Create a zero-filled buffer
Allocate a known-size tensor for padding, masks, accumulators, or other initial values.
TensorFlow · Function
tf.zeros
Creates a tensor with a specified shape whose elements are all zero. The output dtype defaults to float32 and can be changed with dtype.
How it works
tf.zeros constructs a Tensor from the requested shape and fills every element with zero. The element type defaults to float32, but dtype can be changed; an optional layout can make the result a DTensor.
Syntax
tf.zeros(
shape,
dtype=tf.dtypes.float32,
name=None,
layout=None
)Parameters
| Name | Type | Description |
|---|---|---|
shape | list[int] | tuple[int, ...] | Tensor | A list or tuple of integers, or a one-dimensional int32 Tensor describing the output dimensions. |
dtype | tf.dtypes.DType | The data type of an element in the resulting Tensor; defaults to tf.float32. |
name | str | None | Optional operation name. |
layout | tf.experimental.dtensor.Layout | None | Optional tf.experimental.dtensor.Layout; if supplied, the result is a DTensor with that layout. |
Returns
A Tensor with all elements set to zero.
Key points
- The function produces a tensor filled entirely with zero values.
- The returned tensor uses the requested dtype and shape.
- dtype specifies the type of each element in the output.
- name can optionally assign a name to the operation.
- layout accepts an optional DTensor layout.
Common mistakes
- shape must be a list or tuple of integers, or a one-dimensional int32 Tensor.
- Providing layout changes the result into a DTensor using that layout.
Example: Create an integer zero tensor
import tensorflow as tf
zeros = tf.zeros([3, 4], dtype=tf.int32)
print(zeros)Creates and prints a 3-by-4 integer tensor whose entries are all zero.
When to use it
Build model inputs or intermediates
Provide an explicitly shaped constant tensor as an input or intermediate computation.
Control output dtype
Construct a zero-valued tensor with a non-default numeric type such as int32.
Use a DTensor layout
Create a distributed tensor with an explicitly specified layout.
Interview questions
QWhat does tf.zeros return, and what is its default dtype?
It creates a Tensor whose shape is shape and whose elements are all zero. Its default element type is float32 unless dtype is specified.
Follow-up: How would you create zeros matching an existing tensor? Use tf.zeros_like, which derives the output shape and type from the input.
QWhat forms can the shape argument take?
shape may be a list or tuple of integers, or a one-dimensional int32 Tensor. dtype controls the type of each element in the resulting Tensor.
Follow-up: Can shape be supplied as a scalar? No; the documented shape forms describe dimensions, so a scalar does not represent a valid shape specification.
QWhat are the name and layout arguments used for?
name is an optional operation name. layout is optional; when supplied, the result is a DTensor with that layout.
Follow-up: What alternative creates zeros with the same shape as another value? tf.zeros_like accepts a tensor or array-like input and derives its shape.
Related
tf.zeros_liketf.onestf.filltf.zeros_initializer