tf.transpose() function in TensorFlow

In TensorFlow tf.transpose() is used to rearrange the dimensions of a tensor. It works just like the transpose of a matrix where rows become columns and columns become rows. If you’re working with higher-dimensional data like 3D or 4D tensors tf.transpose() allow you to shuffle the dimensions however you want. The syntax for tf.transpose() is:

Syntax: tf.transpose(input_tensor, perm=None,conjugate=False)

where:

• input_tensor: it is the tensor which is to be transposed.
• perm: This parameters specifies the permutation according to which the input_tensor is to be transposed.
• conjugate: This parameters is set to True if the input_tensor is of type complex.

1. Swapping Rows and Columns in a 2D Tensor

Create a simple 2D tensor which is like a matrix of rows and columns. We will create a constant 2D tensor using tf.constant. Then we’ll use tf.transpose() to swap its rows and columns.

import tensorflow as tf

matrix = tf.constant([[1, 2, 3], [4, 5, 6]])
print("Original Matrix:")
print(matrix.numpy())

transposed_matrix = tf.transpose(matrix)
print("\nTransposed Matrix:")
print(transposed_matrix.numpy())

Output:

Here the rows and columns of the matrix are swapped. This is the most common use case for transposing.

2. Rearranging Dimensions in a 3D Tensor

A 3D tensor could represent something like a collection of images, video frames or other multi-dimensional data. In the below example we’ll manually specify how to rearrange the axes using the perm argument.

tensor = tf.constant([[[1, 2, 3, 4], 
                       [5, 6, 7, 8], 
                       [9, 10, 11, 12]], 

                      [[13, 14, 15, 16], 
                       [17, 18, 19, 20], 
                       [21, 22, 23, 24]]])

print("Original Shape:", tensor.shape)

transposed_tensor = tf.transpose(tensor, perm=[2, 0, 1])
print("\nTransposed Shape:", transposed_tensor.shape)

Output:

The original tensor has 2 blocks each with 3 rows and 4 columns. After transposing we have 4 blocks, each with 2 rows and 3 columns

3. Transposing Without the perm Argument

When you use tf.transpose() without the perm argument TensorFlow will automatically reverse the order of the axes of the tensor.

tensor = tf.constant([[[1, 2], [3, 4]], [[5, 6], [7, 8]]])
print("Original Shape:", tensor.shape)

transposed_tensor = tf.transpose(tensor)
print("\nTransposed Shape:", transposed_tensor.shape)

Output:

In the above output reversing the shape (2, 2, 2) still gives (2, 2, 2) so visually it looks the same but the order of the axes has been flipped:

• Axis 0 becomes axis 2
• Axis 1 stays in the middle
• Axis 2 becomes axis 0

With these methods we can apply transpose to a tensor and can change its dimension according to our needs.