Backend

DL4J Keras model import is backend-agnostic — models trained with TensorFlow, Theano, or CNTK backends can all be imported into Eclipse Deeplearning4j 1.0.0-M2.1.

DL4J Keras model import is backend agnostic. Regardless of which backend was used to train the Keras model — TensorFlow, Theano, or CNTK — the model can be imported into DL4J. The importer reads the serialised HDF5 file and interprets the layer configurations and weights without relying on the original backend being present.

Supported Backends

Keras backend	Import supported	Notes
TensorFlow	Yes	The most common backend; widest test coverage
Theano	Yes	Supported; some weight axis orderings differ
CNTK	Yes	Supported; limited production usage

Theano vs. TensorFlow Weight Ordering

The primary difference between backends relates to weight axis ordering for convolutional layers:

• TensorFlow (channels-last, NHWC): filter shape is (height, width, in_channels, out_channels)

• Theano (channels-first, NCHW): filter shape is (out_channels, in_channels, height, width)

DL4J automatically detects the backend from the backend field stored in the Keras model config and transposes convolutional weights as needed during import. You do not need to configure this manually.

Detecting the Backend from an HDF5 File

The backend is stored in the Keras model config JSON inside the HDF5 file:

import h5py, json

with h5py.File("model.h5", "r") as f:
    config = json.loads(f.attrs["model_config"])
    backend = config.get("backend", "tensorflow")
    print("Backend:", backend)

Keras Version Compatibility

DL4J model import supports both Keras 1 and Keras 2 model formats. The major version is detected automatically from the HDF5 file metadata.

Keras major version	Supported	Notes
Keras 1.x	Yes	Legacy format; field names differ slightly
Keras 2.x	Yes	Current recommended format
tf.keras (TF 2.x)	Yes	Saved with model.save() in HDF5 format

Backend Configuration in DL4J

DL4J itself uses ND4J as its numerical backend, which can be accelerated with either CPU (OpenBLAS/MKL) or GPU (CUDA) via its own backend system. This is independent of the Keras training backend.

To configure ND4J backends, set the appropriate dependency in your pom.xml:

<!-- CPU backend -->
<dependency>
  <groupId>org.nd4j</groupId>
  <artifactId>nd4j-native-platform</artifactId>
  <version>1.0.0-M2.1</version>
</dependency>

<!-- GPU (CUDA) backend -->
<dependency>
  <groupId>org.nd4j</groupId>
  <artifactId>nd4j-cuda-11.6-platform</artifactId>
  <version>1.0.0-M2.1</version>
</dependency>

Saving a Keras Model for Import

Always save Keras models in the HDF5 format using model.save() to ensure both the architecture and weights are available:

# Save the full model (architecture + weights + optimizer config)
model.save("model.h5")

# For inference-only import, weights-only files also work:
model.save_weights("weights.h5")
# Architecture can be saved separately:
with open("architecture.json", "w") as f:
    f.write(model.to_json())

Import Examples

import org.deeplearning4j.nn.modelimport.keras.KerasModelImport;
import org.deeplearning4j.nn.multilayer.MultiLayerNetwork;
import org.deeplearning4j.nn.graph.ComputationGraph;

// Import a Sequential model (full model with weights)
MultiLayerNetwork seqModel = KerasModelImport
    .importKerasSequentialModelAndWeights("sequential_model.h5");

// Import a Functional API model (full model with weights)
ComputationGraph funcModel = KerasModelImport
    .importKerasFunctionalModelAndWeights("functional_model.h5");

// Import architecture only, then load weights separately
MultiLayerNetwork modelNoWeights = KerasModelImport
    .importKerasSequentialModelAndWeights("model.h5", false);

Notes

• Models saved in the TensorFlow SavedModel format (a directory rather than a single .h5 file) are not supported by the Keras import module. Use DL4J's TensorFlow import instead.

• Models saved with tf.saved_model.save() or model.save('dir/') in TF2 use the SavedModel format by default. Pass save_format='h5' to force HDF5 output.

• The keras.json backend configuration file on the training machine does not need to be present on the DL4J inference machine.