Python – tensorflow.GradientTape.batch_jacobian()

TensorFlow 是由 Google 设计的开源Python库，用于开发机器学习模型和深度学习神经网络。

batch_jacobian()用于计算和堆叠每个示例 jacobian。

Syntax: batch_jacobian( target, source, unconnected_gradients, parallel_iterations, experimental_use_pfor )

Parameters:

• target: It is a Tensor having minimum rank 2.
• source: It is a Tensor having minimum rank 2.
• unconnected_gradients(optional): It’s value can either be zero or None. Default value is None.
• parallel_iterations(optional): It is used to control parallel iterations and memory usage.
• experimental_use_pfor(optional): It is a boolean with default value True. It uses pfor to calculate jacobian when set to true otherwise tf.while_loop is used.

Returns: It returns a Tensor. 

示例 1：

Python3

# Importing the library
import tensorflow as tf
  
x = tf.constant([[4, 2],[1, 3]], dtype=tf.dtypes.float32)
  
# Using GradientTape
with tf.GradientTape() as gfg:
  gfg.watch(x)
  y = x * x * x
  
# Computing jacobian
res  = gfg.batch_jacobian(y, x) 
  
# Printing result
print("res: ",res)

输出：

res:  tf.Tensor(
[[[48.  0.]
  [ 0. 12.]]

 [[ 3.  0.]
  [ 0. 27.]]], shape=(2, 2, 2), dtype=float32)

示例 2：

Python3

# Importing the library
import tensorflow as tf
  
x = tf.constant([[4, 2],[1, 3]], dtype=tf.dtypes.float32)
  
# Using GradientTape
with tf.GradientTape() as gfg:
  gfg.watch(x)
  
  # Using nested GradientTape for calculating higher order jacobian
  with tf.GradientTape() as gg:
    gg.watch(x)
    y = x * x * x
  # Computing first order jacobian
  first_order = gg.batch_jacobian(y, x)
  
# Computing Second order jacobian
second_order  = gfg.batch_jacobian(first_order, x) 
  
# Printing result
print("first_order: ",first_order)
print("second_order: ",second_order)

输出：

first_order:  tf.Tensor(
[[[48.  0.]
  [ 0. 12.]]

 [[ 3.  0.]
  [ 0. 27.]]], shape=(2, 2, 2), dtype=float32)
second_order:  tf.Tensor(
[[[[24.  0.]
   [ 0.  0.]]

  [[ 0.  0.]
   [ 0. 12.]]]


 [[[ 6.  0.]
   [ 0.  0.]]

  [[ 0.  0.]
   [ 0. 18.]]]], shape=(2, 2, 2, 2), dtype=float32)