Ctrlk
Community ForumND4J JavadocDL4J Javadoc
For the complete documentation index, see llms.txt. This page is also available as Markdown.

Keras Import

Importing Keras models into Deeplearning4j — supported features, limitations, and getting started

Keras Model Import

Keras model import provides routines for importing neural network models originally configured and trained using Keras, a popular Python deep learning library.

Once you have imported your model into DL4J, the full production stack is available. DL4J supports import of all Keras model types, most layers, and practically all utility functionality. See Supported Features for a complete list.

Quick Start

The following example creates a simple MLP in Keras, saves it to HDF5, and loads it as a DL4J MultiLayerNetwork in three lines of Java.

Python (save the model)

from keras.models import Sequential
from keras.layers import Dense

model = Sequential()
model.add(Dense(units=64, activation='relu', input_dim=100))
model.add(Dense(units=10, activation='softmax'))
model.compile(loss='categorical_crossentropy', optimizer='sgd', metrics=['accuracy'])

model.save('simple_mlp.h5')

Java (load and run the model)

String simpleMlp = new ClassPathResource("simple_mlp.h5").getFile().getPath();
MultiLayerNetwork model = KerasModelImport.importKerasSequentialModelAndWeights(simpleMlp);

INDArray input = Nd4j.create(256, 100);
INDArray output = model.output(input);

KerasModelImport is the main entry point. It handles all Keras-to-DL4J concept mapping internally; you provide the model file and get back a fully initialized DL4J network.

Maven Dependency

Add the following to your pom.xml:

Replace ${dl4j.version} with the version matching the rest of your DL4J dependencies.

Model Types

Keras has two model construction APIs, and DL4J maps each to an equivalent:

Keras model type
Keras class
DL4J equivalent

Sequential

keras.models.Sequential

MultiLayerNetwork

Functional API

keras.models.Model

ComputationGraph

  • Sequential models are linear stacks of layers with exactly one input and one output. They map cleanly to MultiLayerNetwork.

  • Functional API models can have arbitrary graph topologies: multiple inputs, multiple outputs, shared layers, and skip connections. They map to ComputationGraph.

Saving Formats

Keras can serialize a model in several ways. Each is supported for import:

What is saved
Python call
DL4J import method

Full model (config + weights + training config)

model.save('model.h5')

importKerasSequentialModelAndWeights(path) or importKerasModelAndWeights(path)

Config only (JSON)

model.to_json()

importKerasSequentialConfiguration(jsonPath) or importKerasModelConfiguration(jsonPath)

Weights only

model.save_weights('weights.h5')

Pass both JSON and weights paths to the two-argument import methods

If you only save the configuration without a training configuration (i.e., the model was not compiled in Keras), set enforceTrainingConfig=false when calling the import method. Otherwise an exception is thrown when DL4J tries to read a training config that does not exist.

Popular Supported Models

A growing set of well-known architectures has been validated end-to-end. These include:

  • ResNet50

  • SqueezeNet

  • MobileNet

  • Inception

  • Xception

  • UNET

  • Deep convolutional GANs

  • Wasserstein GANs

The full end-to-end test list is maintained in KerasModelEndToEndTest.java.

Supported Features Summary

Category
Coverage

Core layers

Dense, Dropout, Flatten, Reshape, Merge, Permute, RepeatVector, Lambda, Masking, SpatialDropout1D/2D/3D

Convolutional layers

Conv1D, Conv2D, Conv3D, SeparableConv2D, Conv2DTranspose, Cropping1/2/3D, UpSampling1/2/3D, ZeroPadding1/2/3D, AtrousConv1D/2D

Pooling layers

MaxPooling1/2/3D, AveragePooling1/2/3D, GlobalMaxPooling1/2/3D, GlobalAveragePooling1/2/3D

Recurrent layers

SimpleRNN, LSTM (GRU and ConvLSTM2D not supported)

Embedding layers

Embedding

Normalization layers

BatchNormalization

Advanced activations

LeakyReLU, PReLU, ELU, ThresholdedReLU

Noise layers

GaussianNoise, GaussianDropout, AlphaDropout

Wrappers

Bidirectional (TimeDistributed not supported)

Losses

13 of 14 standard Keras losses (logcosh excluded)

Activations

softmax, elu, selu, softplus, softsign, relu, tanh, sigmoid, hard_sigmoid, linear

Initializers

All 15 standard Keras initializers

Regularizers

l1, l2, l1_l2

Constraints

max_norm, non_neg, unit_norm, min_max_norm

Optimizers

SGD, RMSprop, Adagrad, Adadelta, Adam, Adamax, Nadam (TFOptimizer excluded)

After Import: Saving the DL4J Model

Once your Keras model is imported, save it using DL4J's ModelSerializer for subsequent loads without the overhead of re-importing from Keras:

Troubleshooting

IncompatibleKerasConfigurationException: the model uses a feature not supported by the importer. Check Supported Features to confirm the layer or configuration in question.

UnsupportedKerasConfigurationException: a recognized feature has parameters that DL4J cannot map. Common causes include unknown regularizer types or unsupported layer parameter combinations.

Weights not loaded correctly: verify that you are using a full model save (not weights-only) or that you are providing both the JSON config and weights paths to the split-file import method.

For additional help, visit the DL4J community forums or open an issue on GitHub.

Why Keras Model Import?

Keras is a user-friendly deep learning library written in Python. Its intuitive API makes rapid prototyping easy. However, Python-trained models often need to be deployed in Java or JVM-based production environments.

Keras model import bridges this gap. Data scientists write and train models in Python; the import module brings them into the DL4J production stack without requiring Python at runtime. The imported model can be:

  • Used for inference directly

  • Fine-tuned with additional DL4J training

  • Integrated with DL4J's distributed training (Spark)

  • Deployed via Konduit Serving

Next Steps

PreviousOverview
NextGetting Started

Last updated

Was this helpful?