Add a batch axis
Prepare a single example for a model that expects a batch dimension.
TensorFlow · Function
tf.expand_dims
Inserts a length-1 axis into a tensor at a chosen position. It uses zero-based Python indexing, with negative axes counted from the end.
How it works
tf.expand_dims creates a new axis of length 1 at the requested position while preserving the input values. The position uses zero-based Python indexing, and negative positions count backward from the end.
Syntax
tf.expand_dims(input, axis, name=None)Parameters
| Name | Type | Description |
|---|---|---|
input | Tensor | A Tensor. |
axis | int | Integer specifying the dimension index at which to expand the shape of input. For an input of D dimensions, it must be in [-(D+1), D], inclusive. |
name | string | Optional string naming the output Tensor. |
Returns
A tensor with the same data as input, with an additional dimension inserted at the index specified by axis.
Key points
- The result gains exactly one new axis, and that axis has length 1.
- The new singleton dimension is inserted into the input shape at the chosen position.
- Axis numbering starts at zero; negative axes select positions relative to the end.
- Typical applications include adding a batch axis, preparing shapes for broadcasting, and adding an inner vector-length axis.
- Passing zero places the new singleton dimension before all existing dimensions.
- Using a negative axis places the singleton dimension relative to the tensor's trailing dimensions.
Common mistakes
axismust be within the input-rank-dependent valid range; an out-of-range value raises an error.- For a rank-D input, valid axes run from
-(D+1)throughD, inclusive.
Example: Add a batch axis
import tensorflow as tf
image = tf.zeros([10, 10, 3])
batched = tf.expand_dims(image, axis=0)
print(batched.shape) # (1, 10, 10, 3)Adds a leading batch dimension to a single image tensor.
When to use it
Align broadcasting axes
Make tensor ranks or axis positions compatible for elementwise broadcasting.
Add an inner singleton axis
Represent scalar values with an explicit final vector-length dimension.
Interview questions
QWhat does tf.expand_dims do?
It inserts a dimension whose length is 1 at the specified axis and returns a tensor with the same data as the input. The axis follows zero-based Python indexing, including negative indices.
Follow-up: How would you add a batch dimension to an image shaped [10, 10, 3]? Use tf.expand_dims(image, axis=0), producing shape [1, 10, 10, 3].
QWhich axis values are valid?
For an input with rank D, axis may be any integer from -(D+1) to D, inclusive. Negative axes count backward from the end, so axis=-1 inserts the innermost dimension.
Follow-up: What shape results from expanding a [10, 10, 3] tensor with axis=-1? The result is [10, 10, 3, 1].
QWhat are common uses for this function?
The operation is useful for adding an outer batch dimension, aligning axes for broadcasting, or adding an inner vector-length axis to scalar tensors. It changes shape but does not alter the tensor's data.
Follow-up: When would an inner axis be useful? It can make tensor shapes compatible with an operation that expects a final singleton feature or vector-length dimension.
Related
tf.squeezetf.reshapetf.sparse.expand_dims