Tensor Broadcasting Explained with Examples

In this post, you will learn about the concepts of Tensor Broadcasting with the help of Python Numpy examples. Recall that Tensor is defined as the container of data (primarily numerical) most fundamental data structure used in Keras and Tensorflow. You may want to check out a related article on Tensor – Tensor explained with Python Numpy examples.

Broadcasting of tensor is borrowed from Numpy broadcasting. Broadcasting is technique used for performing arithmetic operations between Numpy arrays / Tensors having different shapes. In this technique, the smaller array is transformed appropriately according to larger array (broadcasted to large array) such that the arithmetic operations can be performed on these arrays. 

Take a look at the following example where there are two Numpy arrays, X and Y having shape as (2, 2) and (1, ) respectively. Let’s perform the operation to add X and Y and understand the broadcasting concept in detailed manner. Here is the code representing X & Y created using np.random.random  

X = np.random.random((2, 2))
Y = np.random.random((1,))
# Print arrays shape
print('Shape of X: ', X.shape, '\nShape of Y: ', Y.shape, '\n')
# Print the array values
print('Array X: ', X, '\n\nArray Y: ', Y)

Here is the output of X & Y including their shapes. You may note that the shape of X is (2, 2) – 2D Tensor and Y is (1, ) – 1D Tensor. We will try and perform arithmetic operations on 2D Tensor and 1D Tensor. They have different ranks and shapes. This is where Tensor Broadcasting would come into picture where the Tensors / Numpy arrays will be transformed into compatible shapes to perform the arithmetic operations.

Numpy Arrays Shape
Fig 1. Numpy Arrays Shape

In order to bring the Tensors / Numpy Arrays into compatible shape, this is what is done (as part of broadcasting process):

  1. New axes (can be termed as broadcast axes) are added to the smaller tensor to match the ndim of the larger tensor. In the example above, ndim of larger array / Tensor (X) is 2 and ndim of smaller tensor (Y) is 1. New axes will need to be added to Y (1D Tensor).
  2. The smaller tensor is repeated alongside these new axes to match the full shape of the larger tensor (X – 2D tensor).

Based on above, Y will be internally transformed to the following before arithmetic operation is performed:

Y1 = np.array([[0.94992891, 0.94992891], [0.94992891, 0.94992891]])

Lets validate this by performing arithmetic operation X + Y and X + Y1. You may find that both X + Y and X + Y1 ends up giving same result.

Arithmetic Operations on Tensors having different axes / shapes
Fig 2. Arithmetic Operations on Tensors having different axes / shapes


Here is the summary of what you learned about the Tensor broadcasting:

  • Tensor broadcasting concept is borrowed from Numpy broadcasting.
  • Tensor broadcasting is about bringing the tensors of different dimensions / shape to the compatible shape such that arithmetic operations can be performed on them.
  • In broadcasting, the smaller array is found, the new axes are added as per the larger array and data is added appropriately to the transformed array.

Ajitesh Kumar
Follow me
Latest posts by Ajitesh Kumar (see all)

Ajitesh Kumar

I have been recently working in the area of Data Science and Machine Learning / Deep Learning. In addition, I am also passionate about various different technologies including programming languages such as Java/JEE, Javascript, Python, R, Julia etc and technologies such as Blockchain, mobile computing, cloud-native technologies, application security, cloud computing platforms, big data etc. I would love to connect with you on Linkedin.
Posted in Data Science, Deep Learning, Machine Learning, Python. Tagged with , , , , .

Leave a Reply

Your email address will not be published. Required fields are marked *

Time limit is exhausted. Please reload the CAPTCHA.