← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8934 · 生成模型:GPT-6

逐像素预测移除全局汇聚

理解等级:中级

让网络输出空间类别图,需要保留适当分辨率并移除把空间压成向量的全局池化;最后按每个像素的类别通道做Softmax。每个位置的类别分布独立归一化,不应沿整张空间图做一次总归一化。上采样可还原网格,但不能凭空找回编码中丢掉的细节。

这张卡片看不懂?

已有反馈 · 0 条

    还没有人提交反馈。

    在整书笔记中的位置

    原书出处 · 1 处

    • · 查看原书依据

      In section 24.4 we saw a CNN that outputs a single class probability vector for an image. What if we want to output a spatially varying map of predictions, like we discussed in the intro to this chapter? To achieve this, we can simply downsample less, so that the final layer of the CNN is a feature map that maintains higher spatial resolution. It is

    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:摘录中关于逐像素预测移除全局汇聚的定义、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:In section 24.4 we saw a CNN that outputs a single class probability vector for an image. What if we want to output a spatially varying map of predictions, like we discussed in the intro to this chapter? To achieve this, we can simply downsample less, so that the final layer of the CNN is a feature map that maintains higher spatial resolution. It is