tf.constant() creates an immutable tensor from a Python list, NumPy array, or scalar value, with a fixed value that can never be changed after creation.

Syntax

tf.constant(value, dtype=None, shape=None)

Deep Dive Course

A constant tensor's value is fixed at creation time and cannot be reassigned or updated afterward — any operation that appears to modify it actually produces a brand-new tensor instead, leaving the original constant untouched. This immutability makes tf.constant() the right choice for fixed input data, hyperparameters, or any value that should never change during a model's execution, in contrast to tf.Variable(), which is specifically designed to hold values that do need to change, like a model's trainable weights.

1Understanding tf.constant()

💡

Use tf.constant() for genuinely fixed values, like fixed input data or hyperparameters, and tf.Variable() specifically for values that need to be updated during training, like model weights — mixing the two up, like trying to use a constant for trainable weights, will fail, since gradients can only be applied to variables.

editor.html

import tensorflow as tf

x = tf.constant([1, 2, 3])
print(x)

localhost:3000

2Practical Example

Here is a real-world application of tf.constant() showing how it is used in production TensorFlow code.

editor.html

import tensorflow as tf

matrix = tf.constant([[1, 2], [3, 4]], dtype=tf.float32)
print(matrix)

localhost:3000

3Best Practices

Follow these guidelines when working with tf.constant():

1. Use tf.constant() for input data, fixed configuration values, or anything that shouldn't change during execution

2. Use tf.Variable() instead whenever a value needs to be updated over time, like model weights, since constants cannot be reassigned

3. Specify dtype explicitly when the automatically-inferred type from your input data doesn't match what your model or subsequent operations actually expect

⚠️

Tip: Use tf.constant() for genuinely fixed values, like fixed input data or hyperparameters, and tf.Variable() specifically for values that need to be updated during training, like model weights — mixing the two up, like trying to use a constant for trainable weights, will fail, since gradients can only be applied to variables.

editor.html

import tensorflow as tf

x = tf.constant([1, 2, 3])
print(x)

localhost:3000

Examples

Example 01Basic Usage

import tensorflow as tf

x = tf.constant([1, 2, 3])
print(x)

Example 02Advanced Example

import tensorflow as tf

matrix = tf.constant([[1, 2], [3, 4]], dtype=tf.float32)
print(matrix)

Best Practices

Use tf.constant() for input data, fixed configuration values, or anything that shouldn't change during execution
Use tf.Variable() instead whenever a value needs to be updated over time, like model weights, since constants cannot be reassigned
Specify dtype explicitly when the automatically-inferred type from your input data doesn't match what your model or subsequent operations actually expect

Interview Question

Why can't tf.constant() be used to hold a neural network's trainable weights?▼

Hint: Think about what training a model actually requires doing to its weight values over time.

Training a model works by repeatedly updating its weights based on computed gradients, adjusting each weight slightly after every training step to reduce the loss — but a tf.constant()'s value is fixed permanently at creation and can never be reassigned or modified in place. tf.Variable() exists specifically to support this kind of mutable, trainable state: it holds a value that can be updated over time via its assign methods, which optimizers use internally, which is exactly the capability model weights require and constants deliberately don't have.

Exercises

MediumPractice using tf.constant() in a real scenario.

View Solution

import tensorflow as tf

x = tf.constant([1, 2, 3])
print(x)

Frequently Asked Questions

Why can't tf.constant() be used to hold a neural network's trainable weights?▼

Related Functions

tf-variabletf-zerostf-cast