GRU - 1 vs 14#
Next section compares an older to a newer version of the same operator after both definition are converted into markdown text. Green means an addition to the newer version, red means a deletion. Anything else is unchanged.
- GRU1 → GRU14 +10 -27
GRU1 → GRU14
RENAMED
@@ -1 +1 @@
|
|
1
1
|
Computes an one-layer GRU. This operator is usually supported via some custom
|
2
2
|
implementation such as CuDNN.
|
3
3
|
Notations:
|
4
4
|
X - input tensor
|
5
5
|
z - update gate
|
6
6
|
r - reset gate
|
7
7
|
h - hidden gate
|
8
8
|
t - time step (t-1 means previous time step)
|
9
9
|
W[zrh] - W parameter weight matrix for update, reset, and hidden gates
|
10
10
|
R[zrh] - R recurrence weight matrix for update, reset, and hidden gates
|
11
11
|
Wb[zrh] - W bias vectors for update, reset, and hidden gates
|
12
12
|
Rb[zrh] - R bias vectors for update, reset, and hidden gates
|
13
13
|
WB[zrh] - W parameter weight matrix for backward update, reset, and hidden gates
|
14
14
|
RB[zrh] - R recurrence weight matrix for backward update, reset, and hidden gates
|
15
15
|
WBb[zrh] - W bias vectors for backward update, reset, and hidden gates
|
16
16
|
RBb[zrh] - R bias vectors for backward update, reset, and hidden gates
|
17
17
|
H - Hidden state
|
18
18
|
num_directions - 2 if direction == bidirectional else 1
|
19
19
|
Activation functions:
|
20
20
|
Relu(x) - max(0, x)
|
21
21
|
Tanh(x) - (1 - e^{-2x})/(1 + e^{-2x})
|
22
22
|
Sigmoid(x) - 1/(1 + e^{-x})
|
23
23
|
(NOTE: Below are optional)
|
24
24
|
Affine(x) - alpha*x + beta
|
25
25
|
LeakyRelu(x) - x if x >= 0 else alpha * x
|
26
26
|
ThresholdedRelu(x) - x if x >= alpha else 0
|
27
27
|
ScaledTanh(x) - alpha*Tanh(beta*x)
|
28
28
|
HardSigmoid(x) - min(max(alpha*x + beta, 0), 1)
|
29
29
|
Elu(x) - x if x >= 0 else alpha*(e^x - 1)
|
30
30
|
Softsign(x) - x/(1 + |x|)
|
31
31
|
Softplus(x) - log(1 + e^x)
|
32
32
|
Equations (Default: f=Sigmoid, g=Tanh):
|
33
|
-
- zt = f(Xt*(Wz^T) + Ht-1*
|
33
|
+
- zt = f(Xt*(Wz^T) + Ht-1*Rz + Wbz + Rbz)
|
34
|
-
- rt = f(Xt*(Wr^T) + Ht-1*
|
34
|
+
- rt = f(Xt*(Wr^T) + Ht-1*Rr + Wbr + Rbr)
|
35
|
-
- ht = g(Xt*(Wh^T) + (rt (.) Ht-1)*
|
35
|
+
- ht = g(Xt*(Wh^T) + (rt (.) Ht-1)*Rh + Rbh + Wbh) # default, when linear_before_reset = 0
|
36
|
-
- ht = g(Xt*(Wh^T) + (rt (.) (Ht-1*
|
36
|
+
- ht = g(Xt*(Wh^T) + (rt (.) (Ht-1*Rh + Rbh) + Wbh) # when linear_before_reset != 0
|
37
37
|
- Ht = (1 - zt) (.) ht + zt (.) Ht-1
|
38
|
-
This operator has **optional** inputs/outputs. See ONNX <https://github.com/onnx/onnx/blob/master/docs/IR.md>_ 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.
|
39
38
|
**Attributes**
|
40
39
|
* **activation_alpha**:
|
41
40
|
Optional scaling values used by some activation functions. The
|
42
41
|
values are consumed in the order of activation functions, for
|
42
|
+
example (f, g, h) in LSTM.
|
43
|
-
example (f, g, h) in LSTM. Default values are the same as of
|
44
|
-
corresponding ONNX operators.For example with LeakyRelu, the default
|
45
|
-
alpha is 0.01.
|
46
43
|
* **activation_beta**:
|
47
44
|
Optional scaling values used by some activation functions. The
|
48
45
|
values are consumed in the order of activation functions, for
|
46
|
+
example (f, g, h) in LSTM.
|
49
|
-
example (f, g, h) in LSTM. Default values are the same as of
|
50
|
-
corresponding ONNX operators.
|
51
47
|
* **activations**:
|
52
48
|
A list of 2 (or 4 if bidirectional) activation functions for update,
|
53
49
|
reset, and hidden gates. The activation functions must be one of the
|
54
50
|
activation functions specified above. Optional: See the equations
|
55
51
|
for default if not specified.
|
56
52
|
* **clip**:
|
57
53
|
Cell clip threshold. Clipping bounds the elements of a tensor in the
|
58
54
|
range of [-threshold, +threshold] and is applied to the input of
|
59
55
|
activations. No clip if not specified.
|
60
56
|
* **direction**:
|
61
57
|
Specify if the RNN is forward, reverse, or bidirectional. Must be
|
62
58
|
one of forward (default), reverse, or bidirectional.
|
63
59
|
* **hidden_size**:
|
64
60
|
Number of neurons in the hidden layer
|
61
|
+
* **output_sequence**:
|
62
|
+
The sequence output for the hidden is optional if 0. Default 0.
|
65
|
-
* **layout**:
|
66
|
-
The shape format of inputs X, initial_h and outputs Y, Y_h. If 0,
|
67
|
-
the following shapes are expected: X.shape = [seq_length,
|
68
|
-
batch_size, input_size], Y.shape = [seq_length, num_directions,
|
69
|
-
batch_size, hidden_size], initial_h.shape = Y_h.shape =
|
70
|
-
[num_directions, batch_size, hidden_size]. If 1, the following
|
71
|
-
shapes are expected: X.shape = [batch_size, seq_length, input_size],
|
72
|
-
Y.shape = [batch_size, seq_length, num_directions, hidden_size],
|
73
|
-
initial_h.shape = Y_h.shape = [batch_size, num_directions,
|
74
|
-
hidden_size].
|
75
|
-
* **linear_before_reset**:
|
76
|
-
When computing the output of the hidden gate, apply the linear
|
77
|
-
transformation before multiplying by the output of the reset gate.
|
78
63
|
**Inputs**
|
79
64
|
Between 3 and 6 inputs.
|
80
65
|
* **X** (heterogeneous) - **T**:
|
81
66
|
The input sequences packed (and potentially padded) into one 3-D
|
82
67
|
tensor with the shape of [seq_length, batch_size, input_size].
|
83
68
|
* **W** (heterogeneous) - **T**:
|
84
69
|
The weight tensor for the gates. Concatenation of W[zrh] and
|
85
70
|
WB[zrh] (if bidirectional) along dimension 0. This tensor has
|
86
71
|
shape [num_directions, 3*hidden_size, input_size].
|
87
72
|
* **R** (heterogeneous) - **T**:
|
88
73
|
The recurrence weight tensor. Concatenation of R[zrh] and
|
89
74
|
RB[zrh] (if bidirectional) along dimension 0. This tensor has
|
90
75
|
shape [num_directions, 3*hidden_size, hidden_size].
|
91
76
|
* **B** (optional, heterogeneous) - **T**:
|
92
77
|
The bias tensor for the gates. Concatenation of [Wb[zrh], Rb[zrh]]
|
93
78
|
and [WBb[zrh], RBb[zrh]] (if bidirectional) along dimension 0.
|
94
79
|
This tensor has shape [num_directions, 6*hidden_size]. Optional:
|
95
80
|
If not specified - assumed to be 0
|
96
81
|
* **sequence_lens** (optional, heterogeneous) - **T1**:
|
97
82
|
Optional tensor specifying lengths of the sequences in a batch. If
|
98
83
|
not specified - assumed all sequences in the batch to have length
|
99
84
|
seq_length. It has shape [batch_size].
|
100
85
|
* **initial_h** (optional, heterogeneous) - **T**:
|
101
86
|
Optional initial value of the hidden. If not specified - assumed to
|
102
87
|
be 0. It has shape [num_directions, batch_size, hidden_size].
|
103
88
|
**Outputs**
|
104
|
-
Between 0 and 2 outputs.
|
105
|
-
|
106
89
|
* **Y** (optional, heterogeneous) - **T**:
|
107
90
|
A tensor that concats all the intermediate output values of the
|
108
91
|
hidden. It has shape [seq_length, num_directions, batch_size,
|
109
|
-
hidden_size].
|
92
|
+
hidden_size]. It is optional if output_sequence is 0.
|
110
|
-
* **Y_h** (
|
93
|
+
* **Y_h** (heterogeneous) - **T**:
|
111
94
|
The last output value of the hidden. It has shape [num_directions,
|
112
95
|
batch_size, hidden_size].
|
113
96
|
**Type Constraints**
|
114
97
|
* **T** in (
|
115
98
|
tensor(double),
|
116
99
|
tensor(float),
|
117
100
|
tensor(float16)
|
118
101
|
):
|
119
102
|
Constrain input and output types to float tensors.
|
120
103
|
* **T1** in (
|
121
104
|
tensor(int32)
|
122
105
|
):
|
123
106
|
Constrain seq_lens to integer tensor.
|