Keras GRU Layer

Gated recurrent unit as introduced by Cho et al. There are two variants. The default one is based on 1406.1078v3 and has reset gate applied to hidden state before matrix multiplication. The other one is based on original 1406.1078v1 and has the order reversed. Corresponds to the GRU Keras layer .

Options

Name prefix

The name prefix of the layer. The prefix is complemented by an index suffix to obtain a unique layer name. If this option is unchecked, the name prefix is derived from the layer type.

Input tensor

The tensor to use as input for the layer.

Hidden state tensor

The tensor to use as initial hidden state in case the corresponding port is connected.

Units

Dimensionality of the output space.

Activation

The activation function to use on the input transformations.

Recurrent activation

The activation function to use for the recurrent step.

Use bias

If checked, a bias vector will be used.

Return sequences

Whether to return the last output in the output sequence or the full output sequence. If selected the output will have shape [time, units] otherwise the output will have shape [units].

Return state

Whether to return the hidden state in addition to the layer output. If selected the layer returns two tensors, the normal output and the hidden state of the layer. If sequences are returned, this also applies to the hidden state.

Dropout

Fraction of the units to drop for the linear transformation of the input.

Recurrent dropout

Fraction of the units to drop for the linear transformation of the recurrent state.

Go backwards

Whether to go backwards in time i.e. read the input sequence backwards.

Unroll

Whether to unroll the network i.e. convert it in a feed-forward network that reuses the layer's weights for each timestep. Unrolling can speed up an RNN but it's more memory-expensive and only suitable for short sequences. If the layer is not unrolled, a symbolic loop is used.

Implementation

Mode 1 will structure its operations as a larger number of smaller dot products and additions, whereas mode 2 will batch them into fewer, larger operations. These modes will have different performance profiles on different hardware and for different applications.

Reset after

GRU convention (whether to apply reset gate after or before matrix multiplication).

Initializers

Kernel initializer

Initializer for the weight matrix used for the linear transformations of the input. See initializers for details on the available initializers.

Recurrent initializer

Initializer for the weight matrix used for the linear transformation of the recurrent connection. See initializers for details on the available initializers.

Bias initializer

Initializer for the bias vector (if a bias is used). See initializers for details on the available initializers.

Regularizers

Kernel regularizer

Regularizer function applied to the weight matrix. See regularizers for details on the available regularizers.

Recurrent regularizer

Regularizer function applied to the weight matrix for the recurrent connection. See regularizers for details on the available regularizers.

Bias regularizer

Regularizer function applied to the bias vector. See regularizers for details on the available regularizers.

Activity regularizer

Regularizer function applied to the output of the layer i.e. its activation. See regularizers for details on the available regularizers.

Constraints

Kernel constraint

Constraint on the weight matrix for the input connection. See constraints for details on the available constraints.

Recurrent constraint

Constraint on the weight matrix for the recurrent connection. See constraints for details on the available constraints.

Bias constraint

Constraint on the bias vector. See constraints for details on the available constraints.

Views

This node has no views

Workflows

• No workflows found

Links

• KNIME Deep Learning Keras Integration

  www.knime.com/deeplearning/keras

Developers