tensorflow reduce_mean vs numpy mean

Question

据我了解，tensorflow reduce_mean 和 numpy mean 应该返回相同的值，但下面的示例返回不同的值：

import numpy as np
import tensorflow as tf

t_1 = tf.constant([1,3,4,5])
t_2 = tf.constant([7,8,9,0])
list_t = [t_1, t_2]
reduced_t_list = tf.reduce_mean(list_t)
sess= tf.Session()
print(sess.run(reduced_t_list))
print(np.mean([1,3,4,5,7,8,9,0]))

output:
4
4.625

你猜为什么？

Answer 1

来自tf.constant docs：

If the argument dtype is not specified, then the type is inferred from the type of value.

[1, 2, 3, 4] 的dtype 是int，而np.mean([1, 2, 3]) 默认将其转换为floats 的数组。

试试tf.constant(np.arange(3.0))。