Join along the first dimension
Combine batches or rows with matching feature widths.
TensorFlow · Function
tf.concat
Concatenates tensors along an existing dimension. Inputs must match in rank and in every dimension except the selected axis.
How it works
tf.concat joins a list of tensors along the dimension selected by axis. Inputs must have the same rank and identical sizes everywhere except that dimension; the selected dimension grows by the sum of its input sizes.
Syntax
tf.concat(values, axis, name='concat')Parameters
| Name | Type | Description |
|---|---|---|
values | Tensor or list of Tensor | A list of Tensor objects or a single Tensor. |
axis | 0-D int32 Tensor | The dimension along which to concatenate; a 0-D int32 Tensor in the range [-rank(values), rank(values)). |
name | string | Optional name for the operation. |
Returns
A Tensor resulting from concatenation of the input tensors.
Key points
- Use tf.concat to join tensors without introducing a new dimension.
- values accepts either a tensor list or one tensor.
- The output's selected dimension equals the sum of the corresponding input dimensions.
- Negative axis values count dimensions from the end.
- Dimension indices use zero-based numbering.
- name optionally assigns an operation name.
Common mistakes
- Inputs cannot differ in rank or in any dimension other than the concatenation dimension.
- axis must be a valid dimension index, including the documented negative-index range.
Example: Concatenate rows
import tensorflow as tf
left = tf.constant([[1, 2], [3, 4]])
right = tf.constant([[5, 6]])
result = tf.concat([left, right], axis=0)
print(result.numpy())Joins the two rank-2 tensors by adding the second tensor's rows.
When to use it
Join along a feature dimension
Append feature columns while preserving the batch dimension.
Use a runtime axis
Concatenate tensors when the dimension is computed dynamically.
Interview questions
QWhat does tf.concat do, and what shape constraints apply?
tf.concat joins tensors along an existing dimension, so the rank stays unchanged. The inputs must have matching ranks and equal sizes in every non-concatenated dimension.
Follow-up: How does tf.stack differ? tf.stack creates a new dimension, whereas tf.concat extends an existing one.
QHow is axis interpreted?
axis selects the dimension to extend. It is zero-based, and negative values count from the end of the tensor rank.
Follow-up: For rank 3, which dimension does axis=-1 select? It selects dimension 2, the final dimension.
QHow is the output shape determined?
The output size along the selected dimension is the sum of the corresponding sizes from all inputs. The other dimensions retain their common sizes.
Follow-up: What happens if those other dimensions do not match? The operation cannot satisfy the required shape constraints and is invalid.
Related
tf.stacktf.expand_dimstf.sparse.concat