Normalization Layers

DL4J equivalents and API reference for Keras normalization layers — BatchNormalization.

Keras Normalization Layer Import

Normalization layer support is implemented in the layers/normalization package.

Support Summary

Keras Layer	DL4J Equivalent	Supported
BatchNormalization	BatchNormalization	Yes

---

KerasBatchNormalization

source

Imports a Keras BatchNormalization layer as a DL4J BatchNormalization layer. Both training-mode (with learned gamma/beta) and inference-mode (with running mean/variance) parameters are loaded from the HDF5 file.

Constructor

public KerasBatchNormalization(Integer kerasVersion) throws UnsupportedKerasConfigurationException

Pass-through constructor. kerasVersion is the major Keras version (1 or 2).

getBatchNormalizationLayer

public BatchNormalization getBatchNormalizationLayer()

Returns the configured DL4J BatchNormalization layer. Configuration includes:

• epsilon — small constant added to the variance for numerical stability (default 1e-3 in Keras)

• momentum — momentum for updating running statistics (called decay in Keras 1, momentum in Keras 2)

• center — whether a bias (beta) term is used

• scale — whether a scale (gamma) term is used

getOutputType

public InputType getOutputType(InputType... inputType) throws InvalidKerasConfigurationException

Returns the output InputType. For BatchNormalization, the output shape is the same as the input shape.

Parameters:

• inputType — array of input InputType objects

Returns: output InputType

getNumParams

public int getNumParams()

Returns the number of trainable parameters (4: gamma, beta, running mean, running variance).

setWeights

public void setWeights(Map<String, INDArray> weights) throws InvalidKerasConfigurationException

Loads the four weight tensors from the weight map:

DL4J parameter name	Keras HDF5 tensor
gamma	scale parameter
beta	center parameter
mean	running mean
var	running variance

Parameters:

• weights — map from parameter name to INDArray

---

Keras BatchNormalization Parameters

The following Keras parameters are mapped during import:

Keras parameter	DL4J configuration method
axis	determines which dimension is normalized
momentum	decay() in DL4J builder
epsilon	eps() in DL4J builder
center	whether beta is included
scale	whether gamma is included

Note on axis: Keras BatchNormalization normalizes along the feature axis. In Keras with channels-last (NHWC), this is axis 3. DL4J uses channels-first (NCHW), where the feature axis is 1. The importer translates the axis value automatically.

---

Example

# Python: define and save a model with BatchNormalization
from keras.models import Sequential
from keras.layers import Dense, BatchNormalization

model = Sequential()
model.add(Dense(128, input_dim=64))
model.add(BatchNormalization())
model.add(Dense(10, activation='softmax'))
model.save('bn_model.h5')

// Java: import and run inference
MultiLayerNetwork model = KerasModelImport.importKerasSequentialModelAndWeights("bn_model.h5");
INDArray input = Nd4j.randn(32, 64);
INDArray output = model.output(input);

---

Notes

• At inference time, DL4J uses the running mean and variance (not batch statistics), which matches Keras inference behavior.

• LayerNormalization (Keras) has no direct equivalent in DL4J core and is not currently supported by the importer.