Tương đương với np.std() trong TensorFlow là gì?

Question

Chỉ cần tìm kiếm giá trị tương đương của np.std() trong TensorFlow để tính toán độ lệch chuẩn của một tensor.

Answer 1

Để lấy giá trị trung bình và phương sai chỉ cần sử dụng tf.nn.moments.

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

Để biết thêm về tf.nn.moments params thấy docs

Answer 2

Bạn cũng có thể sử dụng reduce_std trong đoạn mã sau chuyển thể từ Keras:

#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))