GRU#

GRU - 14 #

Version

name: GRU (GitHub)
domain: main
since_version: 14
function: False
support_level: SupportType.COMMON
shape inference: True

This version of the operator has been available since version 14.

Summary

Computes an one-layer GRU. This operator is usually supported via some custom implementation such as CuDNN.

Notations:

X - input tensor

z - update gate

r - reset gate

h - hidden gate

t - time step (t-1 means previous time step)

W[zrh] - W parameter weight matrix for update, reset, and hidden gates

R[zrh] - R recurrence weight matrix for update, reset, and hidden gates

Wb[zrh] - W bias vectors for update, reset, and hidden gates

Rb[zrh] - R bias vectors for update, reset, and hidden gates

WB[zrh] - W parameter weight matrix for backward update, reset, and hidden gates

RB[zrh] - R recurrence weight matrix for backward update, reset, and hidden gates

WBb[zrh] - W bias vectors for backward update, reset, and hidden gates

RBb[zrh] - R bias vectors for backward update, reset, and hidden gates

H - Hidden state

num_directions - 2 if direction == bidirectional else 1

Activation functions:

Relu(x) - max(0, x)

Tanh(x) - (1 - e^{-2x})/(1 + e^{-2x})

Sigmoid(x) - 1/(1 + e^{-x})

(NOTE: Below are optional)

Affine(x) - alpha*x + beta

LeakyRelu(x) - x if x >= 0 else alpha * x

ThresholdedRelu(x) - x if x >= alpha else 0

ScaledTanh(x) - alpha*Tanh(beta*x)

HardSigmoid(x) - min(max(alpha*x + beta, 0), 1)

Elu(x) - x if x >= 0 else alpha*(e^x - 1)

Softsign(x) - x/(1 + |x|)

Softplus(x) - log(1 + e^x)

Equations (Default: f=Sigmoid, g=Tanh):

zt = f(Xt*(Wz^T) + Ht-1*(Rz^T) + Wbz + Rbz)

rt = f(Xt*(Wr^T) + Ht-1*(Rr^T) + Wbr + Rbr)

ht = g(Xt*(Wh^T) + (rt (.) Ht-1)*(Rh^T) + Rbh + Wbh) # default, when linear_before_reset = 0

ht = g(Xt*(Wh^T) + (rt (.) (Ht-1*(Rh^T) + Rbh)) + Wbh) # when linear_before_reset != 0

Ht = (1 - zt) (.) ht + zt (.) Ht-1

This operator has optional inputs/outputs. See ONNX for more details about the representation of optional arguments. An empty string may be used in the place of an actual argument’s name to indicate a missing argument. Trailing optional arguments (those not followed by an argument that is present) may also be simply omitted.

Attributes

activation_alpha: Optional scaling values used by some activation functions. The values are consumed in the order of activation functions, for example (f, g, h) in LSTM. Default values are the same as of corresponding ONNX operators.For example with LeakyRelu, the default alpha is 0.01.
activation_beta: Optional scaling values used by some activation functions. The values are consumed in the order of activation functions, for example (f, g, h) in LSTM. Default values are the same as of corresponding ONNX operators.
activations: A list of 2 (or 4 if bidirectional) activation functions for update, reset, and hidden gates. The activation functions must be one of the activation functions specified above. Optional: See the equations for default if not specified.
clip: Cell clip threshold. Clipping bounds the elements of a tensor in the range of [-threshold, +threshold] and is applied to the input of activations. No clip if not specified.
direction: Specify if the RNN is forward, reverse, or bidirectional. Must be one of forward (default), reverse, or bidirectional.
hidden_size: Number of neurons in the hidden layer
layout: The shape format of inputs X, initial_h and outputs Y, Y_h. If 0, the following shapes are expected: X.shape = [seq_length, batch_size, input_size], Y.shape = [seq_length, num_directions, batch_size, hidden_size], initial_h.shape = Y_h.shape = [num_directions, batch_size, hidden_size]. If 1, the following shapes are expected: X.shape = [batch_size, seq_length, input_size], Y.shape = [batch_size, seq_length, num_directions, hidden_size], initial_h.shape = Y_h.shape = [batch_size, num_directions, hidden_size].
linear_before_reset: When computing the output of the hidden gate, apply the linear transformation before multiplying by the output of the reset gate.

Inputs

Between 3 and 6 inputs.

X (heterogeneous) - T: The input sequences packed (and potentially padded) into one 3-D tensor with the shape of [seq_length, batch_size, input_size].
W (heterogeneous) - T: The weight tensor for the gates. Concatenation of W[zrh] and WB[zrh] (if bidirectional) along dimension 0. This tensor has shape [num_directions, 3*hidden_size, input_size].
R (heterogeneous) - T: The recurrence weight tensor. Concatenation of R[zrh] and RB[zrh] (if bidirectional) along dimension 0. This tensor has shape [num_directions, 3*hidden_size, hidden_size].
B (optional, heterogeneous) - T: The bias tensor for the gates. Concatenation of [Wb[zrh], Rb[zrh]] and [WBb[zrh], RBb[zrh]] (if bidirectional) along dimension 0. This tensor has shape [num_directions, 6*hidden_size]. Optional: If not specified - assumed to be 0
sequence_lens (optional, heterogeneous) - T1: Optional tensor specifying lengths of the sequences in a batch. If not specified - assumed all sequences in the batch to have length seq_length. It has shape [batch_size].
initial_h (optional, heterogeneous) - T: Optional initial value of the hidden. If not specified - assumed to be 0. It has shape [num_directions, batch_size, hidden_size].

Outputs

Between 0 and 2 outputs.

Y (optional, heterogeneous) - T: A tensor that concats all the intermediate output values of the hidden. It has shape [seq_length, num_directions, batch_size, hidden_size].
Y_h (optional, heterogeneous) - T: The last output value of the hidden. It has shape [num_directions, batch_size, hidden_size].

Type Constraints

T in ( tensor(double), tensor(float), tensor(float16) ): Constrain input and output types to float tensors.
T1 in ( tensor(int32) ): Constrain seq_lens to integer tensor.

Examples

_defaults

import numpy as np
import onnx

input = np.array([[[1.0, 2.0], [3.0, 4.0], [5.0, 6.0]]]).astype(np.float32)

input_size = 2
hidden_size = 5
weight_scale = 0.1
number_of_gates = 3

node = onnx.helper.make_node(
    "GRU", inputs=["X", "W", "R"], outputs=["", "Y_h"], hidden_size=hidden_size
)

W = weight_scale * np.ones(
    (1, number_of_gates * hidden_size, input_size)
).astype(np.float32)
R = weight_scale * np.ones(
    (1, number_of_gates * hidden_size, hidden_size)
).astype(np.float32)

gru = GRU_Helper(X=input, W=W, R=R)
_, Y_h = gru.step()
expect(
    node,
    inputs=[input, W, R],
    outputs=[Y_h.astype(np.float32)],
    name="test_gru_defaults",
)

_initial_bias

import numpy as np
import onnx

input = np.array([[[1.0, 2.0, 3.0], [4.0, 5.0, 6.0], [7.0, 8.0, 9.0]]]).astype(
    np.float32
)

input_size = 3
hidden_size = 3
weight_scale = 0.1
custom_bias = 0.1
number_of_gates = 3

node = onnx.helper.make_node(
    "GRU",
    inputs=["X", "W", "R", "B"],
    outputs=["", "Y_h"],
    hidden_size=hidden_size,
)

W = weight_scale * np.ones(
    (1, number_of_gates * hidden_size, input_size)
).astype(np.float32)
R = weight_scale * np.ones(
    (1, number_of_gates * hidden_size, hidden_size)
).astype(np.float32)

# Adding custom bias
W_B = custom_bias * np.ones((1, number_of_gates * hidden_size)).astype(
    np.float32
)
R_B = np.zeros((1, number_of_gates * hidden_size)).astype(np.float32)
B = np.concatenate((W_B, R_B), axis=1)

gru = GRU_Helper(X=input, W=W, R=R, B=B)
_, Y_h = gru.step()
expect(
    node,
    inputs=[input, W, R, B],
    outputs=[Y_h.astype(np.float32)],
    name="test_gru_with_initial_bias",
)

_seq_length

import numpy as np
import onnx

input = np.array(
    [
        [[1.0, 2.0, 3.0], [4.0, 5.0, 6.0], [7.0, 8.0, 9.0]],
        [[10.0, 11.0, 12.0], [13.0, 14.0, 15.0], [16.0, 17.0, 18.0]],
    ]
).astype(np.float32)

input_size = 3
hidden_size = 5
number_of_gates = 3

node = onnx.helper.make_node(
    "GRU",
    inputs=["X", "W", "R", "B"],
    outputs=["", "Y_h"],
    hidden_size=hidden_size,
)

W = np.random.randn(1, number_of_gates * hidden_size, input_size).astype(
    np.float32
)
R = np.random.randn(1, number_of_gates * hidden_size, hidden_size).astype(
    np.float32
)

# Adding custom bias
W_B = np.random.randn(1, number_of_gates * hidden_size).astype(np.float32)
R_B = np.random.randn(1, number_of_gates * hidden_size).astype(np.float32)
B = np.concatenate((W_B, R_B), axis=1)

gru = GRU_Helper(X=input, W=W, R=R, B=B)
_, Y_h = gru.step()
expect(
    node,
    inputs=[input, W, R, B],
    outputs=[Y_h.astype(np.float32)],
    name="test_gru_seq_length",
)

_batchwise

import numpy as np
import onnx

input = np.array([[[1.0, 2.0]], [[3.0, 4.0]], [[5.0, 6.0]]]).astype(np.float32)

input_size = 2
hidden_size = 6
number_of_gates = 3
weight_scale = 0.2
layout = 1

node = onnx.helper.make_node(
    "GRU",
    inputs=["X", "W", "R"],
    outputs=["Y", "Y_h"],
    hidden_size=hidden_size,
    layout=layout,
)

W = weight_scale * np.ones(
    (1, number_of_gates * hidden_size, input_size)
).astype(np.float32)
R = weight_scale * np.ones(
    (1, number_of_gates * hidden_size, hidden_size)
).astype(np.float32)

gru = GRU_Helper(X=input, W=W, R=R, layout=layout)
Y, Y_h = gru.step()
expect(
    node,
    inputs=[input, W, R],
    outputs=[Y.astype(np.float32), Y_h.astype(np.float32)],
    name="test_gru_batchwise",
)

GRU - 7 vs 14

GRU - 7 #

Version

name: GRU (GitHub)
domain: main
since_version: 7
function: False
support_level: SupportType.COMMON
shape inference: True

This version of the operator has been available since version 7.