知识卡片
CLIP用双向配对损失对齐
CLIP分别编码图像和文本,构建批内相似度矩阵,对图到文和文到图各自按行或列做交叉熵,再平均。配对样本为正,其他批内组合通常作为负例;这是相对匹配目标,不是直接生成描述或逐像素标注。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
- · 视觉与语言
原书出处 · 1 处
- · 查看原书依据
where ℓ represents an image, t represents text, represents the space of unit vectors of dimensionality dz (i.e., the surface of the [dz – 1]-dimensional hypersphere), fℓ is the image encoder, ft is the text encoder, image inputs are represented as [3 × N × M] pixel arrays, text inputs are represented as dt dimensional tokens, and dz is the embedding dimensionality. Notice that the
参考来源
- 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第51章《视觉与语言》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)/_epub-src/OPS/xhtml/c051.xhtml)
- 结论依据:摘录中关于CLIP用双向配对损失对齐的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。
- 原始内容:where ℓ represents an image, t represents text, represents the space of unit vectors of dimensionality dz (i.e., the surface of the [dz – 1]-dimensional hypersphere), fℓ is the image encoder, ft is the text encoder, image inputs are represented as [3 × N × M] pixel arrays, text inputs are represented as dt dimensional tokens, and dz is the embedding dimensionality. Notice that the