DL4J Feedforward Learner (Pretraining)

This node performs unsupervised pretraining of a feedforward deep learning model. Thereby, the learning procedure can be adjusted using several training methods and parameters, which can be customized in the node dialog. Additionally, the node supplies further methods for regularization, gradient normalization and learning refinements. The learner node automatically adds an output layer to the network configuration, which can be also configured in the node dialog. For pretraining the network architecture needs to contain layers which are can be trained unsupervised. Such layers are for example an RBM or a Autoencoder. Usually, this node is used together with a classification learner node which performs finetuning of the output layer after the network was pretrained. The output of the node is a pretrained deep learning model.

Options

Learning Parameters

Number of Training Iterations

The number of parameter updates that will be done on one batch of input data.

Optimization Algorithm

The type of optimization method to use. The following algorithms are available:

• Line Gradient Descent
• Conjugate Gradient Descent
• Hessian Free
• LBFGS
• Stochastic Gradient Descent

For Line Gradient Descent, Conjugate Gradient Descent, and LBFGS the maximum number of line search iterations can be specified.

Updater

The type of updater to use. These specify how the raw gradients will be modified. If this option is unchecked the node tries to use an updater from a previously trained network if available. If not available the default will be used (NESTEROVS). Some of the updater types may have additional coefficients which can be adjusted. The The following methods are available:

• SGD
• ADAM (ADAM Mean Decay, ADAM Var Decay)
• ADADELTA (RHO)
• NESTEROVS (Momentum, Schedule)
  Nesterovs Schedule:
  Schedule - Schedule for momentum value change during training. This is specified in the following format:
  'iteration':'momentum rate','iteration':'momentum rate' ...
  This creates a map, which maps the iteration to the momentum rate that should be used. E.g. '2:0.8' means that the rate '0.8' should be used in iteration '2'. Leave empty if you do not want to use a schedule.
• ADAGRAD
• RMSPROP (RMS Decay)

An explanation of these methods and their coefficients can be found here.

Random Seed

The seed value which should be used in order to compare training runs. Any Integer number may be used.

Regularization

The L1 and L2 regularization coefficients.

Gradient Normalization

Gradient normalization strategies. These are applied on raw gradients, before the gradients are passed to the updater. An explanation can be found here.

• Renormalize L2 Per Layer
• Renormalize L2 Per Param Type
• ClipElement Wise Absolute Value
• Clip L2 Per Layer
• Clip L2 Per Param Type

For 'ClipElement Wise Absolute Value', 'Clip L2 Per Layer', and 'Clip L2 Per Param Type' you can additionally specify a threshold value.

Global Parameters

Global Learning Rate

The learning rate for the whole network. If not used the learning rate specified in each layer will be used.

Global Drop-Out Rate

The drop-out rate for the whole network. If not used the drop-out rate specified in each layer will be used.

Use Drop-Connect?

Whether to use Drop Connect.

Global Weight Initialization Strategy

The weight initialization strategy to use for the whole network.

Global Bias - Learning Rate

The bias learning rate for the whole network if you want to use a different learning rate for the bias.

Global Bias - Initialization

The value to initialize all biases with.

Data Parameters

Batch Size

The number of examples used for one minibatch.

Epochs

The number of epochs to train the network, hence the number of training runs on the whole data set.

Column Selection

Feature Column Selection

The columns of the input table containing the training data for the network.

Output Layer Parameter

Number of Output Units

The number of output units of the output layer. This value specifies the length of the output vector of the network.

Learning Rate

The learning rate that should be used for this layer.

Weight Initialization Strategy

The strategy which will be used to set the initial weights for this layer.

Loss Function

The type of loss function that should be used for this layer.

Views

Learning Status

Shows information about the current learning run. Has an option for early stopping of training. If training is stopped before the last epoch the model will be saved in the current status.

Workflows

• No workflows found

Further Links

• No further links available

Developers