tf.keras.layers.Normalization

A preprocessing layer which normalizes continuous features.

Inherits From: PreprocessingLayer, Layer, Module

This layer will shift and scale inputs into a distribution centered around 0 with standard deviation 1. It accomplishes this by precomputing the mean and variance of the data, and calling (input - mean) / sqrt(var) at runtime.

The mean and variance values for the layer must be either supplied on construction or learned via adapt(). adapt() will compute the mean and variance of the data and store them as the layer's weights. adapt() should be called before fit(), evaluate(), or predict().

For an overview and full list of preprocessing layers, see the preprocessing guide.

axis Integer, tuple of integers, or None. The axis or axes that should have a separate mean and variance for each index in the shape. For example, if shape is (None, 5) and axis=1, the layer will track 5 separate mean and variance values for the last axis. If axis is set to None, the layer will normalize all elements in the input by a scalar mean and variance. Defaults to -1, where the last axis of the input is assumed to be a feature dimension and is normalized per index. Note that in the specific case of batched scalar inputs where the only axis is the batch axis, the default will normalize each index in the batch separately. In this case, consider passing axis=None.
mean The mean value(s) to use during normalization. The passed value(s) will be broadcast to the shape of the kept axes above; if the value(s) cannot be broadcast, an error will be raised when this layer's build() method is called.
variance The variance value(s) to use during normalization. The passed value(s) will be broadcast to the shape of the kept axes above; if the value(s) cannot be broadcast, an error will be raised when this layer's build() method is called.

Examples:

Calculate a global mean and variance by analyzing the dataset in adapt().

adapt_data = np.array([1., 2., 3., 4., 5.], dtype='float32')
input_data = np.array([1., 2., 3.], dtype='float32')
layer = tf.keras.layers.Normalization(axis=None)
layer.adapt(adapt_data)
layer(input_data)
<tf.Tensor: shape=(3,), dtype=float32, numpy=
array([-1.4142135, -0.70710677, 0.], dtype=float32)>

Calculate a mean and variance for each index on the last axis.

adapt_data = np.array([[0., 7., 4.],
                       [2., 9., 6.],
                       [0., 7., 4.],
                       [2., 9., 6.]], dtype='float32')
input_data = np.array([[0., 7., 4.]], dtype='float32')
layer = tf.keras.layers.Normalization(axis=-1)
layer.adapt(adapt_data)
layer(input_data)
<tf.Tensor: shape=(1, 3), dtype=float32, numpy=
array([0., 0., 0.], dtype=float32)>

Pass the mean and variance directly.

input_data = np.array([[1.], [2.], [3.]], dtype='float32')
layer = tf.keras.layers.Normalization(mean=3., variance=2.)
layer(input_data)
<tf.Tensor: shape=(3, 1), dtype=float32, numpy=
array([[-1.4142135 ],
       [-0.70710677],
       [ 0.        ]], dtype=float32)>

is_adapted Whether the layer has been fit to data already.

Methods

adapt

View source

Fits the state of the preprocessing layer to the data being passed.

After calling adapt on a layer, a preprocessing layer's state will not update during training. In order to make preprocessing layers efficient in any distribution context, they are kept constant with respect to any compiled tf.Graphs that call the layer. This does not affect the layer use when adapting each layer only once, but if you adapt a layer multiple times you will need to take care to re-compile any compiled functions as follows:

  • If you are adding a preprocessing layer to a keras.Model, you need to call model.compile after each subsequent call to adapt.
  • If you are calling a preprocessing layer inside tf.data.Dataset.map, you should call map again on the input tf.data.Dataset after each adapt.
  • If you are using a tf.function directly which calls a preprocessing layer, you need to call tf.function again on your callable after each subsequent call to adapt.

tf.keras.Model example with multiple adapts:

layer = tf.keras.layers.experimental.preprocessing.Normalization(
    axis=None)
layer.adapt([0, 2])
model = tf.keras.Sequential(layer)
model.predict([0, 1, 2])
array([-1.,  0.,  1.], dtype=float32)
layer.adapt([-1, 1])
model.compile() # This is needed to re-compile model.predict!
model.predict([0, 1, 2])
array([0., 1., 2.], dtype=float32)

tf.data.Dataset example with multiple adapts:

layer = tf.keras.layers.experimental.preprocessing.Normalization(
    axis=None)
layer.adapt([0, 2])
input_ds = tf.data.Dataset.range(3)
normalized_ds = input_ds.map(layer)
list(normalized_ds.as_numpy_iterator())
[array([-1.], dtype=float32),
 array([0.], dtype=float32),
 array([1.], dtype=float32)]
layer.adapt([-1, 1])
normalized_ds = input_ds.map(layer) # Re-map over the input dataset.
list(normalized_ds.as_numpy_iterator())
[array([0.], dtype=float32),
 array([1.], dtype=float32),
 array([2.], dtype=float32)]

Arguments
data The data to train on. It can be passed either as a tf.data Dataset, or as a numpy array.
batch_size Integer or None. Number of samples per state update. If unspecified, batch_size will default to 32. Do not specify the batch_size if your data is in the form of datasets, generators, or keras.utils.Sequence instances (since they generate batches).
steps Integer or None. Total number of steps (batches of samples) When training with input tensors such as TensorFlow data tensors, the default None is equal to the number of samples in your dataset divided by the batch size, or 1 if that cannot be determined. If x is a tf.data dataset, and 'steps' is None, the epoch will run until the input dataset is exhausted. When passing an infinitely repeating dataset, you must specify the steps argument. This argument is not supported with array inputs.

compile

View source

Configures the layer for adapt.

Arguments
run_eagerly Bool. Defaults to False. If True, this Model's logic will not be wrapped in a tf.function. Recommended to leave this as None unless your Model cannot be run inside a tf.function. steps_per_execution: Int. Defaults to 1. The number of batches to run during each tf.function call. Running multiple batches inside a single tf.function call can greatly improve performance on TPUs or small models with a large Python overhead.

reset_state

View source

Resets the statistics of the preprocessing layer.

update_state

View source

Accumulates statistics for the preprocessing layer.

Arguments
data A mini-batch of inputs to the layer.