How tf.transpose works in tensorflow?

Question

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

transposes a. It permutes the dimensions according to perm. So if I am using this matrix to transform:

import tensorflow as tt
import os
os.environ["TF_CPP_MIN_LOG_LEVEL"]="3"
import numpy as bb
ab=([[[1,2,3],[6,5,4]],[[4,5,6],[3,6,3]]])
v=bb.array(ab)
fg=tt.transpose(v)
print(v)

with tt.Session() as df:
    print("\n New tranformed matrix is: \n\n{}".format(df.run(fg)))

Result is :

[[[1 2 3]
  [6 5 4]]

 [[4 5 6]
  [3 6 3]]]

 New tranformed matrix is: 

[[[1 4]
  [6 3]]

 [[2 5]
  [5 6]]

 [[3 6]
  [4 3]]]

Process finished with exit code 0

now if i use perm argument then :

import tensorflow as tt
import os
os.environ["TF_CPP_MIN_LOG_LEVEL"]="3"
import numpy as bb
ab=([[[1,2,3],[6,5,4]],[[4,5,6],[3,6,3]]])
v=bb.array(ab)
fg=tt.transpose(v,perm=[0,2,1])
print(v)

with tt.Session() as df:
    print("\n New tranformed matrix is: \n\n{}".format(df.run(fg)))

Result is :

[[[1 2 3]
  [6 5 4]]

 [[4 5 6]
  [3 6 3]]]

 New tranformed matrix is: 

[[[1 6]
  [2 5]
  [3 4]]

 [[4 3]
  [5 6]
  [6 3]]]

Process finished with exit code 0

Due to this, I am confused and I have two questions :

• Whenever I want to transpose a matrix I have to give perm[0,2,1] as default ?
• What is 0,2,1 here ?

Answer 1

Looking at the numpy.transpose documentation, we find that transpose takes the argument

axes : list of ints, optional
By default, reverse the dimensions, otherwise permute the axes according to the values given.

So the default call to transpose translates into np.transpose(a, axes=[1,0]) for the 2D case, or np.transpose(a, axes=[2,1,0]).

The operation you want to have here, is one that leaves the "depth" dimension unchanged. Therefore in the axes argument, the depth axes, which is the 0th axes, needs to stay unchanged. The axes 1 and 2 (where 1 is the vertical axis), need to change positions. So you change the axes order from the initial [0,1,2] to [0,2,1] ([stays the same, changes with other, changes with other]).

In tensorflow, they have for some reason renamed axes to perm. The argument from above stays the same.

images

Concerning images, they differ from the arrays in the question. Images normally have their x and y stored in the first two dimensions and the channel in the last, [y,x,channel].

In order to "transpose" an image in the sense of a 2D transposition, where horizontal and vertical axes are exchanged, you would need to use

np.transpose(a, axes=[1,0,2])

(channel stays the same, x and y are exchanged).