Regularizers

Mapping of Keras regularizers to DL4J regularization for model import in Eclipse Deeplearning4j 1.0.0-M2.1.

Keras allows L1 and L2 regularization penalties to be applied to layer kernels, biases, and activations. DL4J supports all standard Keras regularizers and maps them to equivalent regularization parameters in the layer configuration.

The mapping is implemented in KerasRegularizerUtils.

Supported Regularizers

Keras regularizer	DL4J equivalent	Description
l1	L1Regularization	L1 (Lasso) weight penalty
l2	L2Regularization	L2 (Ridge) weight penalty
l1_l2	L1L2Regularization	Combined L1 and L2 penalty (Elastic Net)

Regularizer Descriptions

l1

Adds the sum of the absolute values of the weights to the loss function, scaled by the regularization coefficient. L1 regularization encourages sparsity — many weights will be driven to exactly zero.

# Keras
from keras.regularizers import l1
layer = Dense(64, kernel_regularizer=l1(0.01))

// DL4J equivalent (applied via layer configuration)
.l1(0.01)

Keras parameter:

Parameter	Default	Description
l	0.01	Regularization coefficient

l2

Adds the sum of the squared weights to the loss function, scaled by the regularization coefficient. L2 regularization (weight decay) penalises large weights and generally produces smoother, more generalisable models.

# Keras
from keras.regularizers import l2
layer = Dense(64, kernel_regularizer=l2(0.01))

// DL4J equivalent
.l2(0.01)

Keras parameter:

Parameter	Default	Description
l	0.01	Regularization coefficient

l1_l2

Combines L1 and L2 penalties (Elastic Net regularization). Both coefficients can be set independently.

# Keras
from keras.regularizers import l1_l2
layer = Dense(64, kernel_regularizer=l1_l2(l1=0.01, l2=0.001))

// DL4J equivalent
.l1(0.01).l2(0.001)

Keras parameters:

Parameter	Default	Description
l1	0.01	L1 regularization coefficient
l2	0.01	L2 regularization coefficient

Regularization Targets

In Keras, regularization can be applied separately to kernels, biases, and activations:

Dense(64,
      kernel_regularizer=l2(0.01),
      bias_regularizer=l1(0.001),
      activity_regularizer=l2(0.0001))

DL4J maps kernel_regularizer and bias_regularizer directly to the layer's l1/l2 settings. activity_regularizer (applied to the layer output rather than the weights) is not currently supported during import.

Usage Example

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

// Regularizer coefficients are automatically read from the Keras config
MultiLayerNetwork model = KerasModelImport.importKerasSequentialModelAndWeights("model.h5");

Notes

• Regularization affects the training loss, not inference. For inference-only import, the regularizer config is loaded but has no practical effect.

• Custom Python regularizer classes cannot be imported.

• activity_regularizer is not supported; a warning is logged if it is present.