区域顺序不敏感，交换左右后总分不变。多向量并不自动保存关系。","sha256":"6424f8550f5b5c6628f1965c5c1b4aafa939764425130e58f1a53d555fcca12a","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","title":"真实猫图结果：保留模型、照片与历史身份","number":3},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"真实猫图结果：保留模型、照片与历史身份","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"r_j","definitionContext":"真实猫图结果：保留模型、照片与历史身份"},{"latex":"q","definitionContext":"真实猫图结果：保留模型、照片与历史身份"},{"latex":"u_j=q^Tr_j","definitionContext":"真实猫图结果：保留模型、照片与历史身份"},{"latex":"w_j=\\mathrm{softmax}(u_j/.05)","definitionContext":"真实猫图结果：保留模型、照片与历史身份"},{"latex":"\\sum_j w_ju_j","definitionContext":"真实猫图结果：保留模型、照片与历史身份"}],"previousBlock":"b-a401b86ad1bcff4eb7980e3a","nextBlock":"b-ffe973d19777952ad08767e0","prerequisites":[{"id":"b-badbf96a9016a6dfa98db812","url":"#b-badbf96a9016a6dfa98db812","role":"section-title"},{"id":"b-5857325209e2365e5dae8218","url":"#b-5857325209e2365e5dae8218","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-ffe973d19777952ad08767e0","type":"p","scope":"block-b7ab2768ecf04c3aa9e1c6307a4de22b","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-ffe973d19777952ad08767e0","content":"逐对重复调用独立 CLIP 编码器仍是 $f(I)^Tg(T)$，不因此变成交叉编码。联合模型把图文在内部相互条件化，通常需要更多查询相关计算。可用双编码先召回，再少量重排；要判断收益，需要固定检索库、查询和评价协议。这些机制在现有 LAB 12 保留为选做，不在主讲另开一章。","sha256":"75e182f642c1113669edafb04efc844003f32942bf07649c400ab44b5772c86d","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","title":"真实猫图结果：保留模型、照片与历史身份","number":3},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"真实猫图结果：保留模型、照片与历史身份","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"f(I)^Tg(T)","definitionContext":"真实猫图结果：保留模型、照片与历史身份"}],"previousBlock":"b-2827633ff499e0da8d457338","nextBlock":"b-682a7ed23acdeb685cd33332","prerequisites":[{"id":"b-badbf96a9016a6dfa98db812","url":"#b-badbf96a9016a6dfa98db812","role":"section-title"},{"id":"b-5857325209e2365e5dae8218","url":"#b-5857325209e2365e5dae8218","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-682a7ed23acdeb685cd33332","type":"p","scope":"block-b7ab2768ecf04c3aa9e1c6307a4de22b","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-682a7ed23acdeb685cd33332","content":"[选读固定 URL](/course/blocks/retrieval-approximation-and-interaction.html)","sha256":"79b23bf9a1f450a215537903032b76943a29701344a0725834d43be47d02bdb8","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","title":"真实猫图结果：保留模型、照片与历史身份","number":3},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"真实猫图结果：保留模型、照片与历史身份","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-ffe973d19777952ad08767e0","nextBlock":"b-132c629cab172fe46126e461","prerequisites":[{"id":"b-badbf96a9016a6dfa98db812","url":"#b-badbf96a9016a6dfa98db812","role":"section-title"},{"id":"b-5857325209e2365e5dae8218","url":"#b-5857325209e2365e5dae8218","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-132c629cab172fe46126e461","type":"h2","scope":"block-b7ab2768ecf04c3aa9e1c6307a4de22b","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-132c629cab172fe46126e461","content":"## 真实 CLIP 关系反例与规则四格，应分开读","sha256":"7a070e4079949de63d3f40004e06bc2b44eabd82cfe803d7183c8a4e3aa60118","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","title":"真实猫图结果：保留模型、照片与历史身份","number":3},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"真实猫图结果：保留模型、照片与历史身份","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-682a7ed23acdeb685cd33332","nextBlock":"b-e539cf39a5808921df8c79e5","prerequisites":[{"id":"b-badbf96a9016a6dfa98db812","url":"#b-badbf96a9016a6dfa98db812","role":"section-title"},{"id":"b-5857325209e2365e5dae8218","url":"#b-5857325209e2365e5dae8218","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-e539cf39a5808921df8c79e5","type":"p","scope":"block-b7ab2768ecf04c3aa9e1c6307a4de22b","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-e539cf39a5808921df8c79e5","content":"持球四格首先是一份规则预测表；它与 LAB 04 另一组真实 CLIP 关系推理不是同一证据。后者使用猫和宇航员的固定拼贴，交换左右位置，比较“猫在左、宇航员在右”及其相反描述。","sha256":"e58f3bfde63e21eff0229440f4df21ebc4312e7d153980a079fc9492e8714ce5","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","title":"真实猫图结果：保留模型、照片与历史身份","number":3},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"真实猫图结果：保留模型、照片与历史身份","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-132c629cab172fe46126e461","nextBlock":"b-76aee3b094abdc7169424ba0","prerequisites":[{"id":"b-132c629cab172fe46126e461","url":"#b-132c629cab172fe46126e461","role":"section-title"},{"id":"b-5857325209e2365e5dae8218","url":"#b-5857325209e2365e5dae8218","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-76aee3b094abdc7169424ba0","type":"p","scope":"block-b7ab2768ecf04c3aa9e1c6307a4de22b","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-76aee3b094abdc7169424ba0","content":"2026-09-22 保存结果中，猫左图的两句余弦约为 0.37850、0.38184；交换后约为 0.35660、0.36301。两图都偏向“宇航员在左、猫在右”，所以第一张关系判断错误。这说明在该固定模型、拼贴与文字条件下，整体匹配没有正确区分这一次左右绑定。","sha256":"a03704d3c05d5ec2a67c2178b051ac4fdd8b69d54e6f999b8cffb44f26c2335f","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","title":"真实猫图结果：保留模型、照片与历史身份","number":3},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"真实猫图结果：保留模型、照片与历史身份","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-e539cf39a5808921df8c79e5","nextBlock":"b-b7d3739e5ef0b2b3a27df848","prerequisites":[{"id":"b-132c629cab172fe46126e461","url":"#b-132c629cab172fe46126e461","role":"section-title"},{"id":"b-5857325209e2365e5dae8218","url":"#b-5857325209e2365e5dae8218","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-b7d3739e5ef0b2b3a27df848","type":"p","scope":"block-b7ab2768ecf04c3aa9e1c6307a4de22b","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-b7d3739e5ef0b2b3a27df848","content":"但同一模型在猫图候选分类和小图库检索中有真实成功，不能用这个失败抹去对象匹配能力。更准确的研究问题是：哪些对象和语义可以由全局表示稳定匹配，哪些细粒度关系需要更合适的空间组织、交互或训练资料？","sha256":"5bc03f71d0e73034dde1ae62b299fce1769bd31220bdeb1b8cbc64c7b96cfa93","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","title":"真实猫图结果：保留模型、照片与历史身份","number":3},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"真实猫图结果：保留模型、照片与历史身份","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-76aee3b094abdc7169424ba0","nextBlock":"b-810f813400087ff1c2d70a37","prerequisites":[{"id":"b-132c629cab172fe46126e461","url":"#b-132c629cab172fe46126e461","role":"section-title"},{"id":"b-5857325209e2365e5dae8218","url":"#b-5857325209e2365e5dae8218","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-810f813400087ff1c2d70a37","type":"p","scope":"block-b7ab2768ecf04c3aa9e1c6307a4de22b","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-810f813400087ff1c2d70a37","content":"LAB 12 的局部加权规则提供另一种解释路径：查询改变了猫/宇航员区域权重，却因汇总公式忽略坐标而对交换左右严格不敏感。这个失败来自可指出的结构对称性。它帮助区分“区域更多”与“关系已经被表达”，再自然接入 P63 的定位任务。以上均为已有历史记录，本轮没有重跑或扩大结论。","sha256":"8017cdceef3018595cd3113e4fc86bca4d857d85b80fd4e531a4badfb1d7c20b","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","title":"真实猫图结果：保留模型、照片与历史身份","number":3},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"真实猫图结果：保留模型、照片与历史身份","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-b7d3739e5ef0b2b3a27df848","nextBlock":"b-b4a8aeee378e54afae9f4905","prerequisites":[{"id":"b-132c629cab172fe46126e461","url":"#b-132c629cab172fe46126e461","role":"section-title"},{"id":"b-5857325209e2365e5dae8218","url":"#b-5857325209e2365e5dae8218","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-b4a8aeee378e54afae9f4905","type":"p","scope":"block-b7ab2768ecf04c3aa9e1c6307a4de22b","url":"https://codingai-lec03.pages.dev/course/notebooks/d890081ddce04dac9aec8238b2809250.html#b-b4a8aeee378e54afae9f4905","content":"[本节固定网页 URL](/course/blocks/clip-spatial-counterexample.html)","sha256":"950dd0014bf5ab61cb0bc697072261d2c0637d1caec586a17dcaee2049d756ca","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"reading-block-b7ab2768ecf04c3aa9e1c6307a4de22b","title":"真实猫图结果：保留模型、照片与历史身份","number":3},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"真实猫图结果：保留模型、照片与历史身份","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-810f813400087ff1c2d70a37","nextBlock":null,"prerequisites":[{"id":"b-132c629cab172fe46126e461","url":"#b-132c629cab172fe46126e461","role":"section-title"},{"id":"b-5857325209e2365e5dae8218","url":"#b-5857325209e2365e5dae8218","role":"problem-context"}],"relatedBlocks":[]}]}