- html - 出于某种原因,IE8 对我的 Sass 文件中继承的 html5 CSS 不友好?
- JMeter 在响应断言中使用 span 标签的问题
- html - 在 :hover and :active? 上具有不同效果的 CSS 动画
- html - 相对于居中的 html 内容固定的 CSS 重复背景?
请帮助理解将元组 (i, j) 编入 ndarray 的 Numpy 索引的设计决策或合理性。
当索引是单个元组(4, 2)时,则(i=row, j=column)。
shape = (6, 7)
X = np.zeros(shape, dtype=int)
X[(4, 2)] = 1
X[(5, 3)] = 1
print("X is :\n{}\n".format(X))
---
X is :
[[0 0 0 0 0 0 0]
[0 0 0 0 0 0 0]
[0 0 0 0 0 0 0]
[0 0 0 0 0 0 0]
[0 0 1 0 0 0 0] <--- (4, 2)
[0 0 0 1 0 0 0]] <--- (5, 3)
但是,当索引是多个元组 (4, 2), (5, 3) 时,则 (i=row, j=row) for (4, 2) and (i=column, j=column) for (5, 3).
shape = (6, 7)
Y = np.zeros(shape, dtype=int)
Y[(4, 2), (5, 3)] = 1
print("Y is :\n{}\n".format(Y))
---
Y is :
[[0 0 0 0 0 0 0]
[0 0 0 0 0 0 0]
[0 0 0 1 0 0 0] <--- (2, 3)
[0 0 0 0 0 0 0]
[0 0 0 0 0 1 0] <--- (4, 5)
[0 0 0 0 0 0 0]]
It means you are constructing a 2d array R, such that
R=A[B, C]
.This means that the value forrij=abijcij.So it means that the item located at
R[0,0]
is the item inA
withas row indexB[0,0]
and as column indexC[0,0]
. The itemR[0,1]
is the item inA
with row indexB[0,1]
and as column indexC[0,1]
, etc.
multi_index: A tuple of integer arrays, one array for each dimension.
为什么不总是(i=行,j=列)?如果一直是(i=row, j=column)会怎么样?
根据 Akshay 和@DaniMesejo 的回答,理解:
X[
(4), # dimension 2 indices with only 1 element
(2) # dimension 1 indices with only 1 element
] = 1
Y[
(4, 2, ...), # dimension 2 indices
(5, 3, ...) # dimension 1 indices (dimension 0 is e.g. np.array(3) whose shape is (), in my understanding)
] = 1
最佳答案
很容易理解它是如何工作的(以及这个设计决策背后的动机)。
Numpy 将其 ndarray 存储为连续的内存块。每个元素在前一个元素之后每隔 n 个字节按顺序存储。
(引用自 excellent SO post 的图片)
所以如果你的 3D 数组看起来像这样 -
然后在内存中将其存储为 -
当检索一个元素(或一个元素 block )时,NumPy 计算有多少 strides
(bytes) 需要遍历得到下一个元素in that direction/axis
.所以,对于上面的例子,axis=2
它必须遍历 8 个字节(取决于 datatype
)但对于 axis=1
它必须遍历 8*4
字节,和 axis=0
它需要8*8
字节。
考虑到这一点,让我们看看您要尝试做什么。
print(X)
print(X.strides)
[[0 0 0 0 0 0 0]
[0 0 0 0 0 0 0]
[0 0 0 0 0 0 0]
[0 0 0 0 0 0 0]
[0 0 1 0 0 0 0]
[0 0 0 1 0 0 0]]
#Strides (bytes) required to traverse in each axis.
(56, 8)
对于您的数组,获取 axis=0
中的下一个元素,我们需要遍历56 bytes
,以及 axis=1
中的下一个元素, 我们需要 8 bytes
.
当您索引 (4,2)
时, NumPy 正在运行 56*4
axis=0
中的字节数和 8*2
axis=1
中的字节数访问它。同样,如果你想访问 (4,2)
和 (5,3)
,它将必须访问 56*(4,5)
在axis=0
和 8*(2,3)
在axis=1
.
这就是设计之所以如此的原因,因为它与 NumPy 实际上使用 strides
索引元素的方式一致。 .
X[(axis0_indices), (axis1_indices), ..]
X[(4, 5), (2, 3)] #(row indices), (column indices)
array([1, 1])
通过这种设计,也可以轻松扩展到更高维度的张量(例如 8 维数组)! 如果您分别提及每个索引元组,则需要元素 * 计算的维数才能获取这些元组。使用这种设计时,它可以将步幅值广播到每个轴的元组并更快地获取这些值!
关于python - numpy索引解释 ndarray[(4, 2), (5, 3)],我们在Stack Overflow上找到一个类似的问题: https://stackoverflow.com/questions/65389298/
我在一个项目中工作,该项目需要 SQL 结果的最佳性能,并且希望优化查询,但经过反复试验后,我在 IN 方面遇到了一些问题。 -- THIS RETURNS NO RESULTS AT ALL. SE
在尝试创建一个实际上非常简单的 SQL 语句时,我发现自己迷失了方向。 我有一个包含 3 个表的数据库: 食谱 - 存储一些用于 cooking 的食谱名称 配料食谱 - 将配料与食谱链接 成分 -
我正在尝试理解 PHP 中的 Hebrev 函数。 https://php.net/manual/en/function.hebrevc.php 它说:“将逻辑希伯来语文本转换为视觉文本”。但我不明白
嗨,我在 Grid view 的 android 文档中发现了一段代码对于以下代码。 gridview.setOnItemClickListener(new OnItemClickListener()
谁能解释一下 InfiniBand 是什么?与以太网相比的主要区别是什么,这些差异如何使其比以太网更快? 在官方description从 mellanox 写到 Introduce InfiniBan
这个问题已经有答案了: How are java increment statements evaluated in complex expressions (1 个回答) 已关闭 8 年前。 我知道
我正在阅读 MySQL 教程,我遇到了这个: SELECT /*! SQL_NO_CACHE */ user FROM users; 为什么优化提示 SQL_NO_CACHE 包含在: /*!
我无法理解$(this),我做了一个剪刀石头布的版本,并应用了 jQuery 让用户在计算机上选择按钮选项。我希望有人能解释一下 $(this) 指的是什么,它是 btn-primary 吗?该函数在
我不是很确定 while(choice == 1 || choice ==2);谁能解释一下。我明白这一点 if(choice ==1) displayMonthly(rainfall); e
let flyRight = CABasicAnimation(keyPath: "position.x") flyRight.toValue = view.bounds.size.width/2 f
目录 解释:int型默认值为0 但我们尝试发现并不能通过: 原因: int的默认值为0,而Integer的默认值为null
我正在处理一个查询,自从一个 SSRS 服务器传输到另一个服务器后,它似乎没有按预期执行,并且 where 语句的一部分中出现了以下行 找出不同之处,或者至少从我能找到的地方来看。 where COA
我正在制作一个退回检测程序,读取退回邮件。我们的设置是发送电子邮件,在发送的邮件中添加一个 noreply@domain.tl。一些收件人不再存在,因此我们想要读取退回邮件,并检测它发送给谁。我已经崩
我有一个关于公式通过控制点弯曲的问题。 如您所知,HTML Canvas 有 quadraticCurveTo(x1, y1, x2, y2)与 x1 and x2作为控制点。 但是,当您尝试使用它绘
我有一个 Emakefile看起来像: %% -- %% %% -- {'/Users/user/projects/custom_test/trunk/*', [debug_info, {out
我有一个非常简单的问题。这不仅适用于 spray-json,而且我已经阅读了 argonaut 和 circe 的类似声明。所以请赐教。 在 spray-json 中,我遇到了 There is no
我正在为视频添加水印。我试图让水印与视频尺寸成比例。我已经使用 scale2ref 看到了十几个不同的答案,但没有解释实际发生了什么,所以我发现很难知道如何实现/更改配置以适应我的情况。 当前覆盖命令
因为我正在学习语言,所以我在玩 Haskell,我只是发现了一些我不理解的东西,我找不到解释。如果我尝试运行此代码: map (`div` 0) [1,2,3,4] 我得到一个除以 0 的异常,这是预
我正在寻找解决错误对象引用未设置到对象实例的步骤/指南。以及问题发生原因的解释。 我正在寻找更一般的解释,所以如果我收到错误,我应该采取什么步骤来查找问题。我经常看到有人提供特定代码段的帖子,而其他人
我最近想升级我的知识React ,所以我从组件生命周期方法开始。让我好奇的第一件事是这个componentWillReceiveProps .所以,文档说当组件接收新的(不一定是更新的) Prop 时
我是一名优秀的程序员,十分优秀!