【Python】Numpy使用reshape和newaxis改变形状

问题描述

numpy中np.reshape()和np.newaxis都能增加array的维度，两者有何区别？

import numpy as np

a = np.ones(12)
b = a.reshape(-1, 1)
c = a.reshape(-1, 1, 1)
d = a[:, np.newaxis]
e = a[:, np.newaxis, np.newaxis]

print(a.shape)  # (12,)
print(b.shape)  # (12, 1)
print(c.shape)  # (12, 1, 1)
print(d.shape)  # (12, 1)
print(e.shape)  # (12, 1, 1)

---

异同

两者没有特别大的区别。两者都创建一个view，使用与原始data使用相同的缓冲区。

print(a.__array_interface__)
print(b.__array_interface__)
print(d.__array_interface__)

"""
{'data': (2012817932320, False), 'strides': None, 'descr': [('', '<f8')], 'typestr': '<f8', 'shape': (12,), 'version': 3}
{'data': (2012817932320, False), 'strides': None, 'descr': [('', '<f8')], 'typestr': '<f8', 'shape': (12, 1), 'version': 3}
{'data': (2012817932320, False), 'strides': None, 'descr': [('', '<f8')], 'typestr': '<f8', 'shape': (12, 1), 'version': 3}
"""

但是在使用时有一个小区别：当要操作多个未知维度时，使用numpy.newaxis而不是ndarray.reshape()。
例如：

import numpy as np

a = np.ones((3, 3))
b = a[:, np.newaxis, :]  
print(b.shape)  # (3, 1, 3)

c = a.reshape(-1, 1, -1)  # ValueError: can only specify one unknown dimension