RBMs are no longer supported as of version 0.9.x. They are no longer best-in-class for most machine learning problems.

[source]

Autoencoder layer. Adds noise to input and learn a reconstruction function.

corruptionLevel

Level of corruption - 0.0 (none) to 1.0 (all values corrupted)

sparsity

Autoencoder sparity parameter

  • param sparsity Sparsity

BernoulliReconstructionDistribution

[source]

Bernoulli reconstruction distribution for variational autoencoder.Outputs are modelled by a Bernoulli distribution - i.e., the Bernoulli distribution should be used for binary data (allvalues 0 or 1); the VAE models the probability of the output being 0 or 1.Consequently, the sigmoid activation function should be used to bound activations to the range of 0 to 1. Activationfunctions that do not produce outputs in the range of 0 to 1 (including relu, tanh, and many others) should be avoided.

hasLossFunction
  1. public boolean hasLossFunction()

Create a BernoulliReconstructionDistribution with the default Sigmoid activation function

[source]

CompositeReconstructionDistribution is a reconstruction distribution built from multiple other ReconstructionDistributioninstances.The typical use is to combine for example continuous and binary data in the same model, or to combine differentdistributions for continuous variables. In either case, this class allows users to model (for example) the first 10 values

addDistribution
  1. public Builder addDistribution(int distributionSize, ReconstructionDistribution distribution)

Add another distribution to the composite distribution. This will add the distribution for the next ‘distributionSize’values, after any previously added.For example, calling addDistribution(10, X) once will result in values 0 to 9 (inclusive) being modelledby the specified distribution X. Calling addDistribution(10, Y) after that will result in values 10 to 19 (inclusive)being modelled by distribution Y.

  • param distributionSize Number of values to model with the specified distribution
  • param distribution Distribution to model data with

ExponentialReconstructionDistribution

Exponential reconstruction distribution.Supports data in range [0,infinity)

This means that an input from the decoder of gamma = 0 gives lambda = 1which corresponds to a mean value for the expontial distribution of 1/lambda = 1

Regarding the choice of activation function: the parameterization above supports gamma in the range (-infinity,infinity)therefore a symmetric activation function such as “identity” or perhaps “tanh” is preferred.

hasLossFunction
  1. public boolean hasLossFunction()
  • deprecated Use {- link #ExponentialReconstructionDistribution(Activation)}

Gaussian reconstruction distribution for variational autoencoder.Outputs are modelled by a Gaussian distribution, with the mean and variances (diagonal covariance matrix) for eachoutput determined by the network forward pass.

Specifically, the GaussianReconstructionDistribution models mean and log(stdev^2). This parameterization gives log(1) = 0,and inputs can be in range (-infinity,infinity). Other parameterizations for variance are of course possible but may beproblematic with respect to the average pre-activation function values and activation function ranges.For activation functions, identity and perhaps tanh are typical - though tanh (unlike identity) implies a minimum/maximumpossible value for mean and log variance. Asymmetric activation functions such as sigmoid or relu should be avoided.

hasLossFunction
  1. public boolean hasLossFunction()

Create a GaussianReconstructionDistribution with the default identity activation function.

LossFunctionWrapper

[source]

LossFunctionWrapper allows training of a VAE model with a standard (possibly deterministic) neural network loss function

Note: most functionality is supported, but clearly reconstruction log probability cannot be calculated when usingLossFunctionWrapper, as ILossFunction instances do not have either (a) a probabilistic interpretation, or (b) ameans of calculating the negative log probability.

VariationalAutoencoder

Variational Autoencoder layer

See: Kingma & Welling, 2013: Auto-Encoding Variational Bayes - https://arxiv.org/abs/1312.6114

A note on scores during pretraining: This implementation minimizes the negative of the variational lower boundobjective as described in Kingma & Welling; the mathematics in that paper is based on maximization of the variationallower bound instead. Thus, scores reported during pretraining in DL4J are the negative of the variational lower boundequation in the paper. The backpropagation and learning procedure is otherwise as described there.

encoderLayerSizes

Size of the encoder layers, in units. Each encoder layer is functionally equivalent to a {- linkorg.deeplearning4j.nn.conf.layers.DenseLayer}. Typically the number and size of the decoder layers (set via{- link #decoderLayerSizes(int…)} is similar to the encoder layers.

setEncoderLayerSizes

Size of the encoder layers, in units. Each encoder layer is functionally equivalent to a {- linkorg.deeplearning4j.nn.conf.layers.DenseLayer}. Typically the number and size of the decoder layers (set via{- link #decoderLayerSizes(int…)} is similar to the encoder layers.

  • param encoderLayerSizes Size of each encoder layer in the variational autoencoder
decoderLayerSizes

    Size of the decoder layers, in units. Each decoder layer is functionally equivalent to a {- linkorg.deeplearning4j.nn.conf.layers.DenseLayer}. Typically the number and size of the decoder layers is similarto the encoder layers (set via {- link #encoderLayerSizes(int…)}.

    • param decoderLayerSizes Size of each deccoder layer in the variational autoencoder
    setDecoderLayerSizes
    1. public void setDecoderLayerSizes(int... decoderLayerSizes)

    Size of the decoder layers, in units. Each decoder layer is functionally equivalent to a {- linkorg.deeplearning4j.nn.conf.layers.DenseLayer}. Typically the number and size of the decoder layers is similarto the encoder layers (set via {- link #encoderLayerSizes(int…)}.

    • param decoderLayerSizes Size of each deccoder layer in the variational autoencoder
    reconstructionDistribution
    1. public Builder reconstructionDistribution(ReconstructionDistribution distribution)

    The reconstruction distribution for the data given the hidden state - i.e., P(data|Z). This should beselected carefully based on the type of data being modelled. For example: - {- linkGaussianReconstructionDistribution} + {identity or tanh} for real-valued (Gaussian) data - {- linkBernoulliReconstructionDistribution} + sigmoid for binary-valued (0 or 1) data

    • param distribution Reconstruction distribution
    lossFunction
    1. public Builder lossFunction(IActivation outputActivationFn, LossFunctions.LossFunction lossFunction)

    Configure the VAE to use the specified loss function for the reconstruction, instead of aReconstructionDistribution. Note that this is NOT following the standard VAE design (as per Kingma &Welling), which assumes a probabilistic output - i.e., some p(x|z). It is however a valid networkconfiguration, allowing for optimization of more traditional objectives such as mean squared error. Note:clearly, setting the loss function here will override any previously set recontruction distribution

    • param outputActivationFn Activation function for the output/reconstruction
    • param lossFunction Loss function to use
    lossFunction

    Configure the VAE to use the specified loss function for the reconstruction, instead of aReconstructionDistribution. Note that this is NOT following the standard VAE design (as per Kingma &Welling), which assumes a probabilistic output - i.e., some p(x|z). It is however a valid networkconfiguration, allowing for optimization of more traditional objectives such as mean squared error. Note:clearly, setting the loss function here will override any previously set recontruction distribution

    • param outputActivationFn Activation function for the output/reconstruction
    • param lossFunction Loss function to use
    lossFunction

    Configure the VAE to use the specified loss function for the reconstruction, instead of aReconstructionDistribution. Note that this is NOT following the standard VAE design (as per Kingma &Welling), which assumes a probabilistic output - i.e., some p(x|z). It is however a valid networkconfiguration, allowing for optimization of more traditional objectives such as mean squared error. Note:clearly, setting the loss function here will override any previously set recontruction distribution

    • param outputActivationFn Activation function for the output/reconstruction
    • param lossFunction Loss function to use
    pzxActivationFn
    1. public Builder pzxActivationFn(IActivation activationFunction)

    Activation function for the input to P(z|data). Care should be taken with this, as some activationfunctions (relu, etc) are not suitable due to being bounded in range [0,infinity).

    • param activationFunction Activation function for p(zx)
    pzxActivationFunction
    1. public Builder pzxActivationFunction(Activation activation)

    Activation function for the input to P(z|data). Care should be taken with this, as some activationfunctions (relu, etc) are not suitable due to being bounded in range [0,infinity).

    • param activation Activation function for p(zx)
    nOut
    1. public Builder nOut(int nOut)

    Set the size of the VAE state Z. This is the output size during standard forward pass, and the size of thedistribution P(Z|data) during pretraining.

    • param nOut Size of P(Zdata) and output size
    numSamples

      Set the number of samples per data point (from VAE state Z) used when doing pretraining. Default value: 1.

      • param numSamples Number of samples per data point for pretraining