tf.zeros(shape) allocates a tensor of the requested shape and initializes every element to 0, using float32 by default, TensorFlow's standard default numeric type, unless a different dtype is specified. It's commonly used to initialize things like a bias vector before training begins, or to pre-allocate a tensor of a known shape that will be filled in or accumulated into as computation proceeds.

1Understanding tf.zeros()

💡

Use tf.zeros_like(other_tensor) instead of tf.zeros(other_tensor.shape) when you want a zero-filled tensor matching another tensor's shape and dtype exactly, without needing to read and pass those properties yourself.

editor.html

import tensorflow as tf

zeros = tf.zeros([3])
print(zeros)

localhost:3000

2Practical Example

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

editor.html

import tensorflow as tf

zeros_matrix = tf.zeros([2, 3], dtype=tf.int32)
print(zeros_matrix)

localhost:3000

3Best Practices

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

1. Use tf.zeros() to initialize bias vectors or other values that should conventionally start at zero, rather than an arbitrary placeholder

2. Use tf.zeros_like() instead of manually reading and passing another tensor's shape and dtype, when you want a matching zero-filled tensor

3. Specify dtype explicitly when the default float32 doesn't match what a specific operation or layer actually expects

⚠️

Tip: Use tf.zeros_like(other_tensor) instead of tf.zeros(other_tensor.shape) when you want a zero-filled tensor matching another tensor's shape and dtype exactly, without needing to read and pass those properties yourself.

editor.html

import tensorflow as tf

zeros = tf.zeros([3])
print(zeros)

localhost:3000

Examples

Example 01Basic Usage

import tensorflow as tf

zeros = tf.zeros([3])
print(zeros)

Example 02Advanced Example

import tensorflow as tf

zeros_matrix = tf.zeros([2, 3], dtype=tf.int32)
print(zeros_matrix)

Best Practices

Use tf.zeros() to initialize bias vectors or other values that should conventionally start at zero, rather than an arbitrary placeholder
Use tf.zeros_like() instead of manually reading and passing another tensor's shape and dtype, when you want a matching zero-filled tensor
Specify dtype explicitly when the default float32 doesn't match what a specific operation or layer actually expects

Interview Question

Why does tf.zeros() default to float32 rather than a lower-precision or integer type?▼

Hint: Think about what TensorFlow's computations, like gradient-based training, typically require numerically.

float32 is TensorFlow's standard default numeric type because most machine learning computations, especially gradient-based training via backpropagation, require floating-point arithmetic to represent the small, continuous weight updates and gradients involved — an integer type couldn't meaningfully represent those fractional adjustments at all. float32 specifically balances enough numerical precision for stable training against reasonable memory usage and computational speed, which is why it's the conventional default across nearly all of TensorFlow's tensor-creation functions, including zeros(), unless a specific reason calls for a different precision.

Exercises

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

View Solution

import tensorflow as tf

zeros = tf.zeros([3])
print(zeros)

Frequently Asked Questions

Why does tf.zeros() default to float32 rather than a lower-precision or integer type?▼

Related Functions

tf-onestf-constanttf-cast