Embedding Layers

DL4J equivalents and API reference for Keras embedding layers — Embedding mapped to EmbeddingSequenceLayer.

Keras Embedding Layer Import

Embedding layer support is implemented in the layers/embeddings package.

Support Summary

Keras Layer	DL4J Equivalent	Supported
Embedding	EmbeddingSequenceLayer	Yes

---

KerasEmbedding

source

Imports a Keras Embedding layer as a DL4J EmbeddingSequenceLayer. The embedding matrix (shape [vocab_size, embedding_dim]) is loaded from the HDF5 file.

Constructor (no-arg, for unit tests)

public KerasEmbedding() throws UnsupportedKerasConfigurationException

Pass-through constructor used in unit tests.

getEmbeddingLayer

public EmbeddingSequenceLayer getEmbeddingLayer()

Returns the configured DL4J EmbeddingSequenceLayer. Configuration parameters derived from Keras:

• input_dim → vocabulary size

• output_dim → embedding dimension

• input_length → sequence length (optional)

• embeddings_initializer → weight initializer

• embeddings_regularizer → L1/L2 regularization on the embedding matrix

• mask_zero → whether to mask the zero index (affects masking in downstream RNN layers)

getOutputType

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

Returns an InputType.recurrent(embeddingDim) reflecting the sequence output of the embedding layer.

Parameters:

• inputType — array of input InputType objects

Returns: output InputType

getNumParams

public int getNumParams()

Returns 1 (the embedding matrix is a single parameter tensor).

setWeights

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

Loads the embedding matrix from the weight map. In the HDF5 file, Keras stores this as embeddings_0 or a similar key depending on the version.

Parameters:

• weights — map from parameter name to INDArray containing the embedding matrix

---

Keras Embedding Parameters

Keras parameter	DL4J equivalent
input_dim	nIn() (vocabulary size)
output_dim	nOut() (embedding dimension)
input_length	sequence length
embeddings_initializer	weight initializer
embeddings_regularizer	weight regularization
mask_zero	masking configuration
trainable	weight update flag

---

Example

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

model = Sequential()
model.add(Embedding(input_dim=10000, output_dim=128, input_length=50))
model.add(LSTM(64))
model.add(Dense(2, activation='softmax'))
model.compile(optimizer='adam', loss='binary_crossentropy')
model.save('embedding_lstm.h5')

// Java: import and run inference
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("embedding_lstm.h5");

// Input: integer token indices, shape [batch, sequence_length]
// DL4J EmbeddingSequenceLayer expects float representation of integer indices
INDArray input = Nd4j.create(new float[]{1, 5, 3, 42, 0}, new int[]{1, 5});
INDArray output = model.output(input);

---

Notes

• DL4J's EmbeddingSequenceLayer handles sequence inputs natively and is preferred over EmbeddingLayer for Keras Embedding import because it preserves the temporal dimension for downstream RNN layers.

• Sparse embedding updates (Keras embeddings_constraint) are read but DL4J uses dense gradient updates for the embedding matrix.

• If mask_zero=True was used in Keras, downstream LSTM/SimpleRNN layers will apply masking. Verify that the DL4J model handles masking as expected for your data.