Broadcasting rules in TensorFlow

Introduction

Broadcasting lets TensorFlow do math on arrays of different shapes easily, like stretching smaller arrays to match bigger ones without copying data.

Adding a vector to each row of a matrix without writing loops.

Multiplying a tensor by a scalar value to scale all elements.

Applying a bias vector to each example in a batch during neural network training.

Combining images and filters of different shapes in convolution operations.

Performing element-wise operations on tensors with compatible but different shapes.

Syntax

TensorFlow

result = tensor1 + tensor2

TensorFlow automatically applies broadcasting when shapes differ but are compatible.

Broadcasting follows specific rules to align shapes from the right side.

Examples

Adds a vector of shape (4,) to each row of a matrix shape (3,4) by broadcasting the vector.

TensorFlow

import tensorflow as tf

# tensor1 shape: (3, 4)
tensor1 = tf.constant([[1, 2, 3, 4],
                       [5, 6, 7, 8],
                       [9, 10, 11, 12]])

# tensor2 shape: (4,)
tensor2 = tf.constant([1, 0, 1, 0])

result = tensor1 + tensor2
print(result)

Broadcasts tensor2 shape (3,) to (1,3) and then to (2,3,3) to add to tensor1 shape (2,3,1).

TensorFlow

import tensorflow as tf

# tensor1 shape: (2, 3, 1)
tensor1 = tf.constant([[[1], [2], [3]],
                       [[4], [5], [6]]])

# tensor2 shape: (3,)
tensor2 = tf.constant([10, 20, 30])

result = tensor1 + tensor2
print(result)

Sample Model

This program shows how TensorFlow adds a vector to each row of a matrix by broadcasting the vector shape (3,) to match the matrix shape (2,3). It prints the shapes, the result, and the mean value of the result tensor.

TensorFlow

import tensorflow as tf

# Define a 2D tensor (matrix) of shape (2, 3)
matrix = tf.constant([[1, 2, 3],
                      [4, 5, 6]])

# Define a 1D tensor (vector) of shape (3,)
vector = tf.constant([10, 20, 30])

# Add vector to each row of matrix using broadcasting
result = matrix + vector

# Print shapes and result
print(f"matrix shape: {matrix.shape}")
print(f"vector shape: {vector.shape}")
print("result:")
print(result)

# Calculate mean of result tensor
mean_value = tf.reduce_mean(result)
print(f"mean of result: {mean_value.numpy():.2f}")

Important Notes

Broadcasting compares shapes from right to left and matches dimensions if they are equal or one of them is 1.

If shapes are not compatible, TensorFlow will raise an error.

Broadcasting avoids copying data and makes code simpler and faster.

Summary

Broadcasting lets you do math on tensors with different but compatible shapes.

TensorFlow automatically stretches smaller tensors to match bigger ones following simple rules.

This helps write clean and efficient code without loops.

Practice

1. What does broadcasting in TensorFlow allow you to do?

easy

A. Perform math operations on tensors with different but compatible shapes

B. Convert tensors into Python lists automatically

C. Change the data type of tensors without copying data

D. Create new tensors with random values

Solution

1. Step 1: Understand broadcasting concept

   Broadcasting lets TensorFlow perform element-wise operations on tensors even if their shapes differ, as long as they are compatible.

2. Step 2: Identify the correct description

   Only Perform math operations on tensors with different but compatible shapes correctly describes this feature; others describe unrelated tensor operations.

3. Final Answer:

   Perform math operations on tensors with different but compatible shapes -> Option A

4. Quick Check:

   Broadcasting = math on compatible shapes [OK]

Hint: Broadcasting = math on tensors with compatible shapes [OK]

Common Mistakes:

• Thinking broadcasting changes data types
• Confusing broadcasting with tensor creation
• Assuming broadcasting converts tensors to lists

2. Which of the following TensorFlow code snippets correctly broadcasts a tensor of shape (3, 1) with a tensor of shape (1, 4)?

easy

A. tf.constant([1, 2, 3]) + tf.constant([4, 5, 6, 7])

B. tf.constant([[1], [2], [3]]) + tf.constant([[4, 5, 6, 7]])

C. tf.constant([[1, 2, 3]]) + tf.constant([[4], [5], [6], [7]])

D. tf.constant([[1], [2], [3]]) + tf.constant([[4], [5], [6], [7]])

Solution

1. Step 1: Check shapes of tensors in each option

   tf.constant([[1], [2], [3]]) + tf.constant([[4, 5, 6, 7]]) adds (3,1) tensor to (1,4) tensor, which are compatible for broadcasting.

2. Step 2: Verify broadcasting rules

   Shapes (3,1) and (1,4) broadcast to (3,4). Other options have incompatible shapes or wrong dimensions.

3. Final Answer:

   tf.constant([[1], [2], [3]]) + tf.constant([[4, 5, 6, 7]]) -> Option B

4. Quick Check:

   Shapes (3,1) + (1,4) broadcast correctly [OK]

Hint: Match trailing dims: 1 and N broadcast fine [OK]

Common Mistakes:

• Ignoring shape dimensions order
• Assuming 1D tensors broadcast like 2D
• Mixing up rows and columns in shapes

3. What is the output shape of the following TensorFlow operation?

import tensorflow as tf
x = tf.constant([[1, 2, 3]])  # shape (1, 3)
y = tf.constant([4, 5, 6, 7])  # shape (4,)
z = x + y

medium

A. (1, 3, 4)

B. (4, 3)

C. (1, 4)

D. Error due to incompatible shapes

Solution

1. Step 1: Analyze shapes of x and y

   x has shape (1,3), y has shape (4,). TensorFlow aligns shapes from the right.

2. Step 2: Apply broadcasting rules

   y's shape (4,) is treated as (1,4). Shapes (1,3) and (1,4) are incompatible because 3 != 4 and neither is 1.

3. Step 3: Check if broadcasting possible

   Since last dimensions differ and neither is 1, broadcasting fails, causing an error.

4. Final Answer:

   Error due to incompatible shapes -> Option D

5. Quick Check:

   Incompatible shapes cause error [OK]

Hint: Broadcast dims must be equal or 1 from right [OK]

Common Mistakes:

• Assuming (4,) broadcasts to (3,)
• Ignoring dimension order in broadcasting
• Expecting automatic reshaping without error

4. You have two tensors:
a = tf.constant([[1, 2, 3], [4, 5, 6]]) (shape (2, 3))
b = tf.constant([1, 2]) (shape (2,))
Why does a + b raise an error, and how can you fix it?

medium

A. Shapes are incompatible; reshape b to (2,1) before adding

B. Data types differ; cast b to a's dtype

C. Tensors must be same shape; reshape a to (2,2)

D. Broadcasting always works; error is from another cause

Solution

1. Step 1: Check shapes of a and b

   a is (2,3), b is (2,). Broadcasting compares from right: 3 vs 2 incompatible.

2. Step 2: Fix shape for broadcasting

   Reshape b to (2,1) so shapes become (2,3) and (2,1), which broadcast to (2,3).

3. Final Answer:

   Shapes are incompatible; reshape b to (2,1) before adding -> Option A

4. Quick Check:

   Reshape b to (2,1) fixes broadcasting [OK]

Hint: Match dims from right; add missing dims with reshape [OK]

Common Mistakes:

• Ignoring shape mismatch causes error
• Trying to reshape a incorrectly
• Assuming broadcasting fixes all shape issues automatically

5. Given a tensor t of shape (5, 1, 3), you want to add a bias tensor b of shape (3,) to each element along the last dimension. Which code correctly applies broadcasting to add b to t?

hard

A. t + tf.reshape(b, (3, 1))

B. t + tf.reshape(b, (3, 1, 1))

C. t + tf.reshape(b, (1, 1, 3))

D. t + tf.reshape(b, (5, 1, 3))

Solution

1. Step 1: Understand shapes and broadcasting

   t shape is (5,1,3), b shape is (3,). To add b along last dim, b must broadcast to (5,1,3).

2. Step 2: Reshape b for broadcasting

   Reshape b to (1,1,3) so it broadcasts correctly across first two dims.

3. Step 3: Check other options

   A reshapes to (3,1), which pads to (1,3,1) and mismatches middle dim. B reshapes to (3,1,1), mismatching first dim. D fails due to element count mismatch (3 vs 15).

4. Final Answer:

   t + tf.reshape(b, (1, 1, 3)) -> Option C

5. Quick Check:

   Reshape bias to (1,1,3) for last-dim addition [OK]

Hint: Reshape bias to add dims before last dimension [OK]

Common Mistakes:

• Not reshaping bias tensor correctly
• Assuming 1D tensor broadcasts without reshape
• Reshaping bias to wrong shape