如何在NumPy中为算术运算指定结果数据类型？

似乎从int数组中减去float32数字会导致默认情况下导致float64数组.有没有办法绕过这个并获得float32作为结果的数据类型？

numpy.subtract不允许指定dtype参数.

实现这一点的唯一方法是将int数组转换为float32之前,实际上减去了两个似乎相当慢的float32数组.这是应该的方式吗？

示例代码:

import time
import numpy as np

if __name__ == '__main__':

    # some int32 array
    a = np.arange(1e7)
    print('a.dtype={}'.format(a.dtype)) # int32

    # subtraction with a python float
    t0 = time.clock()
    b = a - 5.5
    t1 = time.clock()
    print('b.dtype={}, took {}s'.format(b.dtype, t1 - t0)) # float64

    # a numpy float32
    c = np.array(5.5, dtype=np.float32)
    print('c.dtype={}'.format(c.dtype))  # float32

    # subtraction with the numpy float32
    t0 = time.clock()
    d = a - c
    t1 = time.clock()
    print('d.dtype={}, took {}s'.format(d.dtype, t1 - t0)) # float64! why not float32

    # convert the int32 to float32
    e = a.astype(dtype=np.float32)
    print('e.dtype={}'.format(e.dtype)) # float32

    # subtract two float32 array
    t0 = time.clock()
    e = a.astype(dtype=np.float32)
    f = e - c
    t1 = time.clock()
    print('f.dtype={}, took {}s'.format(f.dtype, t1 - t0)) # float32 (finally)

版画

a.dtype=float64
b.dtype=float64, took 0.0229595559008s
c.dtype=float32
d.dtype=float64, took 0.0223958136306s
e.dtype=float32
f.dtype=float32, took 0.0334388477586s

手动转换为float32之前似乎比自动转换为float64慢.

作为ufunc,np.subtract接受几个未在其主文档页面上列出的关键字参数.其中一个是dtype,所以你可以这么说,np.subtract(a, b, dtype='float32')并且:实现你想要的结果.

如果它在将来有用,np.promote_types 会告诉你两个dtypes都可以安全地投射的最小dtype,并np.result_type告诉你dtype Numpy的投射规则默认会产生.