{"version":2,"relationshipURLBase":"https://codingai-lec03.pages.dev/course/blocks/sam-task-training.html","relationshipResolution":"Resolve relative relationship URLs against relationshipURLBase; complete records and absolute URLs are in blocks.","title":"MAE 提供起点，掩码任务训练学会怎样响应提示","url":"https://codingai-lec03.pages.dev/course/blocks/sam-task-training.html","markdown":"https://codingai-lec03.pages.dev/course/blocks/sam-task-training.md","pages":[{"id":"reading-opening","title":"MAE 提供起点，掩码任务训练学会怎样响应提示","unit":"提示怎样指定这一次要分的区域？","unitId":"prompted-spatial-output","chapter":"II","number":1}],"aliases":[{"oldId":"b-ab6fd575202cd6a355e14eb6","targetId":"b-ab6fd575202cd6a355e14eb6","page":"reading-opening","operation":"move","beforeSHA256":"be697915dbb20e206a160720448d056508cd6a5faab412d6b7737460f9dfccc7","afterSHA256":"be697915dbb20e206a160720448d056508cd6a5faab412d6b7737460f9dfccc7","reason":"Same source body in authored semantic page"},{"oldId":"b-512e94b0c5990fc5e166403e","targetId":"b-512e94b0c5990fc5e166403e","page":"reading-opening","operation":"move","beforeSHA256":"3f74312c58902b7fc5b6c50187eec80b4436bbe4c0d78080270d152c18752588","afterSHA256":"3f74312c58902b7fc5b6c50187eec80b4436bbe4c0d78080270d152c18752588","reason":"Same source body in authored semantic page"},{"oldId":"b-f7aac8f5ebed29807ea7c067","targetId":"b-f7aac8f5ebed29807ea7c067","page":"reading-opening","operation":"move","beforeSHA256":"d81be8895ad7921895d1f461f04b87ae6a1812cacfdfd8bae29397181d035b00","afterSHA256":"d81be8895ad7921895d1f461f04b87ae6a1812cacfdfd8bae29397181d035b00","reason":"Same source body in authored semantic page"},{"oldId":"b-0e83c4569394dab3c0552443","targetId":"b-0e83c4569394dab3c0552443","page":"reading-opening","operation":"move","beforeSHA256":"67b3d3c3aad52c0d80e2601a93b15e852b3304b341f8f30e0eaa71ce59012241","afterSHA256":"67b3d3c3aad52c0d80e2601a93b15e852b3304b341f8f30e0eaa71ce59012241","reason":"Same source body in authored semantic page"},{"oldId":"b-08040fbec867229eb8828cb1","targetId":"b-08040fbec867229eb8828cb1","page":"reading-opening","operation":"move","beforeSHA256":"288b13fb11a7b4bf50ae1b07b0d59fe594cb1b283e580e44cb03d1c60f4dcdf6","afterSHA256":"288b13fb11a7b4bf50ae1b07b0d59fe594cb1b283e580e44cb03d1c60f4dcdf6","reason":"Same source body in authored semantic page"},{"oldId":"b-a83a5afa0ed74ca552a19c22","targetId":"b-a83a5afa0ed74ca552a19c22","page":"reading-opening","operation":"move","beforeSHA256":"d09830d2ab723a8ef4a85da99e5bb96bc3f95fcd6e26eeaad7e1b354c766da11","afterSHA256":"d09830d2ab723a8ef4a85da99e5bb96bc3f95fcd6e26eeaad7e1b354c766da11","reason":"Same source body in authored semantic page"},{"oldId":"b-24e585dca580f34f722537c8","targetId":"b-24e585dca580f34f722537c8","page":"reading-opening","operation":"rewrite","beforeSHA256":"20690c55c04cc2b0fd8f8c838924f5be9800af0871ddda78fe78a32b62c1fc95","afterSHA256":"05e1a3d0840ee29fa44fc66e8c9542cc733145315abd9168ef5dae4f8aaa3052","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-48337a685d805af2845d80af","targetId":"b-48337a685d805af2845d80af","page":"reading-opening","operation":"move","beforeSHA256":"2b9ccef7e1220e2250cceff1abd8bdd736b42f8da5a223f26b029f44b962e807","afterSHA256":"2b9ccef7e1220e2250cceff1abd8bdd736b42f8da5a223f26b029f44b962e807","reason":"Same source body in authored semantic page"}],"blocks":[{"id":"b-ab6fd575202cd6a355e14eb6","type":"h1","scope":"sam-task-training","url":"https://codingai-lec03.pages.dev/course/blocks/sam-task-training.html#b-ab6fd575202cd6a355e14eb6","content":"# MAE 提供起点，掩码任务训练学会怎样响应提示","sha256":"be697915dbb20e206a160720448d056508cd6a5faab412d6b7737460f9dfccc7","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"prompted-spatial-output","title":"提示怎样指定这一次要分的区域？"},"page":{"id":"reading-opening","title":"MAE 提供起点，掩码任务训练学会怎样响应提示","number":1},"sectionTitle":"MAE 提供起点，掩码任务训练学会怎样响应提示","currentQuestion":"MAE 提供起点，掩码任务训练学会怎样响应提示","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-512e94b0c5990fc5e166403e","prerequisites":[{"id":"b-512e94b0c5990fc5e166403e","url":"#b-512e94b0c5990fc5e166403e","role":"problem-context"}],"relatedBlocks":[{"id":"b-24e585dca580f34f722537c8","url":"#b-24e585dca580f34f722537c8","relation":"figure-and-explanation"}]},{"id":"b-512e94b0c5990fc5e166403e","type":"p","scope":"sam-task-training","url":"https://codingai-lec03.pages.dev/course/blocks/sam-task-training.html#b-512e94b0c5990fc5e166403e","content":"把三个阶段分开：MAE 初始化视觉编码器，提示式 mask 训练建立任务能力，使用时缓存同图表示。具体训练目标和数据组织使提示能够起作用；这座桥把第一部分的表示来源与第二部分的空间接口连接起来。","sha256":"3f74312c58902b7fc5b6c50187eec80b4436bbe4c0d78080270d152c18752588","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"prompted-spatial-output","title":"提示怎样指定这一次要分的区域？"},"page":{"id":"reading-opening","title":"MAE 提供起点，掩码任务训练学会怎样响应提示","number":1},"sectionTitle":"MAE 提供起点，掩码任务训练学会怎样响应提示","currentQuestion":"MAE 提供起点，掩码任务训练学会怎样响应提示","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-ab6fd575202cd6a355e14eb6","nextBlock":"b-f7aac8f5ebed29807ea7c067","prerequisites":[{"id":"b-ab6fd575202cd6a355e14eb6","url":"#b-ab6fd575202cd6a355e14eb6","role":"section-title"}],"relatedBlocks":[{"id":"b-24e585dca580f34f722537c8","url":"#b-24e585dca580f34f722537c8","relation":"figure-and-explanation"}]},{"id":"b-f7aac8f5ebed29807ea7c067","type":"p","scope":"sam-task-training","url":"https://codingai-lec03.pages.dev/course/blocks/sam-task-training.html#b-f7aac8f5ebed29807ea7c067","content":"原始 SAM 的图像编码器使用 MAE 预训练初始化。这连接了第一部分的表示学习，但“遮挡重建”目标并没有直接提供“某个正点应对应哪个对象 mask”的全部监督。后续还要进行提示式分割训练。","sha256":"d81be8895ad7921895d1f461f04b87ae6a1812cacfdfd8bae29397181d035b00","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"prompted-spatial-output","title":"提示怎样指定这一次要分的区域？"},"page":{"id":"reading-opening","title":"MAE 提供起点，掩码任务训练学会怎样响应提示","number":1},"sectionTitle":"MAE 提供起点，掩码任务训练学会怎样响应提示","currentQuestion":"MAE 提供起点，掩码任务训练学会怎样响应提示","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-512e94b0c5990fc5e166403e","nextBlock":"b-0e83c4569394dab3c0552443","prerequisites":[{"id":"b-ab6fd575202cd6a355e14eb6","url":"#b-ab6fd575202cd6a355e14eb6","role":"section-title"},{"id":"b-512e94b0c5990fc5e166403e","url":"#b-512e94b0c5990fc5e166403e","role":"problem-context"}],"relatedBlocks":[{"id":"b-24e585dca580f34f722537c8","url":"#b-24e585dca580f34f722537c8","relation":"figure-and-explanation"}]},{"id":"b-0e83c4569394dab3c0552443","type":"p","scope":"sam-task-training","url":"https://codingai-lec03.pages.dev/course/blocks/sam-task-training.html#b-0e83c4569394dab3c0552443","content":"训练样本包含图像、目标 mask 与模拟提示。模型预测后，可以从错误区域追加正点或负点，再要求它修正；原论文训练模拟每个 mask 的多轮交互，掩码损失组合 focal 与 dice。这样，推理时的“改点、改框”对应训练中学过的任务，而不是凭空获得一个通用按钮。","sha256":"67b3d3c3aad52c0d80e2601a93b15e852b3304b341f8f30e0eaa71ce59012241","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"prompted-spatial-output","title":"提示怎样指定这一次要分的区域？"},"page":{"id":"reading-opening","title":"MAE 提供起点，掩码任务训练学会怎样响应提示","number":1},"sectionTitle":"MAE 提供起点，掩码任务训练学会怎样响应提示","currentQuestion":"MAE 提供起点，掩码任务训练学会怎样响应提示","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-f7aac8f5ebed29807ea7c067","nextBlock":"b-08040fbec867229eb8828cb1","prerequisites":[{"id":"b-ab6fd575202cd6a355e14eb6","url":"#b-ab6fd575202cd6a355e14eb6","role":"section-title"},{"id":"b-512e94b0c5990fc5e166403e","url":"#b-512e94b0c5990fc5e166403e","role":"problem-context"}],"relatedBlocks":[{"id":"b-24e585dca580f34f722537c8","url":"#b-24e585dca580f34f722537c8","relation":"figure-and-explanation"}]},{"id":"b-08040fbec867229eb8828cb1","type":"p","scope":"sam-task-training","url":"https://codingai-lec03.pages.dev/course/blocks/sam-task-training.html#b-08040fbec867229eb8828cb1","content":"数据组织也属于能力来源。SA-1B 来自模型辅助标注、半自动补充、全自动生成的逐步演进，最终收集约 1100 万张图像与 11 亿个 mask。这里的数据单位是区域轮廓，并非现成的图文语句对。","sha256":"288b13fb11a7b4bf50ae1b07b0d59fe594cb1b283e580e44cb03d1c60f4dcdf6","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"prompted-spatial-output","title":"提示怎样指定这一次要分的区域？"},"page":{"id":"reading-opening","title":"MAE 提供起点，掩码任务训练学会怎样响应提示","number":1},"sectionTitle":"MAE 提供起点，掩码任务训练学会怎样响应提示","currentQuestion":"MAE 提供起点，掩码任务训练学会怎样响应提示","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-0e83c4569394dab3c0552443","nextBlock":"b-a83a5afa0ed74ca552a19c22","prerequisites":[{"id":"b-ab6fd575202cd6a355e14eb6","url":"#b-ab6fd575202cd6a355e14eb6","role":"section-title"},{"id":"b-512e94b0c5990fc5e166403e","url":"#b-512e94b0c5990fc5e166403e","role":"problem-context"}],"relatedBlocks":[{"id":"b-24e585dca580f34f722537c8","url":"#b-24e585dca580f34f722537c8","relation":"figure-and-explanation"}]},{"id":"b-a83a5afa0ed74ca552a19c22","type":"p","scope":"sam-task-training","url":"https://codingai-lec03.pages.dev/course/blocks/sam-task-training.html#b-a83a5afa0ed74ca552a19c22","content":"预训练参数、任务监督和使用接口可以分别复用：先以自监督表示为起点，再学习图像/提示到掩码，使用时缓存同图表示。如果只说“MAE 会补图所以自然会分割”，就跳过了任务训练这一关键桥梁。注意此处是原始 SAM；文字提示在论文中有另行探索，不等于公开几何提示 checkpoint 默认已具备任意语言理解。","sha256":"d09830d2ab723a8ef4a85da99e5bb96bc3f95fcd6e26eeaad7e1b354c766da11","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"prompted-spatial-output","title":"提示怎样指定这一次要分的区域？"},"page":{"id":"reading-opening","title":"MAE 提供起点，掩码任务训练学会怎样响应提示","number":1},"sectionTitle":"MAE 提供起点，掩码任务训练学会怎样响应提示","currentQuestion":"MAE 提供起点，掩码任务训练学会怎样响应提示","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-08040fbec867229eb8828cb1","nextBlock":"b-24e585dca580f34f722537c8","prerequisites":[{"id":"b-ab6fd575202cd6a355e14eb6","url":"#b-ab6fd575202cd6a355e14eb6","role":"section-title"},{"id":"b-512e94b0c5990fc5e166403e","url":"#b-512e94b0c5990fc5e166403e","role":"problem-context"}],"relatedBlocks":[{"id":"b-24e585dca580f34f722537c8","url":"#b-24e585dca580f34f722537c8","relation":"figure-and-explanation"}]},{"id":"b-24e585dca580f34f722537c8","type":"figure","scope":"sam-task-training","url":"https://codingai-lec03.pages.dev/course/blocks/sam-task-training.html#b-24e585dca580f34f722537c8","content":"图内标签与数值：Lecture 03 课堂图解 P47；训练：初始化 ≠ 完整分割能力；MAE 预训练；提供图像编码器起点；分割任务训练；SA-1B 掩码 + 提示监督；SAM；图像 / 提示 / 掩码模块；使用：图像编码可复用；图像表示；新点提示 → 一个掩码；新框提示 → 另一个掩码\n\n初始化、任务训练与使用时缓存是三个不同阶段。 查看对应图解状态\n\n- 训练：初始化 ≠ 完整分割能力\n- MAE 预训练\n- 提供图像编码器起点\n- 分割任务训练\n- SA-1B 掩码 + 提示监督\n- SAM\n- 图像 / 提示 / 掩码模块\n- 使用：图像编码可复用\n- 图像表示\n- 新点提示 → 一个掩码\n- 新框提示 → 另一个掩码初始化、任务训练与使用时缓存是三个不同阶段。 [查看对应图解状态](/#/l3-r2-p47/0)[查看完整图解](/#/l3-r2-p47/0)","sha256":"05e1a3d0840ee29fa44fc66e8c9542cc733145315abd9168ef5dae4f8aaa3052","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"prompted-spatial-output","title":"提示怎样指定这一次要分的区域？"},"page":{"id":"reading-opening","title":"MAE 提供起点，掩码任务训练学会怎样响应提示","number":1},"sectionTitle":"MAE 提供起点，掩码任务训练学会怎样响应提示","currentQuestion":"MAE 提供起点，掩码任务训练学会怎样响应提示","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"figure":{"caption":"初始化、任务训练与使用时缓存是三个不同阶段。 查看对应图解状态","labels":["训练：初始化 ≠ 完整分割能力","MAE 预训练","提供图像编码器起点","分割任务训练","SA-1B 掩码 + 提示监督","SAM","图像 / 提示 / 掩码模块","使用：图像编码可复用","图像表示","新点提示 → 一个掩码","新框提示 → 另一个掩码"],"description":"初始化、任务训练与使用时缓存是三个不同阶段。 查看对应图解状态训练：初始化 ≠ 完整分割能力MAE 预训练提供图像编码器起点分割任务训练SA-1B 掩码 + 提示监督SAM图像 / 提示 / 掩码模块使用：图像编码可复用图像表示新点提示 → 一个掩码新框提示 → 另一个掩码","evidenceIdentity":"初始化、任务训练与使用时缓存是三个不同阶段。 查看对应图解状态","images":[{"src":"/assets/display/8ced939a1fd55047fbf73ea5.webp","alt":""}]},"previousBlock":"b-a83a5afa0ed74ca552a19c22","nextBlock":"b-48337a685d805af2845d80af","prerequisites":[{"id":"b-ab6fd575202cd6a355e14eb6","url":"#b-ab6fd575202cd6a355e14eb6","role":"section-title"},{"id":"b-512e94b0c5990fc5e166403e","url":"#b-512e94b0c5990fc5e166403e","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-48337a685d805af2845d80af","type":"h2","scope":"sam-task-training","url":"https://codingai-lec03.pages.dev/course/blocks/sam-task-training.html#b-48337a685d805af2845d80af","content":"## 继续阅读与实验","sha256":"2b9ccef7e1220e2250cceff1abd8bdd736b42f8da5a223f26b029f44b962e807","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"prompted-spatial-output","title":"提示怎样指定这一次要分的区域？"},"page":{"id":"reading-opening","title":"MAE 提供起点，掩码任务训练学会怎样响应提示","number":1},"sectionTitle":"MAE 提供起点，掩码任务训练学会怎样响应提示","currentQuestion":"MAE 提供起点，掩码任务训练学会怎样响应提示","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-24e585dca580f34f722537c8","nextBlock":null,"prerequisites":[{"id":"b-ab6fd575202cd6a355e14eb6","url":"#b-ab6fd575202cd6a355e14eb6","role":"section-title"},{"id":"b-512e94b0c5990fc5e166403e","url":"#b-512e94b0c5990fc5e166403e","role":"problem-context"}],"relatedBlocks":[{"id":"b-24e585dca580f34f722537c8","url":"#b-24e585dca580f34f722537c8","relation":"figure-and-explanation"}]}]}