NumPy API详解大全（持续更新ing...）

诸神缄默不语-个人优快云博文目录

官方文档：NumPy Reference — NumPy v1.23 Manual

1. Array Objects

1.1 The N-dimensional array ( ndarray )

1.1.1 numpy.ndarray

1. astype(dtype)：复制array对象，并重置为指定type
   实例代码：array2=array1.astype(np.float32)
   https://numpy.org/doc/stable/reference/generated/numpy.ndarray.astype.html
2. shape：由array每一维度组成的tuple
   如果想要改变array维度，可以直接改变shape，但是不建议这么干
   
   https://numpy.org/doc/stable/reference/generated/numpy.ndarray.shape.html
3. mean() 等如numpy.mean()

2. Constants

3. Universal Functions (ufunc)

4. Routines

4.1 Array creation routines

1. array(object)：创建一个值与object相同的numpy.ndarray对象（元素保持最高精度）
   
2. numpy.diag(v, k=0)
   1. 如v是一维向量：构建以v为值的对角阵
   2. k控制在对角线上方几行
      https://numpy.org/doc/stable/reference/generated/numpy.diag.html

4.2 Array manipulation routines

1. numpy.expand_dims(a, axis)：增加一维（感觉上比较类似PyTorch的unsqueeze()）https://numpy.org/doc/stable/reference/generated/numpy.expand_dims.html
2. numpy.concatenate((a1, a2, ...), axis=0, out=None, dtype=None, casting="same_kind")：合并多个ndarray
   https://numpy.org/doc/stable/reference/generated/numpy.concatenate.html
   常见问题：入参ndarray必须用稠密矩阵，如果使用了稀疏矩阵，将会报ValueError: zero-dimensional arrays cannot be concatenated bug
3. numpy.transpose(a, axes=None)
   https://numpy.org/doc/stable/reference/generated/numpy.transpose.html
   1. 对二维矩阵：转置
4. numpy.reshape(a, /, shape=None, *, newshape=None, order='C', copy=None)：改变张量的形状，但不改变数据。

4.3 Binary operations

4.4 String operations

4.5 C-Types Foreign Function Interface (numpy.ctypeslib)

4.6 Datetime Support Functions

4.7 Data type routines

1. class numpy.finfo(dtype)：机器对浮点类型的各项限制

4.8 Optionally SciPy-accelerated routines (numpy.dual)

4.9 Mathematical functions

1. numpy.power(x1, x2, /, out=None, *, where=True, casting='same_kind', order='K', dtype=None, subok=True[, signature, extobj])：element-wise计算幂
2. numpy.sum(a, axis=None, dtype=None, out=None, keepdims=<no value>, initial=<no value>, where=<no value>)：当axis=None时计算张量所有值的和

4.10 Floating point error handling

1. numpy.errstate(**kwargs)：上下文管理器或者装饰器
   kwargs{divide, over, under, invalid}
   https://numpy.org/doc/stable/reference/generated/numpy.errstate.html
   
   示例一：

with np.errstate(divide='ignore'):
    d_inv_sqrt = np.power(rowsum, -0.5)

4.11 Discrete Fourier Transform (numpy.fft)

4.12 Functional programming

4.13 NumPy-specific help functions

4.14 Indexing routines

4.15 Input and output

1. ndarray.tolist()：将ndarray对象转为nested list，元素为Python标量
   https://numpy.org/doc/stable/reference/generated/numpy.ndarray.tolist.html
   如果直接用list(ndarray)会导致元素是numpy类型的数字

4.16 Linear algebra (numpy.linalg)

1. numpy.dot(a, b, out=None)
   1. 如果a和b都是二维矩阵：返回a和b的矩阵乘法结果

4.17 Logic functions

1. numpy.allclose(a, b, rtol=1e-05, atol=1e-08, equal_nan=False)：返回2个array是否所有元素都相同（允许一定误差）
   可用于检测一个矩阵是否是对称阵：np.allclose(X, X.T)
2. numpy.isinf(x, /, out=None, *, where=True, casting='same_kind', order='K', dtype=None, subok=True[, signature, extobj])：element-wise计算矩阵元素是否无限（可能是正负无限）

4.18 Masked array operations

4.19 Mathematical functions

4.20 Matrix library (numpy.matlib)

4.21 Miscellaneous routines

4.22 Padding Arrays

4.23 Polynomials

4.24 Random sampling (numpy.random)

https://numpy.org/doc/stable/reference/random/index.html

1. Generator
   https://numpy.org/doc/stable/reference/random/generator.html
   Generator是代替RandomState的。
   1. numpy.random.default_rng(seed=None)
      参数：seed{None, int, array_like[ints], SeedSequence, BitGenerator, Generator}, optional
      示例代码：

   import numpy as np
   rng = np.random.default_rng(12345)
   
   rfloat = rng.random()
   #生成一个随机float
   
   rints = rng.integers(low=0, high=10, size=3)
   #生成0（包含）-10（不包含）之间的3个integer
   

2. Generator.choice(a, size=None, replace=True, p=None, axis=0, shuffle=True)
   https://numpy.org/doc/stable/reference/random/generated/numpy.random.Generator.choice.html
   从传入的一维数组（a）中产生随机抽样
   参数：
   size {int, tuple[int]}, optional：输出的shape
   replace bool, optional：有放回抽样 / 无放回抽样
   示例代码：
   ①从np.arange(5)中生成一个大小为3的抽样

   rng.choice(5, 3)
   #输出：array([0, 3, 4]) # random
   #This is equivalent to rng.integers(0,5,3)
   

3. random.random(size=None)
   返回 [0.0, 1.0) 范围内的一个随机浮点数
4. random.rand(d0, d1, ..., dn)
   返回给定尺寸的随机数ndarray

4.24 Set routines

1. numpy.unique(arr)：直接的默认返回值是np.ndarray对象，arr中唯一值从小到大排序
   https://numpy.org/doc/stable/reference/generated/numpy.unique.html

4.25 Sorting, searching, and counting

1. numpy.where(condition, [x, y, ]/)：condition是一个张量，返回的张量用x覆盖condition中值为True的元素，y覆盖值为False的
   例子1，直接对每个元素进行替换：

   import numpy as np
   a = np.arange(10)
   a
   # 输出：array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
   np.where(a < 5, a, 10*a)
   # 输出：array([ 0,  1,  2,  3,  4, 50, 60, 70, 80, 90])
   

   例子2，用整个张量去覆盖（就相当于直接挨个儿对应元素）：

   np.where([[True, False], [True, True]],
            [[1, 2], [3, 4]],
            [[9, 8], [7, 6]])
   array([[1, 8],
          [3, 4]])
   

4.26 Statistics

1. numpy.mean(a)：直接求整个array-like对象的平均值
   https://numpy.org/doc/stable/reference/generated/numpy.mean.html

5. 其他功能的实现

1. 切片：
   按标准来将ndarray变成布尔张量：example_matrix[example_matrix>=0.98]

其他正文及脚注未提及的参考资料

1. numpy和tensor增加、删除一个维度 - 简书
2. Numpy矩阵操作总结 | zdaiot