GRU

namespace: Rindow\NeuralNetworks\Layer
classname: GRU

Gated Recurrent Unit RNN Layer.

Methods

constructor

$builer->GRU(
    int $units,
    array $input_shape=null,
    string|object $activation='tanh',
    string|object $recurrent_activation='sigmoid',
    bool $use_bias=true,
    string|callable $kernel_initializer='glorot_uniform',
    string|callable $recurrent_initializer='orthogonal',
    string|callable $bias_initializer='zeros',
    bool $return_sequences=false,
    bool $return_state=false,
    bool $go_backwards=false,
    bool $stateful=false,
    bool $reset_after=true,
    string $name=null,
)

You can create a GRU layer instances with the Layer Builder.

Arguments

units: Dimensionality of the output space.

Options

input_shape: Tell the first layer the shape of the input data. In input_shape, the batch dimension is not included.
activation: Activation function. Default is ‘tanh’.
recurrent_activation: Recurrent Activation function. Default is ‘sigmoid’.
use_bias: Use bias. Default is true.
kernel_initializer: Default is ‘glorot_uniform’.
recurrent_initializer: Default is ‘orthogonal’.
bias_initializer: Default is ‘zeros’.
return_sequences: Whether to return the last output in the output sequence, or the full sequence. Default is false.
return_state: Whether to return the last state in addition to the output. Default is false.
go_backwards: If True, process the input sequence backwards and return the reversed sequence. Default is false.
stateful: N/A.
reset_after: whether to apply reset gate after or before matrix multiplication. Default is true.

forward

public function forward(
    NDArray $inputs,
    Variable|bool $training,
    array $initialStates=null,
    ?NDArray $mask=null,
) : Variable|array

Arguments

inputs: A 3D NDArray with shape (batch, timesteps, feature).
training: When training, it is true.
initialStates: List of initial state. Number of the state is one. Shape of the state is [batch, units]. When it have no state, give a null.
mask: A boolean tensor with shape (batch, timesteps) that indicates whether specific timesteps should be masked (optional). Individual true entries indicate that the corresponding timestep should be utilized, while false entries indicate that the corresponding timestep should be ignored. The default is null.

Input shape

3D array of shape.

[batchsize, timesteps, features]

Output shape

If “return_state” is true then it returns the list of [outputs, states]. If “return_state” is false then it returns just outputs.

If “return_sequences” is false then the shape of outputs is 2D [batchsize, units]. If “return_sequences” is true then the shape of outputs is 3D [batchsize, timesteps, units].

States shape

The states is list of state. Shape of it is 2D [batchsize, units].

$rnn = $builder->layers()->GRU($units=256,
    return_sequences:true,return_state:true);
....
$inputs = $mo->ones([64,32,128]);
$initialStates = [$mo->ones([64,256])];
....
[$outputs,$states] = $rnn->forward($inputs,true,$initialStates);
# $outputs->shape() : [64,32,256]
# $states[0]->shape() : [64,256]

Example of usage

class Foo extends AbstractModel
{
    public function __construct($backend,$builder)
    {
        ...
        $this->rnn = $builder->layers()->GRU($units=256,
            return_sequences:true,return_state:true);
        ....
    }

    protected function call(.....) 
    {
        ...
        [$outputs,$states] = $this->rnn->forward($inputs,$training,$initialStates);
        ...
    }
}

Rindow Neural Networks

GRU