{"version":2,"relationshipURLBase":"https://codingai-lec03.pages.dev/course/blocks/mae-masked-loss-denominator.html","relationshipResolution":"Resolve relative relationship URLs against relationshipURLBase; complete records and absolute URLs are in blocks.","title":"只在隐藏位置算损失，分母也要一致","url":"https://codingai-lec03.pages.dev/course/blocks/mae-masked-loss-denominator.html","markdown":"https://codingai-lec03.pages.dev/course/blocks/mae-masked-loss-denominator.md","pages":[{"id":"reading-opening","title":"只在隐藏位置算损失，分母也要一致","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-34caec975ef78467614459e7","targetId":"b-34caec975ef78467614459e7","page":"reading-opening","operation":"move","beforeSHA256":"eae23d5eeb78ff444844389ed0747b8e1270853eb8f75b0f3cbb7777e9e0af0f","afterSHA256":"eae23d5eeb78ff444844389ed0747b8e1270853eb8f75b0f3cbb7777e9e0af0f","reason":"Same source body in authored semantic page"},{"oldId":"b-5b9a34a57656f7105fb6e052","targetId":"b-5b9a34a57656f7105fb6e052","page":"reading-opening","operation":"move","beforeSHA256":"aa49ae401ddbd7e2d722e8c4b5c79e44357422e339b0bfce9729a2c00bb5f2ae","afterSHA256":"aa49ae401ddbd7e2d722e8c4b5c79e44357422e339b0bfce9729a2c00bb5f2ae","reason":"Same source body in authored semantic page"},{"oldId":"b-8bff5e68272ecd90f23256d3","targetId":"b-8bff5e68272ecd90f23256d3","page":"reading-opening","operation":"move","beforeSHA256":"e482018404dbe89a2de6e4a4a8a91074f75ca7ee4d82b34fc8e1edcf5b9e391c","afterSHA256":"e482018404dbe89a2de6e4a4a8a91074f75ca7ee4d82b34fc8e1edcf5b9e391c","reason":"Same source body in authored semantic page"},{"oldId":"b-d857586b7c6ce8221e0f017d","targetId":"b-d857586b7c6ce8221e0f017d","page":"reading-opening","operation":"move","beforeSHA256":"ba89d4a5e0a04dd83a9edafbf7dcab3afe7f0b478627768042f8f370dd1f31ae","afterSHA256":"ba89d4a5e0a04dd83a9edafbf7dcab3afe7f0b478627768042f8f370dd1f31ae","reason":"Same source body in authored semantic page"},{"oldId":"b-43ad759cfeab1b9672f4bdcf","targetId":"b-43ad759cfeab1b9672f4bdcf","page":"reading-opening","operation":"move","beforeSHA256":"1471ce4a8000b27c653742e3702754f6b6db46f6e878b8e12901808c4df9c073","afterSHA256":"1471ce4a8000b27c653742e3702754f6b6db46f6e878b8e12901808c4df9c073","reason":"Same source body in authored semantic page"},{"oldId":"b-7b9d1e8caed161ca7492b335","targetId":"b-7b9d1e8caed161ca7492b335","page":"reading-opening","operation":"move","beforeSHA256":"2b9ccef7e1220e2250cceff1abd8bdd736b42f8da5a223f26b029f44b962e807","afterSHA256":"2b9ccef7e1220e2250cceff1abd8bdd736b42f8da5a223f26b029f44b962e807","reason":"Same source body in authored semantic page"}],"blocks":[{"id":"b-34caec975ef78467614459e7","type":"h1","scope":"mae-masked-loss-denominator","url":"https://codingai-lec03.pages.dev/course/blocks/mae-masked-loss-denominator.html#b-34caec975ef78467614459e7","content":"# 只在隐藏位置算损失，分母也要一致","sha256":"eae23d5eeb78ff444844389ed0747b8e1270853eb8f75b0f3cbb7777e9e0af0f","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":"只在隐藏位置算损失，分母也要一致","number":1},"sectionTitle":"只在隐藏位置算损失，分母也要一致","currentQuestion":"只在隐藏位置算损失，分母也要一致","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-5b9a34a57656f7105fb6e052","prerequisites":[{"id":"b-5b9a34a57656f7105fb6e052","url":"#b-5b9a34a57656f7105fb6e052","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-5b9a34a57656f7105fb6e052","type":"p","scope":"mae-masked-loss-denominator","url":"https://codingai-lec03.pages.dev/course/blocks/mae-masked-loss-denominator.html#b-5b9a34a57656f7105fb6e052","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":"只在隐藏位置算损失，分母也要一致","number":1},"sectionTitle":"只在隐藏位置算损失，分母也要一致","currentQuestion":"只在隐藏位置算损失，分母也要一致","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-34caec975ef78467614459e7","nextBlock":"b-8bff5e68272ecd90f23256d3","prerequisites":[{"id":"b-34caec975ef78467614459e7","url":"#b-34caec975ef78467614459e7","role":"section-title"}],"relatedBlocks":[]},{"id":"b-8bff5e68272ecd90f23256d3","type":"p","scope":"mae-masked-loss-denominator","url":"https://codingai-lec03.pages.dev/course/blocks/mae-masked-loss-denominator.html#b-8bff5e68272ecd90f23256d3","content":"令隐藏集合为 $\\mathcal M$，每个patch含P个像素值。一个清晰的口径是先在patch内平均，再在隐藏patch间平均：$L=|\\mathcal M|^{-1}\\sum_{i\\in\\mathcal M}P^{-1}\\sum_{p=1}^{P}(\\hat x_{ip}-x_{ip})^2$。16块中隐藏12块时，外层分母就是12。","sha256":"e482018404dbe89a2de6e4a4a8a91074f75ca7ee4d82b34fc8e1edcf5b9e391c","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":"只在隐藏位置算损失，分母也要一致","number":1},"sectionTitle":"只在隐藏位置算损失，分母也要一致","currentQuestion":"只在隐藏位置算损失，分母也要一致","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"\\mathcal M","definitionContext":"只在隐藏位置算损失，分母也要一致"},{"latex":"L=|\\mathcal M|^{-1}\\sum_{i\\in\\mathcal M}P^{-1}\\sum_{p=1}^{P}(\\hat x_{ip}-x_{ip})^2","definitionContext":"只在隐藏位置算损失，分母也要一致"}],"previousBlock":"b-5b9a34a57656f7105fb6e052","nextBlock":"b-d857586b7c6ce8221e0f017d","prerequisites":[{"id":"b-34caec975ef78467614459e7","url":"#b-34caec975ef78467614459e7","role":"section-title"},{"id":"b-5b9a34a57656f7105fb6e052","url":"#b-5b9a34a57656f7105fb6e052","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-d857586b7c6ce8221e0f017d","type":"p","scope":"mae-masked-loss-denominator","url":"https://codingai-lec03.pages.dev/course/blocks/mae-masked-loss-denominator.html#b-d857586b7c6ce8221e0f017d","content":"若把已知可见像素直接拼回预测图，它们误差天然为零。此时对整图平均会使指标更小，却没有改善隐藏预测。图像面板可以为方便阅读拼回可见真值，但必须说明；数值则继续评价隐藏集合。","sha256":"ba89d4a5e0a04dd83a9edafbf7dcab3afe7f0b478627768042f8f370dd1f31ae","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":"只在隐藏位置算损失，分母也要一致","number":1},"sectionTitle":"只在隐藏位置算损失，分母也要一致","currentQuestion":"只在隐藏位置算损失，分母也要一致","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-8bff5e68272ecd90f23256d3","nextBlock":"b-43ad759cfeab1b9672f4bdcf","prerequisites":[{"id":"b-34caec975ef78467614459e7","url":"#b-34caec975ef78467614459e7","role":"section-title"},{"id":"b-5b9a34a57656f7105fb6e052","url":"#b-5b9a34a57656f7105fb6e052","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-43ad759cfeab1b9672f4bdcf","type":"p","scope":"mae-masked-loss-denominator","url":"https://codingai-lec03.pages.dev/course/blocks/mae-masked-loss-denominator.html#b-43ad759cfeab1b9672f4bdcf","content":"原论文还讨论对每个patch的像素做归一化后作为目标。比较不同实验时，要同时核对预测位置、像素缩放及归一化；同样叫“MSE”的数，分母和目标一变就未必可比。","sha256":"1471ce4a8000b27c653742e3702754f6b6db46f6e878b8e12901808c4df9c073","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":"只在隐藏位置算损失，分母也要一致","number":1},"sectionTitle":"只在隐藏位置算损失，分母也要一致","currentQuestion":"只在隐藏位置算损失，分母也要一致","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-d857586b7c6ce8221e0f017d","nextBlock":"b-7b9d1e8caed161ca7492b335","prerequisites":[{"id":"b-34caec975ef78467614459e7","url":"#b-34caec975ef78467614459e7","role":"section-title"},{"id":"b-5b9a34a57656f7105fb6e052","url":"#b-5b9a34a57656f7105fb6e052","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-7b9d1e8caed161ca7492b335","type":"h2","scope":"mae-masked-loss-denominator","url":"https://codingai-lec03.pages.dev/course/blocks/mae-masked-loss-denominator.html#b-7b9d1e8caed161ca7492b335","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":"只在隐藏位置算损失，分母也要一致","number":1},"sectionTitle":"只在隐藏位置算损失，分母也要一致","currentQuestion":"只在隐藏位置算损失，分母也要一致","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-43ad759cfeab1b9672f4bdcf","nextBlock":null,"prerequisites":[{"id":"b-34caec975ef78467614459e7","url":"#b-34caec975ef78467614459e7","role":"section-title"},{"id":"b-5b9a34a57656f7105fb6e052","url":"#b-5b9a34a57656f7105fb6e052","role":"problem-context"}],"relatedBlocks":[]}]}