{"version":2,"relationshipURLBase":"https://codingai-lec03.pages.dev/course/notebook-blocks/18ba37010b894b2893332630cf8b3070.html","relationshipResolution":"Resolve relative relationship URLs against relationshipURLBase; complete records and absolute URLs are in blocks.","title":"与原MAE逐项对齐","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/18ba37010b894b2893332630cf8b3070.html","markdown":"https://codingai-lec03.pages.dev/course/notebook-blocks/18ba37010b894b2893332630cf8b3070.md","pages":[{"id":"reading-opening","title":"与原MAE逐项对齐","unit":"与原MAE逐项对齐","unitId":"notebook-blocks/18ba37010b894b2893332630cf8b3070","chapter":"course","number":1}],"aliases":[{"oldId":"b-05a0557e22f0ab88bb639a9a","targetId":"b-05a0557e22f0ab88bb639a9a","page":"reading-opening","operation":"move","beforeSHA256":"4730e90285e279b2564c179018b9df76aafd9db5dd2729f0f687ac380af591ed","afterSHA256":"4730e90285e279b2564c179018b9df76aafd9db5dd2729f0f687ac380af591ed","reason":"Same source body in authored semantic page"},{"oldId":"b-70d050f563823a9cb0041275","targetId":"b-70d050f563823a9cb0041275","page":"reading-opening","operation":"move","beforeSHA256":"4706db814514ddbeda7e689dd620d749fe5b748b37a11687dd5fccd9f7b3902c","afterSHA256":"4706db814514ddbeda7e689dd620d749fe5b748b37a11687dd5fccd9f7b3902c","reason":"Same source body in authored semantic page"},{"oldId":"b-7a2382b3ed930da1ae8575b0","targetId":"b-7a2382b3ed930da1ae8575b0","page":"reading-opening","operation":"move","beforeSHA256":"e91f24e4e4ca40e2099a0d81d30db6568f4bf2074b499eeee6e3959617bd97ac","afterSHA256":"e91f24e4e4ca40e2099a0d81d30db6568f4bf2074b499eeee6e3959617bd97ac","reason":"Same source body in authored semantic page"},{"oldId":"b-e8e73551faf8d117019f3cc7","targetId":"b-e8e73551faf8d117019f3cc7","page":"reading-opening","operation":"move","beforeSHA256":"0f845ef31b98ed8c96f689fea3088b0a7482be32b2d5510c83f9fcbb05da4c9d","afterSHA256":"0f845ef31b98ed8c96f689fea3088b0a7482be32b2d5510c83f9fcbb05da4c9d","reason":"Same source body in authored semantic page"},{"oldId":"b-b76016ab76a0c4519edc4baa","targetId":"b-b76016ab76a0c4519edc4baa","page":"reading-opening","operation":"move","beforeSHA256":"4f73b68ffb171baaf3e8ee901f1f998d5f303e020953190ae14e5ebf9490b28d","afterSHA256":"4f73b68ffb171baaf3e8ee901f1f998d5f303e020953190ae14e5ebf9490b28d","reason":"Same source body in authored semantic page"},{"oldId":"b-fc43e6ad4f58c1a61a4c2b2b","targetId":"b-fc43e6ad4f58c1a61a4c2b2b","page":"reading-opening","operation":"rewrite","beforeSHA256":"b698b3c0b2c289e3361a9f965f15fd604390a5b0c63e820d3597c8e79d8db761","afterSHA256":"17d2aced04bdd810d8e50a9eecd85a7ba53c403c24a82f75f269a691b99b9509","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-ef8b09b1a865d4d358c5aab2","targetId":"b-ef8b09b1a865d4d358c5aab2","page":"reading-opening","operation":"move","beforeSHA256":"f36623b4a66bc23f3e6de0948398eac4da9de12f2b1cb7dee15f169d61a3c0af","afterSHA256":"f36623b4a66bc23f3e6de0948398eac4da9de12f2b1cb7dee15f169d61a3c0af","reason":"Same source body in authored semantic page"},{"oldId":"b-67eaf7af86e903a46bf53728","targetId":"b-67eaf7af86e903a46bf53728","page":"reading-opening","operation":"move","beforeSHA256":"a0876f8de075bc82b9ef180984eea28985c8ea153021910f079bdd919e4c0538","afterSHA256":"a0876f8de075bc82b9ef180984eea28985c8ea153021910f079bdd919e4c0538","reason":"Same source body in authored semantic page"},{"oldId":"b-e4d189019a25a8ca6babd69e","targetId":"b-e4d189019a25a8ca6babd69e","page":"reading-opening","operation":"move","beforeSHA256":"1d07670d7c0674d4a921d9458ac6aca1e7f8c2c46deb6309e83d78c9c8146c58","afterSHA256":"1d07670d7c0674d4a921d9458ac6aca1e7f8c2c46deb6309e83d78c9c8146c58","reason":"Same source body in authored semantic page"},{"oldId":"block-18ba37010b894b2893332630cf8b3070--与原mae逐项对齐","targetId":"b-b76016ab76a0c4519edc4baa","operation":"retain-source-identity"}],"blocks":[{"id":"b-05a0557e22f0ab88bb639a9a","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/18ba37010b894b2893332630cf8b3070.html#b-05a0557e22f0ab88bb639a9a","content":"[LAB 08 · 图像重建：遮挡损失与表示评价](/course/notebooks/b5ba27f2c5ba4ec0922208a7d5dc80a4.html)","sha256":"4730e90285e279b2564c179018b9df76aafd9db5dd2729f0f687ac380af591ed","chapter":{"id":"course","title":"与原MAE逐项对齐"},"unit":{"id":"notebook-blocks/18ba37010b894b2893332630cf8b3070","title":"与原MAE逐项对齐"},"page":{"id":"reading-opening","title":"与原MAE逐项对齐","number":1},"sectionTitle":"与原MAE逐项对齐","currentQuestion":"与原MAE逐项对齐","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-70d050f563823a9cb0041275","prerequisites":[],"relatedBlocks":[]},{"id":"b-70d050f563823a9cb0041275","type":"h1","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/18ba37010b894b2893332630cf8b3070.html#b-70d050f563823a9cb0041275","content":"# 与原MAE逐项对齐","sha256":"4706db814514ddbeda7e689dd620d749fe5b748b37a11687dd5fccd9f7b3902c","chapter":{"id":"course","title":"与原MAE逐项对齐"},"unit":{"id":"notebook-blocks/18ba37010b894b2893332630cf8b3070","title":"与原MAE逐项对齐"},"page":{"id":"reading-opening","title":"与原MAE逐项对齐","number":1},"sectionTitle":"与原MAE逐项对齐","currentQuestion":"与原MAE逐项对齐","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-05a0557e22f0ab88bb639a9a","nextBlock":"b-7a2382b3ed930da1ae8575b0","prerequisites":[{"id":"b-05a0557e22f0ab88bb639a9a","url":"#b-05a0557e22f0ab88bb639a9a","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-7a2382b3ed930da1ae8575b0","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/18ba37010b894b2893332630cf8b3070.html#b-7a2382b3ed930da1ae8575b0","content":"按问题继续","sha256":"e91f24e4e4ca40e2099a0d81d30db6568f4bf2074b499eeee6e3959617bd97ac","chapter":{"id":"course","title":"与原MAE逐项对齐"},"unit":{"id":"notebook-blocks/18ba37010b894b2893332630cf8b3070","title":"与原MAE逐项对齐"},"page":{"id":"reading-opening","title":"与原MAE逐项对齐","number":1},"sectionTitle":"与原MAE逐项对齐","currentQuestion":"与原MAE逐项对齐","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-70d050f563823a9cb0041275","nextBlock":"b-e8e73551faf8d117019f3cc7","prerequisites":[{"id":"b-70d050f563823a9cb0041275","url":"#b-70d050f563823a9cb0041275","role":"section-title"},{"id":"b-05a0557e22f0ab88bb639a9a","url":"#b-05a0557e22f0ab88bb639a9a","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-e8e73551faf8d117019f3cc7","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/18ba37010b894b2893332630cf8b3070.html#b-e8e73551faf8d117019f3cc7","content":"索引恢复、去噪/遮挡ConvAE训练、PCA基线、独立冻结评价","sha256":"0f845ef31b98ed8c96f689fea3088b0a7482be32b2d5510c83f9fcbb05da4c9d","chapter":{"id":"course","title":"与原MAE逐项对齐"},"unit":{"id":"notebook-blocks/18ba37010b894b2893332630cf8b3070","title":"与原MAE逐项对齐"},"page":{"id":"reading-opening","title":"与原MAE逐项对齐","number":1},"sectionTitle":"与原MAE逐项对齐","currentQuestion":"与原MAE逐项对齐","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-7a2382b3ed930da1ae8575b0","nextBlock":"b-b76016ab76a0c4519edc4baa","prerequisites":[{"id":"b-70d050f563823a9cb0041275","url":"#b-70d050f563823a9cb0041275","role":"section-title"},{"id":"b-05a0557e22f0ab88bb639a9a","url":"#b-05a0557e22f0ab88bb639a9a","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-b76016ab76a0c4519edc4baa","type":"h2","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/18ba37010b894b2893332630cf8b3070.html#b-b76016ab76a0c4519edc4baa","content":"## 与原MAE逐项对齐","sha256":"4f73b68ffb171baaf3e8ee901f1f998d5f303e020953190ae14e5ebf9490b28d","chapter":{"id":"course","title":"与原MAE逐项对齐"},"unit":{"id":"notebook-blocks/18ba37010b894b2893332630cf8b3070","title":"与原MAE逐项对齐"},"page":{"id":"reading-opening","title":"与原MAE逐项对齐","number":1},"sectionTitle":"与原MAE逐项对齐","currentQuestion":"与原MAE逐项对齐","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-e8e73551faf8d117019f3cc7","nextBlock":"b-fc43e6ad4f58c1a61a4c2b2b","prerequisites":[{"id":"b-70d050f563823a9cb0041275","url":"#b-70d050f563823a9cb0041275","role":"section-title"},{"id":"b-05a0557e22f0ab88bb639a9a","url":"#b-05a0557e22f0ab88bb639a9a","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-fc43e6ad4f58c1a61a4c2b2b","type":"table","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/18ba37010b894b2893332630cf8b3070.html#b-fc43e6ad4f58c1a61a4c2b2b","content":"| 环节 | 本册masked ConvAE | 原MAE关键设计 |\n| --- | --- | --- |\n| 编码器输入 | 可见像素+mask通道，完整8×8网格 | 只输入可见patch的token |\n| 编码器 | 小卷积网络、16维瓶颈 | ViT编码器 |\n| 位置恢复 | 索引热身精确演示；主模型保持网格位置 | 解码端恢复token顺序并加入mask token |\n| 重建损失 | 隐藏像素MSE | masked patch的像素重建目标 |\n| 计算量随mask变化 | 本实现基本不减少 | 可见token更少会减少编码器计算 |\n| 表示评价 | 小数字集冻结5-NN | 论文有大规模分类与迁移评价 |","sha256":"17d2aced04bdd810d8e50a9eecd85a7ba53c403c24a82f75f269a691b99b9509","chapter":{"id":"course","title":"与原MAE逐项对齐"},"unit":{"id":"notebook-blocks/18ba37010b894b2893332630cf8b3070","title":"与原MAE逐项对齐"},"page":{"id":"reading-opening","title":"与原MAE逐项对齐","number":1},"sectionTitle":"与原MAE逐项对齐","currentQuestion":"与原MAE逐项对齐","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"table":{"headers":["环节","本册masked ConvAE","原MAE关键设计"],"rows":[["环节","本册masked ConvAE","原MAE关键设计"],["编码器输入","可见像素+mask通道，完整8×8网格","只输入可见patch的token"],["编码器","小卷积网络、16维瓶颈","ViT编码器"],["位置恢复","索引热身精确演示；主模型保持网格位置","解码端恢复token顺序并加入mask token"],["重建损失","隐藏像素MSE","masked patch的像素重建目标"],["计算量随mask变化","本实现基本不减少","可见token更少会减少编码器计算"],["表示评价","小数字集冻结5-NN","论文有大规模分类与迁移评价"]]},"previousBlock":"b-b76016ab76a0c4519edc4baa","nextBlock":"b-ef8b09b1a865d4d358c5aab2","prerequisites":[{"id":"b-b76016ab76a0c4519edc4baa","url":"#b-b76016ab76a0c4519edc4baa","role":"section-title"},{"id":"b-05a0557e22f0ab88bb639a9a","url":"#b-05a0557e22f0ab88bb639a9a","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-ef8b09b1a865d4d358c5aab2","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/18ba37010b894b2893332630cf8b3070.html#b-ef8b09b1a865d4d358c5aab2","content":"因此本册完成了真实图像遮挡重建与单独的表示评价，但没有把卷积模型重命名为原MAE复现。索引热身、masked ConvAE和MAE论文提供三种不同粒度的证据。","sha256":"f36623b4a66bc23f3e6de0948398eac4da9de12f2b1cb7dee15f169d61a3c0af","chapter":{"id":"course","title":"与原MAE逐项对齐"},"unit":{"id":"notebook-blocks/18ba37010b894b2893332630cf8b3070","title":"与原MAE逐项对齐"},"page":{"id":"reading-opening","title":"与原MAE逐项对齐","number":1},"sectionTitle":"与原MAE逐项对齐","currentQuestion":"与原MAE逐项对齐","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-fc43e6ad4f58c1a61a4c2b2b","nextBlock":"b-67eaf7af86e903a46bf53728","prerequisites":[{"id":"b-b76016ab76a0c4519edc4baa","url":"#b-b76016ab76a0c4519edc4baa","role":"section-title"},{"id":"b-05a0557e22f0ab88bb639a9a","url":"#b-05a0557e22f0ab88bb639a9a","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-67eaf7af86e903a46bf53728","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/18ba37010b894b2893332630cf8b3070.html#b-67eaf7af86e903a46bf53728","content":"可做的条件修改：保持训练权重，先改变 `patch_mask` 的评估遮挡比例，维持相同随机排列；再选择一个新实验，把训练遮挡率改为75%，另存新目录。先预测低可见证据会怎样改变细笔画和读出，不要只寻找一张好看的重建图。","sha256":"a0876f8de075bc82b9ef180984eea28985c8ea153021910f079bdd919e4c0538","chapter":{"id":"course","title":"与原MAE逐项对齐"},"unit":{"id":"notebook-blocks/18ba37010b894b2893332630cf8b3070","title":"与原MAE逐项对齐"},"page":{"id":"reading-opening","title":"与原MAE逐项对齐","number":1},"sectionTitle":"与原MAE逐项对齐","currentQuestion":"与原MAE逐项对齐","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-ef8b09b1a865d4d358c5aab2","nextBlock":"b-e4d189019a25a8ca6babd69e","prerequisites":[{"id":"b-b76016ab76a0c4519edc4baa","url":"#b-b76016ab76a0c4519edc4baa","role":"section-title"},{"id":"b-05a0557e22f0ab88bb639a9a","url":"#b-05a0557e22f0ab88bb639a9a","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-e4d189019a25a8ca6babd69e","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/18ba37010b894b2893332630cf8b3070.html#b-e4d189019a25a8ca6babd69e","content":"来源：[去噪自编码器原论文](https://www.cs.toronto.edu/~larocheh/publications/icml-2008-denoising-autoencoders.pdf)、[MAE原论文](https://openaccess.thecvf.com/content/CVPR2022/html/He_Masked_Autoencoders_Are_Scalable_Vision_Learners_CVPR_2022_paper.html)、[MAE作者实现的mask/restore/loss](https://github.com/facebookresearch/mae/blob/main/models_mae.py)、[scikit-learn digits数据说明](https://scikit-learn.org/stable/modules/generated/sklearn.datasets.load_digits.html)。本册模型结构、训练权重、图和测量来自本轮CPU运行。","sha256":"1d07670d7c0674d4a921d9458ac6aca1e7f8c2c46deb6309e83d78c9c8146c58","chapter":{"id":"course","title":"与原MAE逐项对齐"},"unit":{"id":"notebook-blocks/18ba37010b894b2893332630cf8b3070","title":"与原MAE逐项对齐"},"page":{"id":"reading-opening","title":"与原MAE逐项对齐","number":1},"sectionTitle":"与原MAE逐项对齐","currentQuestion":"与原MAE逐项对齐","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-67eaf7af86e903a46bf53728","nextBlock":null,"prerequisites":[{"id":"b-b76016ab76a0c4519edc4baa","url":"#b-b76016ab76a0c4519edc4baa","role":"section-title"},{"id":"b-05a0557e22f0ab88bb639a9a","url":"#b-05a0557e22f0ab88bb639a9a","role":"problem-context"}],"relatedBlocks":[]}]}
