GRU layer

GRU class

tf.keras.layers.GRU(
    units,
    activation="tanh",
    recurrent_activation="sigmoid",
    use_bias=True,
    kernel_initializer="glorot_uniform",
    recurrent_initializer="orthogonal",
    bias_initializer="zeros",
    kernel_regularizer=None,
    recurrent_regularizer=None,
    bias_regularizer=None,
    activity_regularizer=None,
    kernel_constraint=None,
    recurrent_constraint=None,
    bias_constraint=None,
    dropout=0.0,
    recurrent_dropout=0.0,
    implementation=2,
    return_sequences=False,
    return_state=False,
    go_backwards=False,
    stateful=False,
    unroll=False,
    time_major=False,
    reset_after=True,
    **kwargs
)

Gated Recurrent Unit - Cho et al. 2014.

See the Keras RNN API guide for details about the usage of RNN API.

Based on available runtime hardware and constraints, this layer will choose different implementations (cuDNN-based or pure-TensorFlow) to maximize the performance. If a GPU is available and all the arguments to the layer meet the requirement of the CuDNN kernel (see below for details), the layer will use a fast cuDNN implementation.

The requirements to use the cuDNN implementation are:

  1. activation == tanh
  2. recurrent_activation == sigmoid
  3. recurrent_dropout == 0
  4. unroll is False
  5. use_bias is True
  6. reset_after is True
  7. Inputs, if use masking, are strictly right-padded.
  8. Eager execution is enabled in the outermost context.

There are two variants of the GRU implementation. The default one is based on v3 and has reset gate applied to hidden state before matrix multiplication. The other one is based on original and has the order reversed.

The second variant is compatible with CuDNNGRU (GPU-only) and allows inference on CPU. Thus it has separate biases for kernel and recurrent_kernel. To use this variant, set 'reset_after'=True and recurrent_activation='sigmoid'.

For example:

>>> inputs = tf.random.normal([32, 10, 8])
>>> gru = tf.keras.layers.GRU(4)
>>> output = gru(inputs)
>>> print(output.shape)
(32, 4)
>>> gru = tf.keras.layers.GRU(4, return_sequences=True, return_state=True)
>>> whole_sequence_output, final_state = gru(inputs)
>>> print(whole_sequence_output.shape)
(32, 10, 4)
>>> print(final_state.shape)
(32, 4)

Arguments

  • units: Positive integer, dimensionality of the output space.
  • activation: Activation function to use. Default: hyperbolic tangent (tanh). If you pass None, no activation is applied (ie. "linear" activation: a(x) = x).
  • recurrent_activation: Activation function to use for the recurrent step. Default: sigmoid (sigmoid). If you pass None, no activation is applied (ie. "linear" activation: a(x) = x).
  • use_bias: Boolean, (default True), whether the layer uses a bias vector.
  • kernel_initializer: Initializer for the kernel weights matrix, used for the linear transformation of the inputs. Default: glorot_uniform.
  • recurrent_initializer: Initializer for the recurrent_kernel weights matrix, used for the linear transformation of the recurrent state. Default: orthogonal.
  • bias_initializer: Initializer for the bias vector. Default: zeros.
  • kernel_regularizer: Regularizer function applied to the kernel weights matrix. Default: None.
  • recurrent_regularizer: Regularizer function applied to the recurrent_kernel weights matrix. Default: None.
  • bias_regularizer: Regularizer function applied to the bias vector. Default: None.
  • activity_regularizer: Regularizer function applied to the output of the layer (its "activation"). Default: None.
  • kernel_constraint: Constraint function applied to the kernel weights matrix. Default: None.
  • recurrent_constraint: Constraint function applied to the recurrent_kernel weights matrix. Default: None.
  • bias_constraint: Constraint function applied to the bias vector. Default: None.
  • dropout: Float between 0 and 1. Fraction of the units to drop for the linear transformation of the inputs. Default: 0.
  • recurrent_dropout: Float between 0 and 1. Fraction of the units to drop for the linear transformation of the recurrent state. Default: 0.
  • implementation: Implementation mode, either 1 or 2. 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. Default: 2.
  • return_sequences: Boolean. Whether to return the last output in the output sequence, or the full sequence. Default: False.
  • return_state: Boolean. Whether to return the last state in addition to the output. Default: False.
  • go_backwards: Boolean (default False). If True, process the input sequence backwards and return the reversed sequence.
  • stateful: 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.
  • unroll: 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.
  • time_major: The shape format of the inputs and outputs tensors. 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.
  • reset_after: GRU convention (whether to apply reset gate after or before matrix multiplication). False = "before", True = "after" (default and CuDNN compatible).

Call arguments

  • inputs: A 3D tensor, with shape [batch, timesteps, feature].
  • mask: Binary tensor of shape [samples, timesteps] indicating whether a given timestep should be masked (optional, defaults to None).
  • training: Python boolean indicating whether the layer should behave in training mode or in inference mode. This argument is passed to the cell when calling it. This is only relevant if dropout or recurrent_dropout is used (optional, defaults to None).
  • initial_state: List of initial state tensors to be passed to the first call of the cell (optional, defaults to None which causes creation of zero-filled initial state tensors).