知识卡片
分类头组合空间与通道
简单CNN分类器先卷积提取C路响应,经ReLU、全局平均池化变为C维向量,再用K×C仿射层和类别Softmax输出K类分布。全连接描述连接结构,不要求权重矩阵满秩;GAP是均值而非未归一化求和。
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
- · 卷积神经网络
原书出处 · 1 处
- · 查看原书依据
This network has one convolutional layer with C channels followed by a relu layer. The next layer performs spatial global average pooling (gap), and each channel gets projected into a single number that contains the sum of the outputs of the relu. This results in a representation given by a vector of length C. This vector is then processed by a fully connected layer (fc).
参考来源
- 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第24章《卷积神经网络》(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/c024b.xhtml)
- 结论依据:摘录中关于分类头组合空间与通道的定义、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:This network has one convolutional layer with C channels followed by a relu layer. The next layer performs spatial global average pooling (gap), and each channel gets projected into a single number that contains the sum of the outputs of the relu. This results in a representation given by a vector of length C. This vector is then processed by a fully connected layer (fc).