tf.transpose (API r1.3)

tf.transpose (API r1.3)

1. tf.transpose

transpose(    a,    perm=None,    name='transpose')

Defined in tensorflow/python/ops/array_ops.py.
See the guides: Math > Matrix Math Functions, Tensor Transformations > Slicing and Joining

Transposes a. Permutes the dimensions according to perm.

The returned tensor's dimension i will correspond to the input dimension perm[i]. If perm is not given, it is set to (n-1...0), where n is the rank of the input tensor. Hence by default, this operation performs a regular matrix transpose on 2-D input Tensors.

For example:

x = tf.constant([[1, 2, 3], [4, 5, 6]])tf.transpose(x)  # [[1, 4]                 #  [2, 5]                 #  [3, 6]]# Equivalentlytf.transpose(x, perm=[1, 0])  # [[1, 4]                              #  [2, 5]                              #  [3, 6]]# 'perm' is more useful for n-dimensional tensors, for n > 2x = tf.constant([[[ 1,  2,  3],                  [ 4,  5,  6]],                 [[ 7,  8,  9],                  [10, 11, 12]]])# Take the transpose of the matrices in dimension-0tf.transpose(x, perm=[0, 2, 1])  # [[[1,  4],                                 #   [2,  5],                                 #   [3,  6]],                                 #  [[7, 10],                                 #   [8, 11],                                 #   [9, 12]]]

Args:
a: A Tensor.
perm: A permutation of the dimensions of a.
name: A name for the operation (optional).

Returns:

A transposed Tensor.

2. example 1

import tensorflow as tfimport numpy as npx1 = tf.constant([[1, 2, 3], [4, 5, 6]])y11 = tf.transpose(x1)  # [[1, 4]                        #  [2, 5]                        #  [3, 6]]# Equivalentlyy12 = tf.transpose(x1, perm=[1, 0])  # [[1, 4]                                     #  [2, 5]                                     #  [3, 6]]# 'perm' is more useful for n-dimensional tensors, for n > 2x2 = tf.constant([[[ 1,  2,  3],                   [ 4,  5,  6]],                  [[ 7,  8,  9],                   [10, 11, 12]]])# Take the transpose of the matrices in dimension-0y021 = tf.transpose(x2, perm=[0, 2, 1])  # [[[1,  4],                                         #   [2,  5],                                         #   [3,  6]],                                         #  [[7, 10],                                         #   [8, 11],                                         #   [9, 12]]]y012 = tf.transpose(x2, perm=[0, 1, 2])y102 = tf.transpose(x2, perm=[1, 0, 2])y120 = tf.transpose(x2, perm=[1, 2, 0])y201 = tf.transpose(x2, perm=[2, 0, 1])y210 = tf.transpose(x2, perm=[2, 1, 0])with tf.Session() as sess:    outputy11 = sess.run(y11)    print("outputy11:")    print(outputy11)    print('\n')    outputy12 = sess.run(y12)    print("outputy12:")    print(outputy12)    print('\n')    outputy021 = sess.run(y021)    print("outputy021:")    print(outputy021)    print('\n')    outputy012 = sess.run(y012)    print("outputy012:")    print(outputy012)    print('\n')    outputy102 = sess.run(y102)    print("outputy102:")    print(outputy102)    print('\n')    outputy120 = sess.run(y120)    print("outputy120:")    print(outputy120)    print('\n')    outputy201 = sess.run(y201)    print("outputy201:")    print(outputy201)    print('\n')    outputy210 = sess.run(y210)    print("outputy210:")    print(outputy210)

output:

outputy11:[[1 4] [2 5] [3 6]]outputy12:[[1 4] [2 5] [3 6]]outputy021:[[[ 1  4]  [ 2  5]  [ 3  6]] [[ 7 10]  [ 8 11]  [ 9 12]]]outputy012:[[[ 1  2  3]  [ 4  5  6]] [[ 7  8  9]  [10 11 12]]]outputy102:[[[ 1  2  3]  [ 7  8  9]] [[ 4  5  6]  [10 11 12]]]outputy120:[[[ 1  7]  [ 2  8]  [ 3  9]] [[ 4 10]  [ 5 11]  [ 6 12]]]outputy201:[[[ 1  4]  [ 7 10]] [[ 2  5]  [ 8 11]] [[ 3  6]  [ 9 12]]]outputy210:[[[ 1  7]  [ 4 10]] [[ 2  8]  [ 5 11]] [[ 3  9]  [ 6 12]]]Process finished with exit code 0

3. example 2

import tensorflow as tfimport numpy as npx_anchor = tf.constant([[[0, 1],                         [2, 3],                         [4, 5],                         [6, 7],                         [8, 9],                         [10,11]]], dtype=np.float32)y_anchor102 = tf.transpose(x_anchor, perm=[1, 0, 2])with tf.Session() as sess:    input_anchor = sess.run(x_anchor)    print("input_anchor.shape:")    print(input_anchor.shape)    print('\n')    output_anchor102 = sess.run(y_anchor102)    print("output_anchor102:")    print(output_anchor102)    print('\n')    print("output_anchor102.shape:")    print(output_anchor102.shape)

output:

input_anchor.shape:(1, 6, 2)output_anchor102:[[[  0.   1.]] [[  2.   3.]] [[  4.   5.]] [[  6.   7.]] [[  8.   9.]] [[ 10.  11.]]]output_anchor102.shape:(6, 1, 2)Process finished with exit code 0

4. example 3

import tensorflow as tfimport numpy as np# 'perm' is more useful for n-dimensional tensors, for n > 2x = tf.constant([[[ 1,  2,  3],                  [ 4,  5,  6]],                 [[ 7,  8,  9],                  [10, 11, 12]]])# Take the transpose of the matrices in dimension-0y102 = tf.transpose(x, perm=[1, 0, 2])with tf.Session() as sess:    inputx = sess.run(x)    print("inputx.shape:")    print(inputx.shape)    print('\n')    print("inputx:")    print(inputx)    print('\n')    outputy102 = sess.run(y102)    print("outputy102.shape:")    print(outputy102.shape)    print('\n')    print("outputy102:")    print(outputy102)    print('\n')    print("inputx[1, 0, 0] - 7:")    print(inputx[1, 0, 0])    print("outputy102[0, 1, 0] - 7:")    print(outputy102[0, 1, 0])

output:

inputx.shape:(2, 2, 3)inputx:[[[ 1  2  3]  [ 4  5  6]] [[ 7  8  9]  [10 11 12]]]outputy102.shape:(2, 2, 3)outputy102:[[[ 1  2  3]  [ 7  8  9]] [[ 4  5  6]  [10 11 12]]]inputx[1, 0, 0] - 7:7outputy102[0, 1, 0] - 7:7Process finished with exit code 0

x 中元素7的索引为[1, 0, 0]. 
tf.transpose(x, perm=[1, 0, 2])之后，outputy102中元素7的索引[0, 1, 0].
perm=[1, 0, 2]可以理解为第0维和第1维的元素坐标置换。

outputy102中索引[0, 0, 0]的元素（1）是未transpose之前x中索引[0, 0, 0]元素1.
outputy102中索引[0, 0, 1]的元素（2）是未transpose之前x中索引[0, 0, 1]元素2.
outputy102中索引[0, 0, 2]的元素（3）是未transpose之前x中索引[0, 0, 2]元素3.

outputy102中索引[1, 0, 0]的元素（4）是未transpose之前x中索引[0, 1, 0]元素4.
outputy102中索引[1, 0, 1]的元素（5）是未transpose之前x中索引[0, 1, 1]元素5.
outputy102中索引[1, 0, 2]的元素（6）是未transpose之前x中索引[0, 1, 2]元素6.

outputy102中索引[0, 1, 0]的元素（7）是未transpose之前x中索引[1, 0, 0]元素7.
outputy102中索引[0, 1, 1]的元素（8）是未transpose之前x中索引[1, 0, 1]元素8.
outputy102中索引[0, 1, 2]的元素（9）是未transpose之前x中索引[1, 0, 2]元素9.

outputy102中索引[1, 1, 0]的元素（10）是未transpose之前x中索引[1, 1, 0]元素10.
outputy102中索引[1, 1, 1]的元素（11）是未transpose之前x中索引[1, 1, 1]元素11.
outputy102中索引[1, 1, 2]的元素（12）是未transpose之前x中索引[1, 1, 2]元素12.