{"version":2,"relationshipURLBase":"https://codingai-lec03.pages.dev/course/blocks/mae-shuffle-restore-worked-example.html","relationshipResolution":"Resolve relative relationship URLs against relationshipURLBase; complete records and absolute URLs are in blocks.","title":"补回位置与补回像素，是两件事","url":"https://codingai-lec03.pages.dev/course/blocks/mae-shuffle-restore-worked-example.html","markdown":"https://codingai-lec03.pages.dev/course/blocks/mae-shuffle-restore-worked-example.md","pages":[{"id":"reading-opening","title":"补回位置与补回像素，是两件事","unit":"限制可见信息，怎样构成可验证的重建任务？","unitId":"i-masked-prediction","chapter":"I","number":1}],"aliases":[{"oldId":"b-752ac736084e9f7d7acadc57","targetId":"b-752ac736084e9f7d7acadc57","page":"reading-opening","operation":"move","beforeSHA256":"13fc857ebcced0120990575490f9a3a017ed6805c4938f6b6dd6159ebb0a6a6c","afterSHA256":"13fc857ebcced0120990575490f9a3a017ed6805c4938f6b6dd6159ebb0a6a6c","reason":"Same source body in authored semantic page"},{"oldId":"b-1ce3f669b499e32884d074a0","targetId":"b-1ce3f669b499e32884d074a0","page":"reading-opening","operation":"move","beforeSHA256":"aa49ae401ddbd7e2d722e8c4b5c79e44357422e339b0bfce9729a2c00bb5f2ae","afterSHA256":"aa49ae401ddbd7e2d722e8c4b5c79e44357422e339b0bfce9729a2c00bb5f2ae","reason":"Same source body in authored semantic page"},{"oldId":"b-b44f80db013a5804d735904c","targetId":"b-b44f80db013a5804d735904c","page":"reading-opening","operation":"move","beforeSHA256":"93c9a763d7a471efada8debd8c55d13c7a88fe0cd6661a7baf997fe14a5ef0fa","afterSHA256":"93c9a763d7a471efada8debd8c55d13c7a88fe0cd6661a7baf997fe14a5ef0fa","reason":"Same source body in authored semantic page"},{"oldId":"b-260bf96e5ee917092234a206","targetId":"b-260bf96e5ee917092234a206","page":"reading-opening","operation":"move","beforeSHA256":"7de0739a39bb7b319d607392cdf7ba40d0d0602f0380cbeff4eabe1a19d402f2","afterSHA256":"7de0739a39bb7b319d607392cdf7ba40d0d0602f0380cbeff4eabe1a19d402f2","reason":"Same source body in authored semantic page"},{"oldId":"b-b57fc9d66faa0578a08e656f","targetId":"b-b57fc9d66faa0578a08e656f","page":"reading-opening","operation":"move","beforeSHA256":"2e5521ede2f6dd07a26bd91b6a70919737fa9337d6a5c0631ecda8c61f1d6c8b","afterSHA256":"2e5521ede2f6dd07a26bd91b6a70919737fa9337d6a5c0631ecda8c61f1d6c8b","reason":"Same source body in authored semantic page"},{"oldId":"b-983fa2ea245fa3bfe62dc392","targetId":"b-983fa2ea245fa3bfe62dc392","page":"reading-opening","operation":"rewrite","beforeSHA256":"b9493baa29654d8708ddbf1791bfbb6c96230d5245264ec2be05c03e6053b728","afterSHA256":"bb200b55086dad763c20ec927b68bf99821f63cdcf6152c0fe035c34ac4e261a","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-2ac982231db228cb0c032b21","targetId":"b-2ac982231db228cb0c032b21","page":"reading-opening","operation":"move","beforeSHA256":"2b9ccef7e1220e2250cceff1abd8bdd736b42f8da5a223f26b029f44b962e807","afterSHA256":"2b9ccef7e1220e2250cceff1abd8bdd736b42f8da5a223f26b029f44b962e807","reason":"Same source body in authored semantic page"}],"blocks":[{"id":"b-752ac736084e9f7d7acadc57","type":"h1","scope":"mae-shuffle-restore-worked-example","url":"https://codingai-lec03.pages.dev/course/blocks/mae-shuffle-restore-worked-example.html#b-752ac736084e9f7d7acadc57","content":"# 补回位置与补回像素，是两件事","sha256":"13fc857ebcced0120990575490f9a3a017ed6805c4938f6b6dd6159ebb0a6a6c","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-masked-prediction","title":"限制可见信息，怎样构成可验证的重建任务？"},"page":{"id":"reading-opening","title":"补回位置与补回像素，是两件事","number":1},"sectionTitle":"补回位置与补回像素，是两件事","currentQuestion":"补回位置与补回像素，是两件事","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-1ce3f669b499e32884d074a0","prerequisites":[{"id":"b-1ce3f669b499e32884d074a0","url":"#b-1ce3f669b499e32884d074a0","role":"problem-context"}],"relatedBlocks":[{"id":"b-983fa2ea245fa3bfe62dc392","url":"#b-983fa2ea245fa3bfe62dc392","relation":"figure-and-explanation"}]},{"id":"b-1ce3f669b499e32884d074a0","type":"p","scope":"mae-shuffle-restore-worked-example","url":"https://codingai-lec03.pages.dev/course/blocks/mae-shuffle-restore-worked-example.html#b-1ce3f669b499e32884d074a0","content":"解码阶段恢复位置，再预测内容；目标从原图来，误差只在隐藏集合比较。","sha256":"aa49ae401ddbd7e2d722e8c4b5c79e44357422e339b0bfce9729a2c00bb5f2ae","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-masked-prediction","title":"限制可见信息，怎样构成可验证的重建任务？"},"page":{"id":"reading-opening","title":"补回位置与补回像素，是两件事","number":1},"sectionTitle":"补回位置与补回像素，是两件事","currentQuestion":"补回位置与补回像素，是两件事","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-752ac736084e9f7d7acadc57","nextBlock":"b-b44f80db013a5804d735904c","prerequisites":[{"id":"b-752ac736084e9f7d7acadc57","url":"#b-752ac736084e9f7d7acadc57","role":"section-title"}],"relatedBlocks":[{"id":"b-983fa2ea245fa3bfe62dc392","url":"#b-983fa2ea245fa3bfe62dc392","relation":"figure-and-explanation"}]},{"id":"b-b44f80db013a5804d735904c","type":"p","scope":"mae-shuffle-restore-worked-example","url":"https://codingai-lec03.pages.dev/course/blocks/mae-shuffle-restore-worked-example.html#b-b44f80db013a5804d735904c","content":"用4个位置缩小算例。原位置是[0,1,2,3]，随机顺序为[2,0,3,1]，取前两项进入编码器。得到[z2,z0]后，解码阶段先加两个mask项：[z2,z0,M,M]。逆排列为[1,3,0,2]，按它取值后成为[z0,M,z2,M]。","sha256":"93c9a763d7a471efada8debd8c55d13c7a88fe0cd6661a7baf997fe14a5ef0fa","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-masked-prediction","title":"限制可见信息，怎样构成可验证的重建任务？"},"page":{"id":"reading-opening","title":"补回位置与补回像素，是两件事","number":1},"sectionTitle":"补回位置与补回像素，是两件事","currentQuestion":"补回位置与补回像素，是两件事","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-1ce3f669b499e32884d074a0","nextBlock":"b-260bf96e5ee917092234a206","prerequisites":[{"id":"b-752ac736084e9f7d7acadc57","url":"#b-752ac736084e9f7d7acadc57","role":"section-title"},{"id":"b-1ce3f669b499e32884d074a0","url":"#b-1ce3f669b499e32884d074a0","role":"problem-context"}],"relatedBlocks":[{"id":"b-983fa2ea245fa3bfe62dc392","url":"#b-983fa2ea245fa3bfe62dc392","relation":"figure-and-explanation"}]},{"id":"b-260bf96e5ee917092234a206","type":"p","scope":"mae-shuffle-restore-worked-example","url":"https://codingai-lec03.pages.dev/course/blocks/mae-shuffle-restore-worked-example.html#b-260bf96e5ee917092234a206","content":"这时我们只是恢复了“谁在什么位置”。位置1和3仍未知，M不是正确像素；解码器要结合可见内容及各位置线索作预测，再与原图相应位置比较。","sha256":"7de0739a39bb7b319d607392cdf7ba40d0d0602f0380cbeff4eabe1a19d402f2","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-masked-prediction","title":"限制可见信息，怎样构成可验证的重建任务？"},"page":{"id":"reading-opening","title":"补回位置与补回像素，是两件事","number":1},"sectionTitle":"补回位置与补回像素，是两件事","currentQuestion":"补回位置与补回像素，是两件事","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-b44f80db013a5804d735904c","nextBlock":"b-b57fc9d66faa0578a08e656f","prerequisites":[{"id":"b-752ac736084e9f7d7acadc57","url":"#b-752ac736084e9f7d7acadc57","role":"section-title"},{"id":"b-1ce3f669b499e32884d074a0","url":"#b-1ce3f669b499e32884d074a0","role":"problem-context"}],"relatedBlocks":[{"id":"b-983fa2ea245fa3bfe62dc392","url":"#b-983fa2ea245fa3bfe62dc392","relation":"figure-and-explanation"}]},{"id":"b-b57fc9d66faa0578a08e656f","type":"p","scope":"mae-shuffle-restore-worked-example","url":"https://codingai-lec03.pages.dev/course/blocks/mae-shuffle-restore-worked-example.html#b-b57fc9d66faa0578a08e656f","content":"如果忘了逆排列，模型可能把正确形状放到错误坐标，loss也会对错位置。LAB 08的索引热身正好检查这个对应关系；后面的卷积重建才包含参数学习。不要把数组重排成功写成模型学会了补图。","sha256":"2e5521ede2f6dd07a26bd91b6a70919737fa9337d6a5c0631ecda8c61f1d6c8b","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-masked-prediction","title":"限制可见信息，怎样构成可验证的重建任务？"},"page":{"id":"reading-opening","title":"补回位置与补回像素，是两件事","number":1},"sectionTitle":"补回位置与补回像素，是两件事","currentQuestion":"补回位置与补回像素，是两件事","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-260bf96e5ee917092234a206","nextBlock":"b-983fa2ea245fa3bfe62dc392","prerequisites":[{"id":"b-752ac736084e9f7d7acadc57","url":"#b-752ac736084e9f7d7acadc57","role":"section-title"},{"id":"b-1ce3f669b499e32884d074a0","url":"#b-1ce3f669b499e32884d074a0","role":"problem-context"}],"relatedBlocks":[{"id":"b-983fa2ea245fa3bfe62dc392","url":"#b-983fa2ea245fa3bfe62dc392","relation":"figure-and-explanation"}]},{"id":"b-983fa2ea245fa3bfe62dc392","type":"figure","scope":"mae-shuffle-restore-worked-example","url":"https://codingai-lec03.pages.dev/course/blocks/mae-shuffle-restore-worked-example.html#b-983fa2ea245fa3bfe62dc392","content":"图内标签与数值：P21 学习表示机制图，状态 2；解码端：所有位置加入 pᵢ；隐藏位置共用一个 mask 向量 M；h0+p0；M+p1；M+p2；h3+p3；M+p4；M+p5；M+p6；M+p7；M+p8；h9+p9；M+p10；M+p11；M+p12；M+p13；h14+p14；M+p15；4 个 h + 12 个 M；轻量解码器；ŷ0；ŷ1；ŷ2；ŷ3；ŷ4；ŷ5；ŷ6；ŷ7；ŷ8；ŷ9；ŷ10；ŷ11；ŷ12；ŷ13；ŷ14；ŷ15；各位置的像素预测（符号示意）；原图目标 x；L = (1/12) Σ隐藏块 MSE(ŷᵢ, xᵢ)；橙框：12 个计入 loss 的位置；预训练后留下编码器；重建用解码器不必进入下游系统。\n\n可见表示和mask token回到完整位置，所有位置有各自的位置线索。 查看对应图解状态\n\n- 解码端：所有位置加入 pᵢ；隐藏位置共用一个 mask 向量 M\n- h0+p0\n- M+p1\n- M+p2\n- h3+p3\n- M+p4\n- M+p5\n- M+p6\n- M+p7\n- M+p8\n- h9+p9\n- M+p10\n- M+p11\n- M+p12\n- M+p13\n- h14+p14\n- M+p15\n- 4 个 h + 12 个 M\n- 轻量解码器\n- ŷ0\n- ŷ1\n- ŷ2\n- ŷ3\n- ŷ4\n- ŷ5\n- ŷ6\n- ŷ7\n- ŷ8\n- ŷ9\n- ŷ10\n- ŷ11\n- ŷ12\n- ŷ13\n- ŷ14\n- ŷ15\n- 各位置的像素预测（符号示意）\n- 原图目标 x\n- L = (1/12) Σ隐藏块 MSE(ŷᵢ, xᵢ)\n- 橙框：12 个计入 loss 的位置\n- 预训练后留下编码器；重建用解码器不必进入下游系统。可见表示和mask token回到完整位置，所有位置有各自的位置线索。 [查看对应图解状态](/#/l3-r2-p21/1)[查看完整图解](/#/l3-r2-p21/1)","sha256":"bb200b55086dad763c20ec927b68bf99821f63cdcf6152c0fe035c34ac4e261a","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-masked-prediction","title":"限制可见信息，怎样构成可验证的重建任务？"},"page":{"id":"reading-opening","title":"补回位置与补回像素，是两件事","number":1},"sectionTitle":"补回位置与补回像素，是两件事","currentQuestion":"补回位置与补回像素，是两件事","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"figure":{"caption":"可见表示和mask token回到完整位置，所有位置有各自的位置线索。 查看对应图解状态","labels":["解码端：所有位置加入 pᵢ；隐藏位置共用一个 mask 向量 M","h0+p0","M+p1","M+p2","h3+p3","M+p4","M+p5","M+p6","M+p7","M+p8","h9+p9","M+p10","M+p11","M+p12","M+p13","h14+p14","M+p15","4 个 h + 12 个 M","轻量解码器","ŷ0","ŷ1","ŷ2","ŷ3","ŷ4","ŷ5","ŷ6","ŷ7","ŷ8","ŷ9","ŷ10","ŷ11","ŷ12","ŷ13","ŷ14","ŷ15","各位置的像素预测（符号示意）","原图目标 x","L = (1/12) Σ隐藏块 MSE(ŷᵢ, xᵢ)","橙框：12 个计入 loss 的位置","预训练后留下编码器；重建用解码器不必进入下游系统。"],"description":"可见表示和mask token回到完整位置，所有位置有各自的位置线索。 查看对应图解状态解码端：所有位置加入 pᵢ；隐藏位置共用一个 mask 向量 Mh0+p0M+p1M+p2h3+p3M+p4M+p5M+p6M+p7M+p8h9+p9M+p10M+p11M+p12M+p13h14+p14M+p154 个 h + 12 个 M轻量解码器ŷ0ŷ1ŷ2ŷ3ŷ4ŷ5ŷ6ŷ7ŷ8ŷ9ŷ10ŷ11ŷ12ŷ13ŷ14ŷ15各位置的像素预测（符号示意）原图目标 xL = (1/12) Σ隐藏块 MSE(ŷᵢ, xᵢ)橙框：12 个计入 loss 的位置预训练后留下编码器；重建用解码器不必进入下游系统。","evidenceIdentity":"可见表示和mask token回到完整位置，所有位置有各自的位置线索。 查看对应图解状态","images":[{"src":"/assets/display/584c419ba431799ff41df983.webp","alt":""}]},"previousBlock":"b-b57fc9d66faa0578a08e656f","nextBlock":"b-2ac982231db228cb0c032b21","prerequisites":[{"id":"b-752ac736084e9f7d7acadc57","url":"#b-752ac736084e9f7d7acadc57","role":"section-title"},{"id":"b-1ce3f669b499e32884d074a0","url":"#b-1ce3f669b499e32884d074a0","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-2ac982231db228cb0c032b21","type":"h2","scope":"mae-shuffle-restore-worked-example","url":"https://codingai-lec03.pages.dev/course/blocks/mae-shuffle-restore-worked-example.html#b-2ac982231db228cb0c032b21","content":"## 继续阅读与实验","sha256":"2b9ccef7e1220e2250cceff1abd8bdd736b42f8da5a223f26b029f44b962e807","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-masked-prediction","title":"限制可见信息，怎样构成可验证的重建任务？"},"page":{"id":"reading-opening","title":"补回位置与补回像素，是两件事","number":1},"sectionTitle":"补回位置与补回像素，是两件事","currentQuestion":"补回位置与补回像素，是两件事","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-983fa2ea245fa3bfe62dc392","nextBlock":null,"prerequisites":[{"id":"b-752ac736084e9f7d7acadc57","url":"#b-752ac736084e9f7d7acadc57","role":"section-title"},{"id":"b-1ce3f669b499e32884d074a0","url":"#b-1ce3f669b499e32884d074a0","role":"problem-context"}],"relatedBlocks":[{"id":"b-983fa2ea245fa3bfe62dc392","url":"#b-983fa2ea245fa3bfe62dc392","relation":"figure-and-explanation"}]}]}