RNN - 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Next section compares an older to a newer version of the same operator after both definition are converted into markdown text.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.

Files changed (1) hide show

1. RNN1 → RNN14 +4 -13

RNN1 → RNN14 RENAMED Viewed

	@@ -1 +1 @@
1 1	Computes an one-layer simple RNN. This operator is usually supported
2 2	via some custom implementation such as CuDNN.
3 3	Notations:
4 4	X - input tensor
5 5	i - input gate
6 6	t - time step (t-1 means previous time step)
7 7	Wi - W parameter weight matrix for input gate
8 8	Ri - R recurrence weight matrix for input gate
9 9	Wbi - W parameter bias vector for input gate
10 10	Rbi - R parameter bias vector for input gate
11 11	WBi - W parameter weight matrix for backward input gate
12 12	RBi - R recurrence weight matrix for backward input gate
13 13	WBbi - WR bias vectors for backward input gate
14 14	RBbi - RR bias vectors for backward input gate
15 15	H - Hidden state
16 16	num_directions - 2 if direction == bidirectional else 1
17 17	Activation functions:
18 18	Relu(x) - max(0, x)
19 19	Tanh(x) - (1 - e^{-2x})/(1 + e^{-2x})
20 20	Sigmoid(x) - 1/(1 + e^{-x})
21 21	(NOTE: Below are optional)
22 22	Affine(x) - alpha*x + beta
23 23	LeakyRelu(x) - x if x >= 0 else alpha * x
24 24	ThresholdedRelu(x) - x if x >= alpha else 0
25 25	ScaledTanh(x) - alpha*Tanh(beta*x)
26 26	HardSigmoid(x) - min(max(alpha*x + beta, 0), 1)
27 27	Elu(x) - x if x >= 0 else alpha*(e^x - 1)
28 28	Softsign(x) - x/(1 + |x|)
29 29	Softplus(x) - log(1 + e^x)
30 30	Equations (Default: f=Tanh):
31	- - Ht = f(Xt*(Wi^T) + Ht-1*(Ri^T) + Wbi + Rbi)
31	+ - Ht = f(Xt*(Wi^T) + Ht-1*Ri + Wbi + Rbi)
32	- 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.
33 32	**Attributes**
34 33	* **activation_alpha**:
35 34	Optional scaling values used by some activation functions. The
36 35	values are consumed in the order of activation functions, for
37 36	example (f, g, h) in LSTM. Default values are the same as of
38 37	corresponding ONNX operators.For example with LeakyRelu, the default
39 38	alpha is 0.01.
40 39	* **activation_beta**:
41 40	Optional scaling values used by some activation functions. The
42 41	values are consumed in the order of activation functions, for
43 42	example (f, g, h) in LSTM. Default values are the same as of
44 43	corresponding ONNX operators.
45 44	* **activations**:
46 45	One (or two if bidirectional) activation function for input gate.
47 46	The activation function must be one of the activation functions
48 47	specified above. Optional: Default Tanh if not specified.
49 48	* **clip**:
50 49	Cell clip threshold. Clipping bounds the elements of a tensor in the
51 50	range of [-threshold, +threshold] and is applied to the input of
52 51	activations. No clip if not specified.
53 52	* **direction**:
54 53	Specify if the RNN is forward, reverse, or bidirectional. Must be
55 54	one of forward (default), reverse, or bidirectional.
56 55	* **hidden_size**:
57 56	Number of neurons in the hidden layer
57	+ * **output_sequence**:
58	+ The sequence output for the hidden is optional if 0. Default 0.
58	- * **layout**:
59	- The shape format of inputs X, initial_h and outputs Y, Y_h. If 0,
60	- the following shapes are expected: X.shape = [seq_length,
61	- batch_size, input_size], Y.shape = [seq_length, num_directions,
62	- batch_size, hidden_size], initial_h.shape = Y_h.shape =
63	- [num_directions, batch_size, hidden_size]. If 1, the following
64	- shapes are expected: X.shape = [batch_size, seq_length, input_size],
65	- Y.shape = [batch_size, seq_length, num_directions, hidden_size],
66	- initial_h.shape = Y_h.shape = [batch_size, num_directions,
67	- hidden_size].
68 59	**Inputs**
69 60	Between 3 and 6 inputs.
70 61	* **X** (heterogeneous) - **T**:
71 62	The input sequences packed (and potentially padded) into one 3-D
72 63	tensor with the shape of [seq_length, batch_size, input_size].
73 64	* **W** (heterogeneous) - **T**:
74 65	The weight tensor for input gate. Concatenation of Wi and WBi
75 66	(if bidirectional). The tensor has shape [num_directions,
76 67	hidden_size, input_size].
77 68	* **R** (heterogeneous) - **T**:
78 69	The recurrence weight tensor. Concatenation of Ri and RBi (if
79 70	bidirectional). The tensor has shape [num_directions, hidden_size,
80 71	hidden_size].
81 72	* **B** (optional, heterogeneous) - **T**:
82 73	The bias tensor for input gate. Concatenation of [Wbi, Rbi] and
83 74	[WBbi, RBbi] (if bidirectional). The tensor has shape
84 75	[num_directions, 2*hidden_size]. Optional: If not specified -
85 76	assumed to be 0.
86 77	* **sequence_lens** (optional, heterogeneous) - **T1**:
87 78	Optional tensor specifying lengths of the sequences in a batch. If
88 79	not specified - assumed all sequences in the batch to have length
89 80	seq_length. It has shape [batch_size].
90 81	* **initial_h** (optional, heterogeneous) - **T**:
91 82	Optional initial value of the hidden. If not specified - assumed to
92 83	be 0. It has shape [num_directions, batch_size, hidden_size].
93 84	**Outputs**
94 85	Between 0 and 2 outputs.
95 86	* **Y** (optional, heterogeneous) - **T**:
96 87	A tensor that concats all the intermediate output values of the
97 88	hidden. It has shape [seq_length, num_directions, batch_size,
98	- hidden_size].
89	+ hidden_size]. It is optional if output_sequence is 0.
99 90	* **Y_h** (optional, heterogeneous) - **T**:
100 91	The last output value of the hidden. It has shape [num_directions,
101 92	batch_size, hidden_size].
102 93	**Type Constraints**
103 94	* **T** in (
104 95	tensor(double),
105 96	tensor(float),
106 97	tensor(float16)
107 98	):
108 99	Constrain input and output types to float tensors.
109 100	* **T1** in (
110 101	tensor(int32)
111 102	):
112 103	Constrain seq_lens to integer tensor.