Noise Layers

DL4J equivalents and API reference for Keras noise layers — GaussianNoise, GaussianDropout, and AlphaDropout.

Keras Noise Layer Import

Noise layers add stochastic regularization during training. Support is implemented in the layers/noise package.

Support Summary

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

---

KerasGaussianNoise

source

Imports a Keras GaussianNoise layer as a DL4J DropoutLayer with GaussianNoise applied to the activations during training. The standard deviation stddev is read from the Keras configuration.

output = input + N(0, stddev)  during training
output = input                 during inference

Constructor

public KerasGaussianNoise(Integer kerasVersion) throws UnsupportedKerasConfigurationException

getGaussianNoiseLayer

public DropoutLayer getGaussianNoiseLayer()

Returns the DL4J DropoutLayer configured with Gaussian noise.

getOutputType

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

---

KerasGaussianDropout

source

Imports a Keras GaussianDropout layer as a DL4J DropoutLayer with GaussianDropout. This multiplies activations by a Gaussian noise factor with mean 1 and variance derived from the dropout rate rate:

stddev = sqrt(rate / (1 - rate))
output = input * N(1, stddev)

Constructor

public KerasGaussianDropout(Integer kerasVersion) throws UnsupportedKerasConfigurationException

getGaussianDropoutLayer

public DropoutLayer getGaussianDropoutLayer()

Returns the DL4J DropoutLayer configured with Gaussian dropout.

getOutputType

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

---

KerasAlphaDropout

source

Imports a Keras AlphaDropout layer as a DL4J DropoutLayer with AlphaDropout. AlphaDropout is designed for use with SELU activations; it preserves the mean and variance of activations after applying dropout by setting dropped-out units to the negative saturation value of SELU.

Constructor

public KerasAlphaDropout(Integer kerasVersion) throws UnsupportedKerasConfigurationException

getAlphaDropoutLayer

public DropoutLayer getAlphaDropoutLayer()

Returns the DL4J DropoutLayer configured with AlphaDropout.

getOutputType

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

---

Notes

• All noise layers are applied only during training. At inference time they are identity functions. DL4J handles this through the LayerWorkspaceMgr training/test mode flag.

• AlphaDropout is intended to be used with SELU activations and the lecun_normal initializer. If your Keras model used this combination, verify that the DL4J model is also initialized with lecun_normal (which it will be if weights are loaded from HDF5).

• Noise layers have no trainable parameters; setWeights is a no-op for these classes.