Local Layers

DL4J equivalents and API reference for Keras locally connected layers — LocallyConnected1D and LocallyConnected2D.

Keras Locally Connected Layer Import

Locally connected layers are like convolutional layers, but do not share weights across spatial locations. Each spatial position has its own distinct set of filters. Support is implemented in the layers/local package.

Support Summary

Keras Layer	DL4J Equivalent	Supported
LocallyConnected1D	LocallyConnected1D	Yes
LocallyConnected2D	LocallyConnected2D	Yes

---

KerasLocallyConnected1D

source

Imports a Keras LocallyConnected1D layer as a DL4J LocallyConnected1D layer.

Constructor

public KerasLocallyConnected1D(Integer kerasVersion) throws UnsupportedKerasConfigurationException

Pass-through constructor. kerasVersion is the major Keras version (1 or 2).

getLocallyConnected1DLayer

public LocallyConnected1D getLocallyConnected1DLayer()

Returns the DL4J LocallyConnected1D layer. Configuration includes:

• filters → number of output filters (nOut)

• kernel_size → filter size

• strides → stride

• padding → "valid" only (same padding is not supported for locally connected)

• activation → per-position activation function

getOutputType

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

Returns the output InputType based on the filter count, kernel size, and stride.

Parameters:

• inputType — array of input InputType objects

setWeights

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

Loads the locally connected weight tensor and bias from the HDF5 file. The weight tensor for LocallyConnected1D has shape [output_length, kernel_size * input_channels, filters]; the importer handles the reshape from Keras's storage format.

Parameters:

• weights — map from parameter name to INDArray

---

KerasLocallyConnected2D

source

Imports a Keras LocallyConnected2D layer as a DL4J LocallyConnected2D layer.

Constructor

public KerasLocallyConnected2D(Integer kerasVersion) throws UnsupportedKerasConfigurationException

getLocallyConnected2DLayer

public LocallyConnected2D getLocallyConnected2DLayer()

Returns the DL4J LocallyConnected2D layer. Configuration includes:

• filters → number of output filters

• kernel_size → 2D filter size (height, width)

• strides → 2D stride

• padding → "valid" only

• activation → activation function

getOutputType

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

setWeights

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

Loads the 2D locally connected weight tensor and bias. The weight tensor has shape [output_rows, output_cols, kernel_rows * kernel_cols * input_channels, filters].

---

Notes

• Locally connected layers have significantly more parameters than equivalent convolutional layers because weights are not shared spatially.

• Only padding="valid" is supported. padding="same" will throw UnsupportedKerasConfigurationException.

• These layers are relatively uncommon in modern architectures; they were more prevalent in early face recognition models (e.g., DeepFace). For most use cases, standard convolutional layers with sufficient depth are preferred.