Wrapper Layers

DL4J equivalents and API reference for Keras wrapper layers — Bidirectional and TimeDistributed.

Keras Wrapper Layer Import

Wrapper layers in Keras apply a transformation around an inner layer. Support is implemented in the layers/wrappers package.

Support Summary

Keras Wrapper	DL4J Equivalent	Supported
Bidirectional	Bidirectional	Yes
TimeDistributed	—	No

---

KerasBidirectional

source

Builds a DL4J Bidirectional layer from a Keras Bidirectional layer wrapper. The importer reads the wrapped RNN layer (LSTM or SimpleRNN) and constructs a DL4J Bidirectional wrapper around it.

Constructor

public KerasBidirectional(Integer kerasVersion) throws UnsupportedKerasConfigurationException

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

getUnderlyingRecurrentLayer

public Layer getUnderlyingRecurrentLayer()

Returns the underlying DL4J recurrent layer (e.g., LSTM or SimpleRnn) that was wrapped by the Keras Bidirectional wrapper.

Throws: InvalidKerasConfigurationException if the inner layer configuration is invalid; UnsupportedKerasConfigurationException if the inner layer type is not supported.

getBidirectionalLayer

public Bidirectional getBidirectionalLayer()

Returns the DL4J Bidirectional layer, which wraps the underlying recurrent layer and runs it in both forward and backward directions. The merge mode (sum, multiply, concat, average) is read from the Keras configuration.

Merge mode mapping:

Keras merge_mode	DL4J Bidirectional.Mode
"sum"	ADD
"mul"	MUL
"concat"	CONCAT
"ave"	AVERAGE

getOutputType

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

Returns the output InputType. For concat merge mode, the output size is doubled relative to the inner RNN's output size.

Parameters:

• inputType — array of input InputType objects

getNumParams

public int getNumParams()

Returns the total number of trainable parameters (twice the underlying layer's count, once for each direction).

getInputPreprocessor

public InputPreProcessor getInputPreprocessor(InputType... inputType)
        throws InvalidKerasConfigurationException

Returns the appropriate InputPreProcessor for format conversion before the bidirectional layer.

setWeights

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

Loads weights for both the forward and backward RNN layers from the weight map. Keras stores bidirectional weights with forward_ and backward_ prefixes; the importer splits these into the appropriate DL4J parameter names.

Parameters:

• weights — map from parameter name to INDArray

---

Example

# Python: Bidirectional LSTM
from keras.models import Sequential
from keras.layers import Embedding, Bidirectional, LSTM, Dense

model = Sequential()
model.add(Embedding(10000, 128, input_length=50))
model.add(Bidirectional(LSTM(64)))
model.add(Dense(1, activation='sigmoid'))
model.compile(optimizer='adam', loss='binary_crossentropy')
model.save('bidirectional_lstm.h5')

// Java: import
import org.deeplearning4j.nn.modelimport.keras.KerasModelImport;
import org.deeplearning4j.nn.multilayer.MultiLayerNetwork;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.linalg.factory.Nd4j;

MultiLayerNetwork model = KerasModelImport
    .importKerasSequentialModelAndWeights("bidirectional_lstm.h5");

// Input: [batch, sequence_length] integer token ids
INDArray input = Nd4j.create(new float[]{1, 5, 3, 0, 0}, new int[]{1, 5});
INDArray output = model.output(input);
// output shape: [1, 1] (single sigmoid output)

---

TimeDistributed

TimeDistributed is not currently supported. This wrapper applies an inner layer (e.g., Dense) independently to each timestep of a sequence.

Workarounds:

• For dense operations on sequences, consider replacing TimeDistributed(Dense(n)) with a 1D convolutional layer with kernel size 1, which applies the same transformation at each timestep.

• For feature extraction followed by an RNN, restructure the architecture to use a convolutional encoder before the recurrent layers.

---

Notes

• The Bidirectional wrapper works with both LSTM and SimpleRNN inner layers. Wrapping a GRU is not supported because GRU itself is unsupported.

• For Bidirectional with return_sequences=True on the inner LSTM (which is needed when stacking bidirectional layers), both the forward and backward sequences are merged and passed to the next layer. Verify the expected shape from the Keras model summary before constructing manual inputs.