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.
layer_gru(
object,
units,
activation = "tanh",
recurrent_activation = "sigmoid",
use_bias = TRUE,
return_sequences = FALSE,
return_state = FALSE,
go_backwards = FALSE,
stateful = FALSE,
unroll = FALSE,
time_major = FALSE,
reset_after = TRUE,
kernel_initializer = "glorot_uniform",
recurrent_initializer = "orthogonal",
bias_initializer = "zeros",
kernel_regularizer = NULL,
recurrent_regularizer = NULL,
bias_regularizer = NULL,
activity_regularizer = NULL,
kernel_constraint = NULL,
recurrent_constraint = NULL,
bias_constraint = NULL,
dropout = 0,
recurrent_dropout = 0,
...
)What to compose the new Layer instance with. Typically a
Sequential model or a Tensor (e.g., as returned by layer_input()).
The return value depends on object. If object is:
missing or NULL, the Layer instance is returned.
a Sequential model, the model with an additional layer is returned.
a Tensor, the output tensor from layer_instance(object) is returned.
Positive integer, dimensionality of the output space.
Activation function to use. Default: hyperbolic tangent
(tanh). If you pass NULL, no activation is applied
(ie. "linear" activation: a(x) = x).
Activation function to use for the recurrent step.
Boolean, whether the layer uses a bias vector.
Boolean. Whether to return the last output in the output sequence, or the full sequence.
Boolean (default FALSE). Whether to return the last state in addition to the output.
Boolean (default FALSE). If TRUE, process the input sequence backwards and return the reversed sequence.
Boolean (default FALSE). If TRUE, the last state for each sample at index i in a batch will be used as initial state for the sample of index i in the following batch.
Boolean (default FALSE). If TRUE, the network will be unrolled, else a symbolic loop will be used. Unrolling can speed-up a RNN, although it tends to be more memory-intensive. Unrolling is only suitable for short sequences.
If True, the inputs and outputs will be in shape
[timesteps, batch, feature], whereas in the False case, it will be
[batch, timesteps, feature]. Using time_major = TRUE is a bit more
efficient because it avoids transposes at the beginning and end of the RNN
calculation. However, most TensorFlow data is batch-major, so by default
this function accepts input and emits output in batch-major form.
GRU convention (whether to apply reset gate after or before matrix multiplication). FALSE = "before" (default), TRUE = "after" (CuDNN compatible).
Initializer for the kernel weights matrix, used
for the linear transformation of the inputs.
Initializer for the recurrent_kernel weights
matrix, used for the linear transformation of the recurrent state.
Initializer for the bias vector.
Regularizer function applied to the kernel
weights matrix.
Regularizer function applied to the
recurrent_kernel weights matrix.
Regularizer function applied to the bias vector.
Regularizer function applied to the output of the layer (its "activation")..
Constraint function applied to the kernel weights
matrix.
Constraint function applied to the
recurrent_kernel weights matrix.
Constraint function applied to the bias vector.
Float between 0 and 1. Fraction of the units to drop for the linear transformation of the inputs.
Float between 0 and 1. Fraction of the units to drop for the linear transformation of the recurrent state.
Standard Layer args.
N-D tensor with shape (batch_size, timesteps, ...),
or (timesteps, batch_size, ...) when time_major = TRUE.
if return_state: a list of tensors. The first tensor is
the output. The remaining tensors are the last states,
each with shape (batch_size, state_size), where state_size
could be a high dimension tensor shape.
if return_sequences: N-D tensor with shape [batch_size, timesteps, output_size], where output_size could be a high dimension tensor shape, or
[timesteps, batch_size, output_size] when time_major is TRUE
else, N-D tensor with shape [batch_size, output_size], where
output_size could be a high dimension tensor shape.
This layer supports masking for input data with a variable number of
timesteps. To introduce masks to your data, use
layer_embedding() with the mask_zero parameter set to TRUE.
You can set RNN layers to be 'stateful', which means that the states computed for the samples in one batch will be reused as initial states for the samples in the next batch. This assumes a one-to-one mapping between samples in different successive batches.
For intuition behind statefulness, there is a helpful blog post here: https://philipperemy.github.io/keras-stateful-lstm/
To enable statefulness:
Specify stateful = TRUE in the layer constructor.
Specify a fixed batch size for your model. For sequential models,
pass batch_input_shape = list(...) to the first layer in your model.
For functional models with 1 or more Input layers, pass
batch_shape = list(...) to all the first layers in your model.
This is the expected shape of your inputs including the batch size.
It should be a list of integers, e.g. list(32, 10, 100).
For dimensions which can vary (are not known ahead of time),
use NULL in place of an integer, e.g. list(32, NULL, NULL).
Specify shuffle = FALSE when calling fit().
To reset the states of your model, call layer$reset_states() on either
a specific layer, or on your entire model.
You can specify the initial state of RNN layers symbolically by calling them
with the keyword argument initial_state. The value of initial_state should
be a tensor or list of tensors representing the initial state of the RNN
layer.
You can specify the initial state of RNN layers numerically by calling
reset_states with the named argument states. The value of states should
be an array or list of arrays representing the initial state of the RNN
layer.
You can pass "external" constants to the cell using the constants named
argument of RNN$__call__ (as well as RNN$call) method. This requires that the
cell$call method accepts the same keyword argument constants. Such constants
can be used to condition the cell transformation on additional static inputs
(not changing over time), a.k.a. an attention mechanism.
The second variant is compatible with CuDNNGRU (GPU-only) and allows
inference on CPU. Thus it has separate biases for kernel and
recurrent_kernel. Use reset_after = TRUE and
recurrent_activation = "sigmoid".
Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation
On the Properties of Neural Machine Translation: Encoder-Decoder Approaches
Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling
A Theoretically Grounded Application of Dropout in Recurrent Neural Networks
Other recurrent layers:
layer_cudnn_gru(),
layer_cudnn_lstm(),
layer_lstm(),
layer_rnn(),
layer_simple_rnn()