Build a leading batch dimension
Combine several same-shaped feature vectors into one batch-like tensor.
TensorFlow · Function
tf.stack
Stacks equally shaped tensors along a new axis, producing a tensor whose rank is one greater than each input. The axis determines where the new dimension is inserted.
How it works
tf.stack takes tensors with matching shapes and combines them along a selected axis. The result gains one rank: stacking N inputs of rank R produces a rank R+1 tensor, while the axis determines where the new dimension appears.
Syntax
tf.stack(values, axis=0, name='stack')Parameters
| Name | Type | Description |
|---|---|---|
values | list of Tensor | A list of Tensor objects with identical shape and type. |
axis | int | Dimension along which to stack; defaults to the first dimension. Negative values are supported within the valid output-rank range. |
name | string | Optional name for the operation. |
Returns
A stacked Tensor with the same type as values.
Key points
- The operation increases the input rank by exactly one.
- It packs the supplied sequence into a single higher-rank tensor.
- With axis 0, the number of input tensors becomes the leading dimension.
- With axis 1, the new input-count dimension follows the first existing dimension.
- NumPy provides the corresponding operation under np.stack.
Common mistakes
- Every input must have the same shape and data type; tensors with mismatches cannot be stacked.
- Negative axes are allowed, but the value must remain within the output-rank bounds.
- An out-of-range axis raises ValueError.
Example: Stack vectors along axis 1
import tensorflow as tf
x = tf.constant([1, 4])
y = tf.constant([2, 5])
z = tf.constant([3, 6])
result = tf.stack([x, y, z], axis=1)
print(result)Creates a rank-2 tensor by inserting a new axis at position 1.
When to use it
Control dimension order
Arrange corresponding tensor slices in a chosen position within a multidimensional result.
Aggregate computed tensors
Convert a sequence of compatible intermediate tensors into one tensor for later computation.
Interview questions
QWhat does tf.stack do to tensor rank?
tf.stack combines a list of equally shaped tensors into one tensor whose rank is one greater than the inputs. It inserts a new dimension at the requested axis, with axis 0 placing the input list dimension first.
Follow-up: How does axis 1 differ from axis 0 for rank-3 inputs? Axis 0 produces shape (N, A, B, C), whereas axis 1 produces (A, N, B, C).
QWhat are tf.stack's parameters and input requirements?
values is the input tensor list, axis selects the new dimension and defaults to 0, and name optionally names the operation. The inputs must share shape and type.
Follow-up: What constraint applies to axis? For input rank R, valid values are from -(R+1) through R inclusive.
QHow does axis affect the output shape?
For N inputs shaped (A, B, C), stacking on axis 0 creates (N, A, B, C), while stacking on axis 1 creates (A, N, B, C). Thus, axis controls the position of the inserted dimension.
Follow-up: What is the NumPy counterpart? The documentation identifies np.stack as the equivalent operation.
Related
tf.parallel_stacktf.concattf.unstacktf.ragged.stacknp.stack