TensorFlow中的np.std()相当于什么?

时间:2016-09-06 17:19:50

标签: python tensorflow

只需在TensorFlow中查找等效的np.std()来计算张量的标准偏差。

2 个答案:

答案 0 :(得分:39)

要获得均值和方差,只需使用tf.nn.moments

mean, var = tf.nn.moments(x, axes=[1])

有关tf.nn.moments参数的详情,请参阅docs

答案 1 :(得分:6)

您还可以在以下由Keras改编的代码中使用reduce_std

#coding=utf-8
import numpy as np
import tensorflow as tf

def reduce_var(x, axis=None, keepdims=False):
    """Variance of a tensor, alongside the specified axis.

    # Arguments
        x: A tensor or variable.
        axis: An integer, the axis to compute the variance.
        keepdims: A boolean, whether to keep the dimensions or not.
            If `keepdims` is `False`, the rank of the tensor is reduced
            by 1. If `keepdims` is `True`,
            the reduced dimension is retained with length 1.

    # Returns
        A tensor with the variance of elements of `x`.
    """
    m = tf.reduce_mean(x, axis=axis, keep_dims=True)
    devs_squared = tf.square(x - m)
    return tf.reduce_mean(devs_squared, axis=axis, keep_dims=keepdims)

def reduce_std(x, axis=None, keepdims=False):
    """Standard deviation of a tensor, alongside the specified axis.

    # Arguments
        x: A tensor or variable.
        axis: An integer, the axis to compute the standard deviation.
        keepdims: A boolean, whether to keep the dimensions or not.
            If `keepdims` is `False`, the rank of the tensor is reduced
            by 1. If `keepdims` is `True`,
            the reduced dimension is retained with length 1.

    # Returns
        A tensor with the standard deviation of elements of `x`.
    """
    return tf.sqrt(reduce_var(x, axis=axis, keepdims=keepdims))

if __name__ == '__main__':
    x_np = np.arange(10).reshape(2, 5).astype(np.float32)
    x_tf = tf.constant(x_np)
    with tf.Session() as sess:
        print(sess.run(reduce_std(x_tf, keepdims=True)))
        print(sess.run(reduce_std(x_tf, axis=0, keepdims=True)))
        print(sess.run(reduce_std(x_tf, axis=1, keepdims=True)))
    print(np.std(x_np, keepdims=True))
    print(np.std(x_np, axis=0, keepdims=True))
    print(np.std(x_np, axis=1, keepdims=True))