Prepare structured features
Convert a flat feature vector into a matrix or higher-rank tensor before subsequent layers.
TensorFlow · Class
layers.Reshape
tf.keras.layers.Reshape changes each input sample to a specified target shape while preserving the batch dimension. It requires fixed input dimensions and can infer one target dimension with -1.
How it works
Reshape changes the non-batch dimensions of each input sample to target_shape while preserving the batch axis. The input dimensions must be known and fixed, and the target shape must contain the same number of elements, with -1 optionally allowing shape inference.
Syntax
tf.keras.layers.Reshape(target_shape, **kwargs)Parameters
| Name | Type | Description |
|---|---|---|
target_shape | tuple of integers | Target shape for each sample, represented as integers and excluding the samples or batch dimension. |
Returns
A tensor with shape (batch_size, *target_shape).
Key points
- The layer changes the arrangement of input elements to the requested per-sample shape.
- Construct it with the desired target_shape and optional standard layer keyword arguments.
- The resulting tensor retains the batch axis and appends the requested target dimensions.
- When used as a model's first layer, input_shape can declare the sample dimensions supplied to the layer.
- A -1 target dimension can be inferred automatically from the other dimensions.
Common mistakes
- Do not put the batch size in target_shape; the layer preserves that dimension automatically.
- Inputs with unknown or dynamic non-batch dimensions are not supported by this layer.
- Use at most one inferred dimension, and ensure the remaining dimensions determine a compatible total size.
Example: Reshape a feature vector
import tensorflow as tf
x = tf.keras.Input(shape=(12,))
y = tf.keras.layers.Reshape((3, 4))(x)
print(y.shape) # (None, 3, 4)This creates a symbolic 12-element input and reshapes each sample into three rows of four elements.
When to use it
Reconstruct decoded outputs
Restore a decoder's flat output to the dimensions of an encoded image or other data representation.
Build model shape pipelines
Insert a symbolic shape transformation inside a Functional or Sequential Keras model.
Interview questions
QWhat does tf.keras.layers.Reshape do to the batch dimension?
Reshape changes each sample's layout without changing its number of elements, and it leaves the batch dimension untouched. For an input shaped (batch_size, 12) and target_shape (3, 4), the output is (batch_size, 3, 4).
Follow-up: What must be true for the reshape to be valid? The input dimensions must be known and fixed, and the target dimensions must represent the same element count, aside from one optionally inferred dimension.
QWhat is the meaning of target_shape?
target_shape is a tuple of integers describing the desired per-sample shape. It excludes the samples or batch axis, which Keras supplies and preserves separately.
Follow-up: How would you represent a batch of 12-element vectors becoming 2-by-6 tensors? Use target_shape=(2, 6), not (batch_size, 2, 6).
QHow do you determine the output shape?
The output shape is (batch_size, *target_shape). Thus, Reshape((3, 4)) applied to an input with shape (12,) produces an output shape of (None, 3, 4) in a symbolic model.
Follow-up: Can a target dimension be inferred? Yes; the documentation demonstrates -1, as in (-1, 2, 2), when the element count allows that dimension to be determined.
Related
tf.keras.layers.Layertf.keras.layers.Flattentf.keras.layers.Dense