Supported Features

Full support matrix for Keras model import — layers, activations, losses, and optimizers

Keras Model Import: Supported Features

This page provides the complete support matrix for Keras model import into DL4J. All mapping is implemented in the deeplearning4j-modelimport module.

• Supported

• Not supported

---

Layers

Mapping of Keras layers to DL4J is implemented in the layers sub-module.

Core Layers

Keras Layer	DL4J Equivalent	Supported
Dense	DenseLayer	Yes
Activation	ActivationLayer	Yes
Dropout	DropoutLayer	Yes
Flatten	CnnToFeedForwardPreProcessor / RnnToFeedForwardPreProcessor	Yes
Reshape	Reshape (via input preprocessor)	Yes
Merge	MergeVertex	Yes
Permute	PermutePreProcessor	Yes
RepeatVector	RepeatVector	Yes
Lambda	SameDiffLambda	Yes
ActivityRegularization	—	No
Masking	MaskZeroLayer	Yes
SpatialDropout1D	DropoutLayer (spatial)	Yes
SpatialDropout2D	DropoutLayer (spatial)	Yes
SpatialDropout3D	DropoutLayer (spatial)	Yes

Convolutional Layers

Keras Layer	DL4J Equivalent	Supported
Conv1D	Convolution1DLayer	Yes
Conv2D	ConvolutionLayer	Yes
Conv3D	ConvolutionLayer (3D)	Yes
AtrousConvolution1D	Convolution1DLayer (with dilation)	Yes
AtrousConvolution2D	ConvolutionLayer (with dilation)	Yes
SeparableConv1D	—	No
SeparableConv2D	SeparableConvolution2D	Yes
DepthwiseConv2D	DepthwiseConvolution2D	Yes
Conv2DTranspose	Deconvolution2D	Yes
Conv3DTranspose	—	No
Cropping1D	Cropping1D	Yes
Cropping2D	Cropping2D	Yes
Cropping3D	Cropping3D	Yes
UpSampling1D	Upsampling1D	Yes
UpSampling2D	Upsampling2D	Yes
UpSampling3D	Upsampling3D	Yes
ZeroPadding1D	ZeroPadding1DLayer	Yes
ZeroPadding2D	ZeroPaddingLayer	Yes
ZeroPadding3D	ZeroPadding3DLayer	Yes

Pooling Layers

Keras Layer	DL4J Equivalent	Supported
MaxPooling1D	Subsampling1DLayer (MAX)	Yes
MaxPooling2D	SubsamplingLayer (MAX)	Yes
MaxPooling3D	Subsampling3DLayer (MAX)	Yes
AveragePooling1D	Subsampling1DLayer (AVG)	Yes
AveragePooling2D	SubsamplingLayer (AVG)	Yes
AveragePooling3D	Subsampling3DLayer (AVG)	Yes
GlobalMaxPooling1D	GlobalPoolingLayer (MAX)	Yes
GlobalMaxPooling2D	GlobalPoolingLayer (MAX)	Yes
GlobalMaxPooling3D	GlobalPoolingLayer (MAX)	Yes
GlobalAveragePooling1D	GlobalPoolingLayer (AVG)	Yes
GlobalAveragePooling2D	GlobalPoolingLayer (AVG)	Yes
GlobalAveragePooling3D	GlobalPoolingLayer (AVG)	Yes

Locally-Connected Layers

Keras Layer	DL4J Equivalent	Supported
LocallyConnected1D	LocallyConnected1D	Yes
LocallyConnected2D	LocallyConnected2D	Yes

Recurrent Layers

Keras Layer	DL4J Equivalent	Supported
SimpleRNN	SimpleRnn	Yes
GRU	—	No
LSTM	LSTM	Yes
ConvLSTM2D	—	No

Embedding Layers

Keras Layer	DL4J Equivalent	Supported
Embedding	EmbeddingSequenceLayer	Yes

Merge Layers

Keras Layer	DL4J Equivalent	Supported
Add / add	MergeVertex (add)	Yes
Multiply / multiply	MergeVertex (mul)	Yes
Subtract / subtract	MergeVertex (sub)	Yes
Average / average	MergeVertex (avg)	Yes
Maximum / maximum	MergeVertex (max)	Yes
Concatenate / concatenate	MergeVertex (concat)	Yes
Dot / dot	—	No

Advanced Activation Layers

Keras Layer	DL4J Equivalent	Supported
LeakyReLU	ActivationLayer (LeakyReLU)	Yes
PReLU	PReLULayer	Yes
ELU	ActivationLayer (ELU)	Yes
ThresholdedReLU	ActivationLayer (ThresholdedReLU)	Yes

Normalization Layers

Keras Layer	DL4J Equivalent	Supported
BatchNormalization	BatchNormalization	Yes

Noise Layers

Keras Layer	DL4J Equivalent	Supported
GaussianNoise	DropoutLayer (GaussianNoise)	Yes
GaussianDropout	DropoutLayer (GaussianDropout)	Yes
AlphaDropout	DropoutLayer (AlphaDropout)	Yes

Layer Wrappers

Keras Wrapper	DL4J Equivalent	Supported
TimeDistributed	—	No
Bidirectional	Bidirectional	Yes

---

Losses

Source: KerasLossUtils

Keras Loss	DL4J Equivalent	Supported
mean_squared_error	MSE	Yes
mean_absolute_error	MAE	Yes
mean_absolute_percentage_error	MAPE	Yes
mean_squared_logarithmic_error	MSLE	Yes
squared_hinge	SquaredHinge	Yes
hinge	Hinge	Yes
categorical_hinge	CategoricalHinge	Yes
logcosh	—	No
categorical_crossentropy	CategoricalCrossEntropy	Yes
sparse_categorical_crossentropy	SparseMCXENT	Yes
binary_crossentropy	BinaryCrossEntropy	Yes
kullback_leibler_divergence	KullbackLeiblerDivergence	Yes
poisson	Poisson	Yes
cosine_proximity	CosineSimilarity	Yes

---

Activations

Source: KerasActivationUtils

All standard Keras activations are supported:

Keras Activation	DL4J Equivalent	Supported
softmax	Softmax	Yes
elu	ELU	Yes
selu	SELU	Yes
softplus	Softplus	Yes
softsign	Softsign	Yes
relu	ReLU	Yes
tanh	Tanh	Yes
sigmoid	Sigmoid	Yes
hard_sigmoid	HardSigmoid	Yes
linear	Identity	Yes

---

Initializers

Source: KerasInitilizationUtils

All standard Keras weight initializers are supported:

Keras Initializer	DL4J Equivalent	Supported
Zeros	ZeroInitScheme	Yes
Ones	OneInitScheme	Yes
Constant	ConstantDistribution	Yes
RandomNormal	NormalDistribution	Yes
RandomUniform	UniformDistribution	Yes
TruncatedNormal	TruncatedNormalDistribution	Yes
VarianceScaling	VarianceScalingInitScheme	Yes
Orthogonal	OrthogonalInitScheme	Yes
Identity	IdentityInitScheme	Yes
lecun_uniform	WeightInitLecunUniform	Yes
lecun_normal	WeightInitLecunNormal	Yes
glorot_normal	GlorotNormal (Xavier)	Yes
glorot_uniform	GlorotUniform (Xavier)	Yes
he_normal	HeNormal	Yes
he_uniform	HeUniform	Yes

---

Regularizers

Source: KerasRegularizerUtils

All standard Keras regularizers are supported:

Keras Regularizer	DL4J Equivalent	Supported
l1	L1Regularization	Yes
l2	L2Regularization	Yes
l1_l2	L1L2Regularization	Yes

---

Constraints

Source: KerasConstraintUtils

All standard Keras constraints are supported:

Keras Constraint	DL4J Equivalent	Supported
max_norm	MaxNormConstraint	Yes
non_neg	NonNegativeConstraint	Yes
unit_norm	UnitNormConstraint	Yes
min_max_norm	MinMaxNormConstraint	Yes

---

Optimizers

Source: KerasOptimizerUtils

All standard Keras optimizers are supported. TFOptimizer (a TensorFlow-specific optimizer wrapper) is not:

Keras Optimizer	DL4J Equivalent	Supported
SGD	Sgd	Yes
RMSprop	RmsProp	Yes
Adagrad	AdaGrad	Yes
Adadelta	AdaDelta	Yes
Adam	Adam	Yes
Adamax	AdaMax	Yes
Nadam	Nadam	Yes
TFOptimizer	—	No

---

Notes on Partial Support

• GRU: not currently supported. The Keras GRU has a slightly different recurrent formulation from DL4J's. Consider replacing with LSTM for import, or implement a custom layer mapper.

• TimeDistributed wrapper: not supported. In many cases, you can restructure your model to achieve the same effect with 1D convolutions or RNNs directly.

• ConvLSTM2D: not supported. No equivalent exists in DL4J core at this time.

• Dot merge layer: the dot product merge mode is not supported. Concatenate followed by a Dense layer is a workable alternative in many cases.

• logcosh loss: not implemented. mean_squared_error is a reasonable substitute for smooth losses.

• ActivityRegularization: this layer type is not supported. Apply regularization directly in the layer configuration instead.

• Custom TensorFlow optimizers (TFOptimizer): cannot be imported. Use a standard Keras optimizer.

---

Keras Version Compatibility

DL4J model import supports both Keras 1.x and Keras 2.x. The importer detects the Keras version from the HDF5 metadata and adjusts config key names accordingly. Both TensorFlow and Theano backends produce compatible HDF5 files.