{"version":2,"relationshipURLBase":"https://codingai-lec03.pages.dev/course/notebook-blocks/fab7a942c9ab4a2ca8dc8d8c5d2877d9.html","relationshipResolution":"Resolve relative relationship URLs against relationshipURLBase; complete records and absolute URLs are in blocks.","title":"4. 推理只接自己的输出，再检查视觉依赖","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/fab7a942c9ab4a2ca8dc8d8c5d2877d9.html","markdown":"https://codingai-lec03.pages.dev/course/notebook-blocks/fab7a942c9ab4a2ca8dc8d8c5d2877d9.md","pages":[{"id":"reading-opening","title":"4. 推理只接自己的输出，再检查视觉依赖","unit":"4. 推理只接自己的输出，再检查视觉依赖","unitId":"notebook-blocks/fab7a942c9ab4a2ca8dc8d8c5d2877d9","chapter":"course","number":1}],"aliases":[{"oldId":"b-f946a4e0ad1d005faf36f974","targetId":"b-f946a4e0ad1d005faf36f974","page":"reading-opening","operation":"move","beforeSHA256":"db08960319942d9119c9cad2cf7d0b78a68c914f8e19ab3bd375531d27c96204","afterSHA256":"db08960319942d9119c9cad2cf7d0b78a68c914f8e19ab3bd375531d27c96204","reason":"Same source body in authored semantic page"},{"oldId":"b-e8820dd5a58ebc7170310586","targetId":"b-e8820dd5a58ebc7170310586","page":"reading-opening","operation":"move","beforeSHA256":"7d9a64bd2956d9687a40de5ee89d75652fc8abcfb269521cd0b8e06b0569abcb","afterSHA256":"7d9a64bd2956d9687a40de5ee89d75652fc8abcfb269521cd0b8e06b0569abcb","reason":"Same source body in authored semantic page"},{"oldId":"b-b50dba0027b1a8e50cbe9fdf","targetId":"b-b50dba0027b1a8e50cbe9fdf","page":"reading-opening","operation":"move","beforeSHA256":"e91f24e4e4ca40e2099a0d81d30db6568f4bf2074b499eeee6e3959617bd97ac","afterSHA256":"e91f24e4e4ca40e2099a0d81d30db6568f4bf2074b499eeee6e3959617bd97ac","reason":"Same source body in authored semantic page"},{"oldId":"b-f175d133c854d6a534f254ca","targetId":"b-f175d133c854d6a534f254ca","page":"reading-opening","operation":"move","beforeSHA256":"6a611aef342a7535362f318ee0c33b63170ce4ab853fa44a05f76a04873fd320","afterSHA256":"6a611aef342a7535362f318ee0c33b63170ce4ab853fa44a05f76a04873fd320","reason":"Same source body in authored semantic page"},{"oldId":"b-2233c3479a3b38c6327111a9","targetId":"b-2233c3479a3b38c6327111a9","page":"reading-opening","operation":"move","beforeSHA256":"f2d01d4cc65a2d6b8086b410e90681528f43d74afafdaecea97b1929b58cc8e9","afterSHA256":"f2d01d4cc65a2d6b8086b410e90681528f43d74afafdaecea97b1929b58cc8e9","reason":"Same source body in authored semantic page"},{"oldId":"b-46a58acc5d434799068fdef4","targetId":"b-46a58acc5d434799068fdef4","page":"reading-opening","operation":"move","beforeSHA256":"afd24af6fb83a24b9e6b1c37d986650cb001aeabdb60796ba074ba78d03f830e","afterSHA256":"afd24af6fb83a24b9e6b1c37d986650cb001aeabdb60796ba074ba78d03f830e","reason":"Same source body in authored semantic page"},{"oldId":"b-b62bb015aacc7495162f9c4b","targetId":"b-b62bb015aacc7495162f9c4b","page":"reading-opening","operation":"move","beforeSHA256":"16f7f71dd346b7d3c2ce1b02f0545c45b258a932b83c6536d6981d6d5af217c7","afterSHA256":"16f7f71dd346b7d3c2ce1b02f0545c45b258a932b83c6536d6981d6d5af217c7","reason":"Same source body in authored semantic page"},{"oldId":"b-9f1779a66a3bd7c069e3820c","targetId":"b-9f1779a66a3bd7c069e3820c","page":"reading-opening","operation":"move","beforeSHA256":"7484037b201c9794c21fc86f5bd33a6f31f9cf5c2c99baf3cd95d5076e2a3da7","afterSHA256":"7484037b201c9794c21fc86f5bd33a6f31f9cf5c2c99baf3cd95d5076e2a3da7","reason":"Same source body in authored semantic page"},{"oldId":"block-fab7a942c9ab4a2ca8dc8d8c5d2877d9--4-推理只接自己的输出再检查视觉依赖","targetId":"b-2233c3479a3b38c6327111a9","operation":"retain-source-identity"}],"blocks":[{"id":"b-f946a4e0ad1d005faf36f974","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/fab7a942c9ab4a2ca8dc8d8c5d2877d9.html#b-f946a4e0ad1d005faf36f974","content":"[LAB 13 · 看图生成一句话：训练、解码与消融](/course/notebooks/3da47f3ee3554ee0a36e0bb6a2fdc61f.html)","sha256":"db08960319942d9119c9cad2cf7d0b78a68c914f8e19ab3bd375531d27c96204","chapter":{"id":"course","title":"4. 推理只接自己的输出，再检查视觉依赖"},"unit":{"id":"notebook-blocks/fab7a942c9ab4a2ca8dc8d8c5d2877d9","title":"4. 推理只接自己的输出，再检查视觉依赖"},"page":{"id":"reading-opening","title":"4. 推理只接自己的输出，再检查视觉依赖","number":1},"sectionTitle":"4. 推理只接自己的输出，再检查视觉依赖","currentQuestion":"4. 推理只接自己的输出，再检查视觉依赖","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-e8820dd5a58ebc7170310586","prerequisites":[],"relatedBlocks":[]},{"id":"b-e8820dd5a58ebc7170310586","type":"h1","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/fab7a942c9ab4a2ca8dc8d8c5d2877d9.html#b-e8820dd5a58ebc7170310586","content":"# 4. 推理只接自己的输出，再检查视觉依赖","sha256":"7d9a64bd2956d9687a40de5ee89d75652fc8abcfb269521cd0b8e06b0569abcb","chapter":{"id":"course","title":"4. 推理只接自己的输出，再检查视觉依赖"},"unit":{"id":"notebook-blocks/fab7a942c9ab4a2ca8dc8d8c5d2877d9","title":"4. 推理只接自己的输出，再检查视觉依赖"},"page":{"id":"reading-opening","title":"4. 推理只接自己的输出，再检查视觉依赖","number":1},"sectionTitle":"4. 推理只接自己的输出，再检查视觉依赖","currentQuestion":"4. 推理只接自己的输出，再检查视觉依赖","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-f946a4e0ad1d005faf36f974","nextBlock":"b-b50dba0027b1a8e50cbe9fdf","prerequisites":[{"id":"b-f946a4e0ad1d005faf36f974","url":"#b-f946a4e0ad1d005faf36f974","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-b50dba0027b1a8e50cbe9fdf","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/fab7a942c9ab4a2ca8dc8d8c5d2877d9.html#b-b50dba0027b1a8e50cbe9fdf","content":"按问题继续","sha256":"e91f24e4e4ca40e2099a0d81d30db6568f4bf2074b499eeee6e3959617bd97ac","chapter":{"id":"course","title":"4. 推理只接自己的输出，再检查视觉依赖"},"unit":{"id":"notebook-blocks/fab7a942c9ab4a2ca8dc8d8c5d2877d9","title":"4. 推理只接自己的输出，再检查视觉依赖"},"page":{"id":"reading-opening","title":"4. 推理只接自己的输出，再检查视觉依赖","number":1},"sectionTitle":"4. 推理只接自己的输出，再检查视觉依赖","currentQuestion":"4. 推理只接自己的输出，再检查视觉依赖","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-e8820dd5a58ebc7170310586","nextBlock":"b-f175d133c854d6a534f254ca","prerequisites":[{"id":"b-e8820dd5a58ebc7170310586","url":"#b-e8820dd5a58ebc7170310586","role":"section-title"},{"id":"b-f946a4e0ad1d005faf36f974","url":"#b-f946a4e0ad1d005faf36f974","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-f175d133c854d6a534f254ca","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/fab7a942c9ab4a2ca8dc8d8c5d2877d9.html#b-f175d133c854d6a534f254ca","content":"真实CNN–GRU条件训练与greedy/beam生成、遮图/错图消融","sha256":"6a611aef342a7535362f318ee0c33b63170ce4ab853fa44a05f76a04873fd320","chapter":{"id":"course","title":"4. 推理只接自己的输出，再检查视觉依赖"},"unit":{"id":"notebook-blocks/fab7a942c9ab4a2ca8dc8d8c5d2877d9","title":"4. 推理只接自己的输出，再检查视觉依赖"},"page":{"id":"reading-opening","title":"4. 推理只接自己的输出，再检查视觉依赖","number":1},"sectionTitle":"4. 推理只接自己的输出，再检查视觉依赖","currentQuestion":"4. 推理只接自己的输出，再检查视觉依赖","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-b50dba0027b1a8e50cbe9fdf","nextBlock":"b-2233c3479a3b38c6327111a9","prerequisites":[{"id":"b-e8820dd5a58ebc7170310586","url":"#b-e8820dd5a58ebc7170310586","role":"section-title"},{"id":"b-f946a4e0ad1d005faf36f974","url":"#b-f946a4e0ad1d005faf36f974","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-2233c3479a3b38c6327111a9","type":"h2","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/fab7a942c9ab4a2ca8dc8d8c5d2877d9.html#b-2233c3479a3b38c6327111a9","content":"## 4. 推理只接自己的输出，再检查视觉依赖","sha256":"f2d01d4cc65a2d6b8086b410e90681528f43d74afafdaecea97b1929b58cc8e9","chapter":{"id":"course","title":"4. 推理只接自己的输出，再检查视觉依赖"},"unit":{"id":"notebook-blocks/fab7a942c9ab4a2ca8dc8d8c5d2877d9","title":"4. 推理只接自己的输出，再检查视觉依赖"},"page":{"id":"reading-opening","title":"4. 推理只接自己的输出，再检查视觉依赖","number":1},"sectionTitle":"4. 推理只接自己的输出，再检查视觉依赖","currentQuestion":"4. 推理只接自己的输出，再检查视觉依赖","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-f175d133c854d6a534f254ca","nextBlock":"b-46a58acc5d434799068fdef4","prerequisites":[{"id":"b-e8820dd5a58ebc7170310586","url":"#b-e8820dd5a58ebc7170310586","role":"section-title"},{"id":"b-f946a4e0ad1d005faf36f974","url":"#b-f946a4e0ad1d005faf36f974","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-46a58acc5d434799068fdef4","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/fab7a942c9ab4a2ca8dc8d8c5d2877d9.html#b-46a58acc5d434799068fdef4","content":"Greedy 每一步取当前分布的 argmax，追加到自己的前缀，遇到 EOS 停止。下面的 beam 真实展开每条未结束假设，累加 log probability，并只保留分数最高的 3 条；结束假设不再扩展，最长生成 8 个新 token。此处没有长度归一化，所以不同长度会受原始句子概率的长度偏好影响。保留数据已足够简单，两种搜索都能答对，并未显示 beam 胜出。","sha256":"afd24af6fb83a24b9e6b1c37d986650cb001aeabdb60796ba074ba78d03f830e","chapter":{"id":"course","title":"4. 推理只接自己的输出，再检查视觉依赖"},"unit":{"id":"notebook-blocks/fab7a942c9ab4a2ca8dc8d8c5d2877d9","title":"4. 推理只接自己的输出，再检查视觉依赖"},"page":{"id":"reading-opening","title":"4. 推理只接自己的输出，再检查视觉依赖","number":1},"sectionTitle":"4. 推理只接自己的输出，再检查视觉依赖","currentQuestion":"4. 推理只接自己的输出，再检查视觉依赖","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-2233c3479a3b38c6327111a9","nextBlock":"b-b62bb015aacc7495162f9c4b","prerequisites":[{"id":"b-2233c3479a3b38c6327111a9","url":"#b-2233c3479a3b38c6327111a9","role":"section-title"},{"id":"b-f946a4e0ad1d005faf36f974","url":"#b-f946a4e0ad1d005faf36f974","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-b62bb015aacc7495162f9c4b","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/fab7a942c9ab4a2ca8dc8d8c5d2877d9.html#b-b62bb015aacc7495162f9c4b","content":"本册的整句精确匹配比较解码后的内容字符串：颜色、形状和位置需与规范句一致。`words()`会忽略BOS/PAD并在EOS处截断，因此这个分数不单独检查EOS出现的具体位置；停止行为应另看生成token序列。它适合唯一规范句子的合成任务。自然图像可有多种正确描述，不能照搬此指标。视觉置零是分布外干预，单独下降还不够；因此再做固定错配：把保留集图像循环移位一个位置，保持真值句子不变。这个排列无固定点，但不等于均匀随机抽样，也不能当作“随机猜测基线”。","sha256":"16f7f71dd346b7d3c2ce1b02f0545c45b258a932b83c6536d6981d6d5af217c7","chapter":{"id":"course","title":"4. 推理只接自己的输出，再检查视觉依赖"},"unit":{"id":"notebook-blocks/fab7a942c9ab4a2ca8dc8d8c5d2877d9","title":"4. 推理只接自己的输出，再检查视觉依赖"},"page":{"id":"reading-opening","title":"4. 推理只接自己的输出，再检查视觉依赖","number":1},"sectionTitle":"4. 推理只接自己的输出，再检查视觉依赖","currentQuestion":"4. 推理只接自己的输出，再检查视觉依赖","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-46a58acc5d434799068fdef4","nextBlock":"b-9f1779a66a3bd7c069e3820c","prerequisites":[{"id":"b-2233c3479a3b38c6327111a9","url":"#b-2233c3479a3b38c6327111a9","role":"section-title"},{"id":"b-f946a4e0ad1d005faf36f974","url":"#b-f946a4e0ad1d005faf36f974","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-9f1779a66a3bd7c069e3820c","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/fab7a942c9ab4a2ca8dc8d8c5d2877d9.html#b-9f1779a66a3bd7c069e3820c","content":"先预测：固定 `<bos> a`，红方块左侧换成蓝圆右侧，哪个词最先应改变？随后检查 `same_prefix_controls`，不要把数据生成器的标签当成模型生成。","sha256":"7484037b201c9794c21fc86f5bd33a6f31f9cf5c2c99baf3cd95d5076e2a3da7","chapter":{"id":"course","title":"4. 推理只接自己的输出，再检查视觉依赖"},"unit":{"id":"notebook-blocks/fab7a942c9ab4a2ca8dc8d8c5d2877d9","title":"4. 推理只接自己的输出，再检查视觉依赖"},"page":{"id":"reading-opening","title":"4. 推理只接自己的输出，再检查视觉依赖","number":1},"sectionTitle":"4. 推理只接自己的输出，再检查视觉依赖","currentQuestion":"4. 推理只接自己的输出，再检查视觉依赖","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-b62bb015aacc7495162f9c4b","nextBlock":null,"prerequisites":[{"id":"b-2233c3479a3b38c6327111a9","url":"#b-2233c3479a3b38c6327111a9","role":"section-title"},{"id":"b-f946a4e0ad1d005faf36f974","url":"#b-f946a4e0ad1d005faf36f974","role":"problem-context"}],"relatedBlocks":[]}]}