{"version":2,"relationshipURLBase":"https://codingai-lec03.pages.dev/course/notebook-blocks/8cfb433dc5cc4c4d9b8052e80463dfb5.html","relationshipResolution":"Resolve relative relationship URLs against relationshipURLBase; complete records and absolute URLs are in blocks.","title":"5 · 查询依赖的交互：一个可运行的区域加权对照","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/8cfb433dc5cc4c4d9b8052e80463dfb5.html","markdown":"https://codingai-lec03.pages.dev/course/notebook-blocks/8cfb433dc5cc4c4d9b8052e80463dfb5.md","pages":[{"id":"reading-opening","title":"5 · 查询依赖的交互：一个可运行的区域加权对照","unit":"5 · 查询依赖的交互：一个可运行的区域加权对照","unitId":"notebook-blocks/8cfb433dc5cc4c4d9b8052e80463dfb5","chapter":"course","number":1}],"aliases":[{"oldId":"b-a63369e3222a54fe9c2432e7","targetId":"b-a63369e3222a54fe9c2432e7","page":"reading-opening","operation":"move","beforeSHA256":"1a4de4df51c1a8787dc96ec401888dfe89d250842519319b9af7a931ca765eee","afterSHA256":"1a4de4df51c1a8787dc96ec401888dfe89d250842519319b9af7a931ca765eee","reason":"Same source body in authored semantic page"},{"oldId":"b-4045e5e04384834221f0eb74","targetId":"b-4045e5e04384834221f0eb74","page":"reading-opening","operation":"move","beforeSHA256":"99db53bf710466e865719979f388bdc986bb82eee0634ae11172cd3f282b4a96","afterSHA256":"99db53bf710466e865719979f388bdc986bb82eee0634ae11172cd3f282b4a96","reason":"Same source body in authored semantic page"},{"oldId":"b-6d5a0205456f80c4c6fba2e2","targetId":"b-6d5a0205456f80c4c6fba2e2","page":"reading-opening","operation":"move","beforeSHA256":"dc06c1e728a4b5740ba634b145bc9a1da6323031626734cb4555db5a744df806","afterSHA256":"dc06c1e728a4b5740ba634b145bc9a1da6323031626734cb4555db5a744df806","reason":"Same source body in authored semantic page"},{"oldId":"b-74c210e69fbb607ff093d3ff","targetId":"b-74c210e69fbb607ff093d3ff","page":"reading-opening","operation":"move","beforeSHA256":"8021c18afc52ba7b6612f43ecfb0fdd4789efe91cc9cd656984357d81a8514de","afterSHA256":"8021c18afc52ba7b6612f43ecfb0fdd4789efe91cc9cd656984357d81a8514de","reason":"Same source body in authored semantic page"},{"oldId":"b-576767ade454e02c9f6fa247","targetId":"b-576767ade454e02c9f6fa247","page":"reading-opening","operation":"move","beforeSHA256":"2fd9111f853e5d93f7d59d72c39d6652bb45bfaa5b4f36353c00e0eca228db69","afterSHA256":"2fd9111f853e5d93f7d59d72c39d6652bb45bfaa5b4f36353c00e0eca228db69","reason":"Same source body in authored semantic page"},{"oldId":"b-f9eddd499b35eb418720faa9","targetId":"b-f9eddd499b35eb418720faa9","page":"reading-opening","operation":"move","beforeSHA256":"1c0a54709b04ee6babf39f66eee1768f60b0f62b5c9cd89bb4c23667b8ade7c6","afterSHA256":"1c0a54709b04ee6babf39f66eee1768f60b0f62b5c9cd89bb4c23667b8ade7c6","reason":"Same source body in authored semantic page"},{"oldId":"b-3ad28c5197476ed2458a019d","targetId":"b-3ad28c5197476ed2458a019d","page":"reading-opening","operation":"move","beforeSHA256":"9131075fd10fc81a22296a0bfdfa26a9b81a5fc2506249842fe1f424e20c20b0","afterSHA256":"9131075fd10fc81a22296a0bfdfa26a9b81a5fc2506249842fe1f424e20c20b0","reason":"Same source body in authored semantic page"},{"oldId":"b-103bbac2ffe097d9cfebd232","targetId":"b-103bbac2ffe097d9cfebd232","page":"reading-opening","operation":"move","beforeSHA256":"7319327aded1ab0df3bfecb88955ac7ec5e20b4a260492e8c92a288c7f8f2c0e","afterSHA256":"7319327aded1ab0df3bfecb88955ac7ec5e20b4a260492e8c92a288c7f8f2c0e","reason":"Same source body in authored semantic page"},{"oldId":"block-8cfb433dc5cc4c4d9b8052e80463dfb5--5-查询依赖的交互一个可运行的区域加权对照","targetId":"b-576767ade454e02c9f6fa247","operation":"retain-source-identity"}],"blocks":[{"id":"b-a63369e3222a54fe9c2432e7","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/8cfb433dc5cc4c4d9b8052e80463dfb5.html#b-a63369e3222a54fe9c2432e7","content":"[LAB 12 · CLIP延伸：缓存与检索（近似索引选读）](/course/notebooks/921eeaf784284ce19568c6ca7619b55a.html)","sha256":"1a4de4df51c1a8787dc96ec401888dfe89d250842519319b9af7a931ca765eee","chapter":{"id":"course","title":"5 · 查询依赖的交互：一个可运行的区域加权对照"},"unit":{"id":"notebook-blocks/8cfb433dc5cc4c4d9b8052e80463dfb5","title":"5 · 查询依赖的交互：一个可运行的区域加权对照"},"page":{"id":"reading-opening","title":"5 · 查询依赖的交互：一个可运行的区域加权对照","number":1},"sectionTitle":"5 · 查询依赖的交互：一个可运行的区域加权对照","currentQuestion":"5 · 查询依赖的交互：一个可运行的区域加权对照","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-4045e5e04384834221f0eb74","prerequisites":[],"relatedBlocks":[]},{"id":"b-4045e5e04384834221f0eb74","type":"h1","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/8cfb433dc5cc4c4d9b8052e80463dfb5.html#b-4045e5e04384834221f0eb74","content":"# 5 · 查询依赖的交互：一个可运行的区域加权对照","sha256":"99db53bf710466e865719979f388bdc986bb82eee0634ae11172cd3f282b4a96","chapter":{"id":"course","title":"5 · 查询依赖的交互：一个可运行的区域加权对照"},"unit":{"id":"notebook-blocks/8cfb433dc5cc4c4d9b8052e80463dfb5","title":"5 · 查询依赖的交互：一个可运行的区域加权对照"},"page":{"id":"reading-opening","title":"5 · 查询依赖的交互：一个可运行的区域加权对照","number":1},"sectionTitle":"5 · 查询依赖的交互：一个可运行的区域加权对照","currentQuestion":"5 · 查询依赖的交互：一个可运行的区域加权对照","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-a63369e3222a54fe9c2432e7","nextBlock":"b-6d5a0205456f80c4c6fba2e2","prerequisites":[{"id":"b-a63369e3222a54fe9c2432e7","url":"#b-a63369e3222a54fe9c2432e7","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-6d5a0205456f80c4c6fba2e2","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/8cfb433dc5cc4c4d9b8052e80463dfb5.html#b-6d5a0205456f80c4c6fba2e2","content":"并入相关学习路径","sha256":"dc06c1e728a4b5740ba634b145bc9a1da6323031626734cb4555db5a744df806","chapter":{"id":"course","title":"5 · 查询依赖的交互：一个可运行的区域加权对照"},"unit":{"id":"notebook-blocks/8cfb433dc5cc4c4d9b8052e80463dfb5","title":"5 · 查询依赖的交互：一个可运行的区域加权对照"},"page":{"id":"reading-opening","title":"5 · 查询依赖的交互：一个可运行的区域加权对照","number":1},"sectionTitle":"5 · 查询依赖的交互：一个可运行的区域加权对照","currentQuestion":"5 · 查询依赖的交互：一个可运行的区域加权对照","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-4045e5e04384834221f0eb74","nextBlock":"b-74c210e69fbb607ff093d3ff","prerequisites":[{"id":"b-4045e5e04384834221f0eb74","url":"#b-4045e5e04384834221f0eb74","role":"section-title"},{"id":"b-a63369e3222a54fe9c2432e7","url":"#b-a63369e3222a54fe9c2432e7","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-74c210e69fbb607ff093d3ff","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/8cfb433dc5cc4c4d9b8052e80463dfb5.html#b-74c210e69fbb607ff093d3ff","content":"同一CLIP缓存、exact/IVF搜索、延迟/扫描量、区域交互基线","sha256":"8021c18afc52ba7b6612f43ecfb0fdd4789efe91cc9cd656984357d81a8514de","chapter":{"id":"course","title":"5 · 查询依赖的交互：一个可运行的区域加权对照"},"unit":{"id":"notebook-blocks/8cfb433dc5cc4c4d9b8052e80463dfb5","title":"5 · 查询依赖的交互：一个可运行的区域加权对照"},"page":{"id":"reading-opening","title":"5 · 查询依赖的交互：一个可运行的区域加权对照","number":1},"sectionTitle":"5 · 查询依赖的交互：一个可运行的区域加权对照","currentQuestion":"5 · 查询依赖的交互：一个可运行的区域加权对照","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-6d5a0205456f80c4c6fba2e2","nextBlock":"b-576767ade454e02c9f6fa247","prerequisites":[{"id":"b-4045e5e04384834221f0eb74","url":"#b-4045e5e04384834221f0eb74","role":"section-title"},{"id":"b-a63369e3222a54fe9c2432e7","url":"#b-a63369e3222a54fe9c2432e7","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-576767ade454e02c9f6fa247","type":"h2","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/8cfb433dc5cc4c4d9b8052e80463dfb5.html#b-576767ade454e02c9f6fa247","content":"## 5 · 查询依赖的交互：一个可运行的区域加权对照","sha256":"2fd9111f853e5d93f7d59d72c39d6652bb45bfaa5b4f36353c00e0eca228db69","chapter":{"id":"course","title":"5 · 查询依赖的交互：一个可运行的区域加权对照"},"unit":{"id":"notebook-blocks/8cfb433dc5cc4c4d9b8052e80463dfb5","title":"5 · 查询依赖的交互：一个可运行的区域加权对照"},"page":{"id":"reading-opening","title":"5 · 查询依赖的交互：一个可运行的区域加权对照","number":1},"sectionTitle":"5 · 查询依赖的交互：一个可运行的区域加权对照","currentQuestion":"5 · 查询依赖的交互：一个可运行的区域加权对照","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-74c210e69fbb607ff093d3ff","nextBlock":"b-f9eddd499b35eb418720faa9","prerequisites":[{"id":"b-4045e5e04384834221f0eb74","url":"#b-4045e5e04384834221f0eb74","role":"section-title"},{"id":"b-a63369e3222a54fe9c2432e7","url":"#b-a63369e3222a54fe9c2432e7","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-f9eddd499b35eb418720faa9","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/8cfb433dc5cc4c4d9b8052e80463dfb5.html#b-f9eddd499b35eb418720faa9","content":"“每对都调用模型”不等于“有图文交互”。我们再定义一个小而完整的 query-dependent 后期交互：先缓存左右两个照片区域的 CLIP 向量 $r_1,r_2$；对每一个查询 $q$，计算 $u_j=q^\\top r_j$，然后用 $w_j=\\mathrm{softmax}(u_j/0.05)$ 加权，得到分数 $s=\\sum_jw_ju_j$。查询不同，区域权重和汇总表示不同，所以它确实在比较时使用了图和文的共同信息；但它是本册定义的加权规则，没有训练新的联合编码器，也不是 CLIP 官方的全图匹配分数。","sha256":"1c0a54709b04ee6babf39f66eee1768f60b0f62b5c9cd89bb4c23667b8ade7c6","chapter":{"id":"course","title":"5 · 查询依赖的交互：一个可运行的区域加权对照"},"unit":{"id":"notebook-blocks/8cfb433dc5cc4c4d9b8052e80463dfb5","title":"5 · 查询依赖的交互：一个可运行的区域加权对照"},"page":{"id":"reading-opening","title":"5 · 查询依赖的交互：一个可运行的区域加权对照","number":1},"sectionTitle":"5 · 查询依赖的交互：一个可运行的区域加权对照","currentQuestion":"5 · 查询依赖的交互：一个可运行的区域加权对照","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"r_1,r_2","definitionContext":"5 · 查询依赖的交互：一个可运行的区域加权对照"},{"latex":"q","definitionContext":"5 · 查询依赖的交互：一个可运行的区域加权对照"},{"latex":"u_j=q^\\top r_j","definitionContext":"5 · 查询依赖的交互：一个可运行的区域加权对照"},{"latex":"w_j=\\mathrm{softmax}(u_j/0.05)","definitionContext":"5 · 查询依赖的交互：一个可运行的区域加权对照"},{"latex":"s=\\sum_jw_ju_j","definitionContext":"5 · 查询依赖的交互：一个可运行的区域加权对照"}],"previousBlock":"b-576767ade454e02c9f6fa247","nextBlock":"b-3ad28c5197476ed2458a019d","prerequisites":[{"id":"b-576767ade454e02c9f6fa247","url":"#b-576767ade454e02c9f6fa247","role":"section-title"},{"id":"b-a63369e3222a54fe9c2432e7","url":"#b-a63369e3222a54fe9c2432e7","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-3ad28c5197476ed2458a019d","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/8cfb433dc5cc4c4d9b8052e80463dfb5.html#b-3ad28c5197476ed2458a019d","content":"在“猫左、宇航员右”拼贴上，cat 查询给左猫区94.15%的权重；astronaut 查询给右宇航员区80.08%。这是一项可观察的查询依赖变化。它的代价是每对查询与图要计算多个区域点积，缓存也从一向量变成区域集合；由于区域特征独立生成，它仍可部分预计算。","sha256":"9131075fd10fc81a22296a0bfdfa26a9b81a5fc2506249842fe1f424e20c20b0","chapter":{"id":"course","title":"5 · 查询依赖的交互：一个可运行的区域加权对照"},"unit":{"id":"notebook-blocks/8cfb433dc5cc4c4d9b8052e80463dfb5","title":"5 · 查询依赖的交互：一个可运行的区域加权对照"},"page":{"id":"reading-opening","title":"5 · 查询依赖的交互：一个可运行的区域加权对照","number":1},"sectionTitle":"5 · 查询依赖的交互：一个可运行的区域加权对照","currentQuestion":"5 · 查询依赖的交互：一个可运行的区域加权对照","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-f9eddd499b35eb418720faa9","nextBlock":"b-103bbac2ffe097d9cfebd232","prerequisites":[{"id":"b-576767ade454e02c9f6fa247","url":"#b-576767ade454e02c9f6fa247","role":"section-title"},{"id":"b-a63369e3222a54fe9c2432e7","url":"#b-a63369e3222a54fe9c2432e7","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-103bbac2ffe097d9cfebd232","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/8cfb433dc5cc4c4d9b8052e80463dfb5.html#b-103bbac2ffe097d9cfebd232","content":"接着使用两条“谁在左/右”的关系句。交换区域顺序后，下述加权分数**完全相同**，因为这个公式对区域排列不敏感；它没有使用坐标。这不是代码 bug，而是设计本身丢失了顺序。如果需要真正的区域关系，需要显式位置、关系模块或经训练的联合模型。一个一般的 cross-encoder 会在内部让图文 token 相互条件化，通常不能将整次交互压成一个与查询无关的图片向量；本册只给其结构和成本解释，没有假装运行或比较某个未加载的 cross-encoder。","sha256":"7319327aded1ab0df3bfecb88955ac7ec5e20b4a260492e8c92a288c7f8f2c0e","chapter":{"id":"course","title":"5 · 查询依赖的交互：一个可运行的区域加权对照"},"unit":{"id":"notebook-blocks/8cfb433dc5cc4c4d9b8052e80463dfb5","title":"5 · 查询依赖的交互：一个可运行的区域加权对照"},"page":{"id":"reading-opening","title":"5 · 查询依赖的交互：一个可运行的区域加权对照","number":1},"sectionTitle":"5 · 查询依赖的交互：一个可运行的区域加权对照","currentQuestion":"5 · 查询依赖的交互：一个可运行的区域加权对照","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-3ad28c5197476ed2458a019d","nextBlock":null,"prerequisites":[{"id":"b-576767ade454e02c9f6fa247","url":"#b-576767ade454e02c9f6fa247","role":"section-title"},{"id":"b-a63369e3222a54fe9c2432e7","url":"#b-a63369e3222a54fe9c2432e7","role":"problem-context"}],"relatedBlocks":[]}]}
