{"version":2,"relationshipURLBase":"https://codingai-lec03.pages.dev/course/blocks/shared-convolution-reuse.html","relationshipResolution":"Resolve relative relationship URLs against relationshipURLBase; complete records and absolute URLs are in blocks.","title":"参数复用与一次前向结果的复用","url":"https://codingai-lec03.pages.dev/course/blocks/shared-convolution-reuse.html","markdown":"https://codingai-lec03.pages.dev/course/blocks/shared-convolution-reuse.md","pages":[{"id":"reading-opening","title":"参数复用与一次前向结果的复用","unit":"多个区域怎样复用同一份视觉计算？","unitId":"region-computation-reuse","chapter":"II","number":1}],"aliases":[{"oldId":"b-6382e5fec8fab2bbdc924988","targetId":"b-6382e5fec8fab2bbdc924988","page":"reading-opening","operation":"move","beforeSHA256":"8d34cacadac664ebb6b3dd4bf9529a5555166859ce8d734ac21d047d8071efa4","afterSHA256":"8d34cacadac664ebb6b3dd4bf9529a5555166859ce8d734ac21d047d8071efa4","reason":"Same source body in authored semantic page"},{"oldId":"b-ce707004250a2139236e0d19","targetId":"b-ce707004250a2139236e0d19","page":"reading-opening","operation":"move","beforeSHA256":"e000fe6edc6fe136c1722ebb0e45cf3feddac5bd500016f5f219b575e35fee52","afterSHA256":"e000fe6edc6fe136c1722ebb0e45cf3feddac5bd500016f5f219b575e35fee52","reason":"Same source body in authored semantic page"},{"oldId":"b-2794b4d70e97cda1db4a00c3","targetId":"b-2794b4d70e97cda1db4a00c3","page":"reading-opening","operation":"move","beforeSHA256":"932bf95223423530aa77b5a75c4d8e358fb6dd24d080002540e34220fbb056c7","afterSHA256":"932bf95223423530aa77b5a75c4d8e358fb6dd24d080002540e34220fbb056c7","reason":"Same source body in authored semantic page"},{"oldId":"b-e806d73ca6775a33c8bdd123","targetId":"b-e806d73ca6775a33c8bdd123","page":"reading-opening","operation":"move","beforeSHA256":"05cefa158150f4696144cbf571a843576e81595b5e71f4d8a2722ff8f50e07bc","afterSHA256":"05cefa158150f4696144cbf571a843576e81595b5e71f4d8a2722ff8f50e07bc","reason":"Same source body in authored semantic page"},{"oldId":"b-6f1e53fc5fd6de1fd101a76b","targetId":"b-6f1e53fc5fd6de1fd101a76b","page":"reading-opening","operation":"move","beforeSHA256":"452a20668dfbbfbbee14fff435225f950ce5eafad56605cba10fbe3678b86520","afterSHA256":"452a20668dfbbfbbee14fff435225f950ce5eafad56605cba10fbe3678b86520","reason":"Same source body in authored semantic page"},{"oldId":"b-b15bacb00fcceeee111703c0","targetId":"b-b15bacb00fcceeee111703c0","page":"reading-opening","operation":"move","beforeSHA256":"2ab749850314a59e27a184813b97c4580dc551dc16a85da42ec8bf2aab26390a","afterSHA256":"2ab749850314a59e27a184813b97c4580dc551dc16a85da42ec8bf2aab26390a","reason":"Same source body in authored semantic page"},{"oldId":"b-cffed4082ebd69ef8a243ef2","targetId":"b-cffed4082ebd69ef8a243ef2","page":"reading-opening","operation":"rewrite","beforeSHA256":"dbb6916e6d2fedb692d6eb333d65a6b097d0a2a9915523c3e0fcf52f170a9692","afterSHA256":"6dbc372607f88adc2634d9914bb937e8dbda92d3e5d12fb5defcc932f42f4953","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-2e588de74846ac6105ef858f","targetId":"b-2e588de74846ac6105ef858f","page":"reading-opening","operation":"move","beforeSHA256":"2b9ccef7e1220e2250cceff1abd8bdd736b42f8da5a223f26b029f44b962e807","afterSHA256":"2b9ccef7e1220e2250cceff1abd8bdd736b42f8da5a223f26b029f44b962e807","reason":"Same source body in authored semantic page"}],"blocks":[{"id":"b-6382e5fec8fab2bbdc924988","type":"h1","scope":"shared-convolution-reuse","url":"https://codingai-lec03.pages.dev/course/blocks/shared-convolution-reuse.html#b-6382e5fec8fab2bbdc924988","content":"# 参数复用与一次前向结果的复用","sha256":"8d34cacadac664ebb6b3dd4bf9529a5555166859ce8d734ac21d047d8071efa4","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"region-computation-reuse","title":"多个区域怎样复用同一份视觉计算？"},"page":{"id":"reading-opening","title":"参数复用与一次前向结果的复用","number":1},"sectionTitle":"参数复用与一次前向结果的复用","currentQuestion":"参数复用与一次前向结果的复用","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-ce707004250a2139236e0d19","prerequisites":[{"id":"b-ce707004250a2139236e0d19","url":"#b-ce707004250a2139236e0d19","role":"problem-context"}],"relatedBlocks":[{"id":"b-cffed4082ebd69ef8a243ef2","url":"#b-cffed4082ebd69ef8a243ef2","relation":"figure-and-explanation"}]},{"id":"b-ce707004250a2139236e0d19","type":"p","scope":"shared-convolution-reuse","url":"https://codingai-lec03.pages.dev/course/blocks/shared-convolution-reuse.html#b-ce707004250a2139236e0d19","content":"同一个 CNN 权重处理多个裁剪，仍可能重复计算共同像素。整图共享改变的是一次推理的计算组织，与预训练参数迁移分开。共享之后每个区域仍需要自己的坐标和读出；计算重排并不保证与原来的裁剪路线数学等价。","sha256":"e000fe6edc6fe136c1722ebb0e45cf3feddac5bd500016f5f219b575e35fee52","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"region-computation-reuse","title":"多个区域怎样复用同一份视觉计算？"},"page":{"id":"reading-opening","title":"参数复用与一次前向结果的复用","number":1},"sectionTitle":"参数复用与一次前向结果的复用","currentQuestion":"参数复用与一次前向结果的复用","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-6382e5fec8fab2bbdc924988","nextBlock":"b-2794b4d70e97cda1db4a00c3","prerequisites":[{"id":"b-6382e5fec8fab2bbdc924988","url":"#b-6382e5fec8fab2bbdc924988","role":"section-title"}],"relatedBlocks":[{"id":"b-cffed4082ebd69ef8a243ef2","url":"#b-cffed4082ebd69ef8a243ef2","relation":"figure-and-explanation"}]},{"id":"b-2794b4d70e97cda1db4a00c3","type":"p","scope":"shared-convolution-reuse","url":"https://codingai-lec03.pages.dev/course/blocks/shared-convolution-reuse.html#b-2794b4d70e97cda1db4a00c3","content":"两张重叠裁剪都使用同一个 CNN，已经共享了参数，但仍分别执行卷积。假如它们都包含杯沿和桌面，这些像素会参与两次运算。SPP-net 和 Fast R-CNN 的路线把较重的卷积移到整图：先计算 $F=f_\\theta(I)$，每个候选再用自己的框 $b_i$ 读取 $r_i=R(F,b_i)$。","sha256":"932bf95223423530aa77b5a75c4d8e358fb6dd24d080002540e34220fbb056c7","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"region-computation-reuse","title":"多个区域怎样复用同一份视觉计算？"},"page":{"id":"reading-opening","title":"参数复用与一次前向结果的复用","number":1},"sectionTitle":"参数复用与一次前向结果的复用","currentQuestion":"参数复用与一次前向结果的复用","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"F=f_\\theta(I)","definitionContext":"参数复用与一次前向结果的复用"},{"latex":"b_i","definitionContext":"参数复用与一次前向结果的复用"},{"latex":"r_i=R(F,b_i)","definitionContext":"参数复用与一次前向结果的复用"}],"previousBlock":"b-ce707004250a2139236e0d19","nextBlock":"b-e806d73ca6775a33c8bdd123","prerequisites":[{"id":"b-6382e5fec8fab2bbdc924988","url":"#b-6382e5fec8fab2bbdc924988","role":"section-title"},{"id":"b-ce707004250a2139236e0d19","url":"#b-ce707004250a2139236e0d19","role":"problem-context"}],"relatedBlocks":[{"id":"b-cffed4082ebd69ef8a243ef2","url":"#b-cffed4082ebd69ef8a243ef2","relation":"figure-and-explanation"}]},{"id":"b-e806d73ca6775a33c8bdd123","type":"p","scope":"shared-convolution-reuse","url":"https://codingai-lec03.pages.dev/course/blocks/shared-convolution-reuse.html#b-e806d73ca6775a33c8bdd123","content":"这里有两种复用。预训练让新的任务借用过去学到的参数；整图共享则让同一次推理中的多个区域借用已算出的中间结果。前者减少从头学习的需要，后者减少重复工作，不能用同一个“迁移”概念含糊带过。","sha256":"05cefa158150f4696144cbf571a843576e81595b5e71f4d8a2722ff8f50e07bc","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"region-computation-reuse","title":"多个区域怎样复用同一份视觉计算？"},"page":{"id":"reading-opening","title":"参数复用与一次前向结果的复用","number":1},"sectionTitle":"参数复用与一次前向结果的复用","currentQuestion":"参数复用与一次前向结果的复用","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-2794b4d70e97cda1db4a00c3","nextBlock":"b-6f1e53fc5fd6de1fd101a76b","prerequisites":[{"id":"b-6382e5fec8fab2bbdc924988","url":"#b-6382e5fec8fab2bbdc924988","role":"section-title"},{"id":"b-ce707004250a2139236e0d19","url":"#b-ce707004250a2139236e0d19","role":"problem-context"}],"relatedBlocks":[{"id":"b-cffed4082ebd69ef8a243ef2","url":"#b-cffed4082ebd69ef8a243ef2","relation":"figure-and-explanation"}]},{"id":"b-6f1e53fc5fd6de1fd101a76b","type":"p","scope":"shared-convolution-reuse","url":"https://codingai-lec03.pages.dev/course/blocks/shared-convolution-reuse.html#b-6f1e53fc5fd6de1fd101a76b","content":"若有 N 个候选，逐区域成本近似为 N 次卷积加区域预测；共享方式近似为一次整图卷积加 N 次轻量读取与预测。这是计算组织的估算，不是实测加速倍数，因为整图与裁剪的分辨率不同，内存、候选产生和预测头也占时间。","sha256":"452a20668dfbbfbbee14fff435225f950ce5eafad56605cba10fbe3678b86520","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"region-computation-reuse","title":"多个区域怎样复用同一份视觉计算？"},"page":{"id":"reading-opening","title":"参数复用与一次前向结果的复用","number":1},"sectionTitle":"参数复用与一次前向结果的复用","currentQuestion":"参数复用与一次前向结果的复用","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-e806d73ca6775a33c8bdd123","nextBlock":"b-b15bacb00fcceeee111703c0","prerequisites":[{"id":"b-6382e5fec8fab2bbdc924988","url":"#b-6382e5fec8fab2bbdc924988","role":"section-title"},{"id":"b-ce707004250a2139236e0d19","url":"#b-ce707004250a2139236e0d19","role":"problem-context"}],"relatedBlocks":[{"id":"b-cffed4082ebd69ef8a243ef2","url":"#b-cffed4082ebd69ef8a243ef2","relation":"figure-and-explanation"}]},{"id":"b-b15bacb00fcceeee111703c0","type":"p","scope":"shared-convolution-reuse","url":"https://codingai-lec03.pages.dev/course/blocks/shared-convolution-reuse.html#b-b15bacb00fcceeee111703c0","content":"先裁再编码与先编码再读取不严格相等：前者改变对象尺度并裁去上下文；后者的感受野可能跨出 RoI。方法的贡献是在这样的取舍下找到能训练、有效率的共享接口。","sha256":"2ab749850314a59e27a184813b97c4580dc551dc16a85da42ec8bf2aab26390a","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"region-computation-reuse","title":"多个区域怎样复用同一份视觉计算？"},"page":{"id":"reading-opening","title":"参数复用与一次前向结果的复用","number":1},"sectionTitle":"参数复用与一次前向结果的复用","currentQuestion":"参数复用与一次前向结果的复用","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-6f1e53fc5fd6de1fd101a76b","nextBlock":"b-cffed4082ebd69ef8a243ef2","prerequisites":[{"id":"b-6382e5fec8fab2bbdc924988","url":"#b-6382e5fec8fab2bbdc924988","role":"section-title"},{"id":"b-ce707004250a2139236e0d19","url":"#b-ce707004250a2139236e0d19","role":"problem-context"}],"relatedBlocks":[{"id":"b-cffed4082ebd69ef8a243ef2","url":"#b-cffed4082ebd69ef8a243ef2","relation":"figure-and-explanation"}]},{"id":"b-cffed4082ebd69ef8a243ef2","type":"figure","scope":"shared-convolution-reuse","url":"https://codingai-lec03.pages.dev/course/blocks/shared-convolution-reuse.html#b-cffed4082ebd69ef8a243ef2","content":"图内标签与数值：Lecture 03 课堂图解 P29；先裁区域：重叠内容反复编码；CNN；CNN；先算全图：再按区域读取；预训练复用的是过去学到的参数；这里复用的是当前图片已经算出的特征。\n\n共享的是整张图的 F；两个候选仍有各自的读出。 查看对应图解状态\n\n- 先裁区域：重叠内容反复编码\n- CNN\n- CNN\n- 先算全图：再按区域读取\n- 预训练复用的是过去学到的参数；这里复用的是当前图片已经算出的特征。共享的是整张图的 F；两个候选仍有各自的读出。 [查看对应图解状态](/#/l3-r2-p29/1)[查看完整图解](/#/l3-r2-p29/1)","sha256":"6dbc372607f88adc2634d9914bb937e8dbda92d3e5d12fb5defcc932f42f4953","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"region-computation-reuse","title":"多个区域怎样复用同一份视觉计算？"},"page":{"id":"reading-opening","title":"参数复用与一次前向结果的复用","number":1},"sectionTitle":"参数复用与一次前向结果的复用","currentQuestion":"参数复用与一次前向结果的复用","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"figure":{"caption":"共享的是整张图的 F；两个候选仍有各自的读出。 查看对应图解状态","labels":["先裁区域：重叠内容反复编码","CNN","CNN","先算全图：再按区域读取","预训练复用的是过去学到的参数；这里复用的是当前图片已经算出的特征。"],"description":"共享的是整张图的 F；两个候选仍有各自的读出。 查看对应图解状态先裁区域：重叠内容反复编码CNNCNN先算全图：再按区域读取预训练复用的是过去学到的参数；这里复用的是当前图片已经算出的特征。","evidenceIdentity":"共享的是整张图的 F；两个候选仍有各自的读出。 查看对应图解状态","images":[{"src":"/assets/display/8ced939a1fd55047fbf73ea5.webp","alt":""},{"src":"/assets/display/8ced939a1fd55047fbf73ea5.webp","alt":""},{"src":"/assets/display/8ced939a1fd55047fbf73ea5.webp","alt":""}]},"previousBlock":"b-b15bacb00fcceeee111703c0","nextBlock":"b-2e588de74846ac6105ef858f","prerequisites":[{"id":"b-6382e5fec8fab2bbdc924988","url":"#b-6382e5fec8fab2bbdc924988","role":"section-title"},{"id":"b-ce707004250a2139236e0d19","url":"#b-ce707004250a2139236e0d19","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-2e588de74846ac6105ef858f","type":"h2","scope":"shared-convolution-reuse","url":"https://codingai-lec03.pages.dev/course/blocks/shared-convolution-reuse.html#b-2e588de74846ac6105ef858f","content":"## 继续阅读与实验","sha256":"2b9ccef7e1220e2250cceff1abd8bdd736b42f8da5a223f26b029f44b962e807","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"region-computation-reuse","title":"多个区域怎样复用同一份视觉计算？"},"page":{"id":"reading-opening","title":"参数复用与一次前向结果的复用","number":1},"sectionTitle":"参数复用与一次前向结果的复用","currentQuestion":"参数复用与一次前向结果的复用","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-cffed4082ebd69ef8a243ef2","nextBlock":null,"prerequisites":[{"id":"b-6382e5fec8fab2bbdc924988","url":"#b-6382e5fec8fab2bbdc924988","role":"section-title"},{"id":"b-ce707004250a2139236e0d19","url":"#b-ce707004250a2139236e0d19","role":"problem-context"}],"relatedBlocks":[{"id":"b-cffed4082ebd69ef8a243ef2","url":"#b-cffed4082ebd69ef8a243ef2","relation":"figure-and-explanation"}]}]}