LSTM - 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Next section compares an older to a newer version of the same operator after both definition are converted into markdown text.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. LSTM1 → LSTM14 +8 -18

LSTM1 → LSTM14 RENAMED Viewed

	@@ -1 +1 @@
1 1	Computes an one-layer LSTM. This operator is usually supported via some
2 2	custom implementation such as CuDNN.
3 3	Notations:
4 4	X - input tensor
5 5	i - input gate
6 6	o - output gate
7 7	f - forget gate
8 8	c - cell gate
9 9	t - time step (t-1 means previous time step)
10 10	W[iofc] - W parameter weight matrix for input, output, forget, and cell gates
11 11	R[iofc] - R recurrence weight matrix for input, output, forget, and cell gates
12 12	Wb[iofc] - W bias vectors for input, output, forget, and cell gates
13 13	Rb[iofc] - R bias vectors for input, output, forget, and cell gates
14 14	P[iof] - P peephole weight vector for input, output, and forget gates
15 15	WB[iofc] - W parameter weight matrix for backward input, output, forget, and cell gates
16 16	RB[iofc] - R recurrence weight matrix for backward input, output, forget, and cell gates
17 17	WBb[iofc] - W bias vectors for backward input, output, forget, and cell gates
18 18	RBb[iofc] - R bias vectors for backward input, output, forget, and cell gates
19 19	PB[iof] - P peephole weight vector for backward input, output, and forget gates
20 20	H - Hidden state
21 21	num_directions - 2 if direction == bidirectional else 1
22 22	Activation functions:
23 23	Relu(x) - max(0, x)
24 24	Tanh(x) - (1 - e^{-2x})/(1 + e^{-2x})
25 25	Sigmoid(x) - 1/(1 + e^{-x})
26 26	(NOTE: Below are optional)
27 27	Affine(x) - alpha*x + beta
28 28	LeakyRelu(x) - x if x >= 0 else alpha * x
29 29	ThresholdedRelu(x) - x if x >= alpha else 0
30 30	ScaledTanh(x) - alpha*Tanh(beta*x)
31 31	HardSigmoid(x) - min(max(alpha*x + beta, 0), 1)
32 32	Elu(x) - x if x >= 0 else alpha*(e^x - 1)
33 33	Softsign(x) - x/(1 + |x|)
34 34	Softplus(x) - log(1 + e^x)
35 35	Equations (Default: f=Sigmoid, g=Tanh, h=Tanh):
36	- - it = f(Xt*(Wi^T) + Ht-1*(Ri^T) + Pi (.) Ct-1 + Wbi + Rbi)
36	+ - it = f(Xt*(Wi^T) + Ht-1*Ri + Pi (.) Ct-1 + Wbi + Rbi)
37	- - ft = f(Xt*(Wf^T) + Ht-1*(Rf^T) + Pf (.) Ct-1 + Wbf + Rbf)
37	+ - ft = f(Xt*(Wf^T) + Ht-1*Rf + Pf (.) Ct-1 + Wbf + Rbf)
38	- - ct = g(Xt*(Wc^T) + Ht-1*(Rc^T) + Wbc + Rbc)
38	+ - ct = g(Xt*(Wc^T) + Ht-1*Rc + Wbc + Rbc)
39 39	- Ct = ft (.) Ct-1 + it (.) ct
40	- - ot = f(Xt*(Wo^T) + Ht-1*(Ro^T) + Po (.) Ct + Wbo + Rbo)
40	+ - ot = f(Xt*(Wo^T) + Ht-1*Ro + Po (.) Ct + Wbo + Rbo)
41 41	- Ht = ot (.) h(Ct)
42	- 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.
43 42	**Attributes**
44 43	* **activation_alpha**:
45 44	Optional scaling values used by some activation functions. The
46 45	values are consumed in the order of activation functions, for
47 46	example (f, g, h) in LSTM. Default values are the same as of
48 47	corresponding ONNX operators.For example with LeakyRelu, the default
49 48	alpha is 0.01.
50 49	* **activation_beta**:
51 50	Optional scaling values used by some activation functions. The
52 51	values are consumed in the order of activation functions, for
53 52	example (f, g, h) in LSTM. Default values are the same as of
54 53	corresponding ONNX operators.
55 54	* **activations**:
56 55	A list of 3 (or 6 if bidirectional) activation functions for input,
57 56	output, forget, cell, and hidden. The activation functions must be
58 57	one of the activation functions specified above. Optional: See the
59 58	equations for default if not specified.
60 59	* **clip**:
61 60	Cell clip threshold. Clipping bounds the elements of a tensor in the
62 61	range of [-threshold, +threshold] and is applied to the input of
63 62	activations. No clip if not specified.
64 63	* **direction**:
65 64	Specify if the RNN is forward, reverse, or bidirectional. Must be
66 65	one of forward (default), reverse, or bidirectional.
67 66	* **hidden_size**:
68 67	Number of neurons in the hidden layer
69 68	* **input_forget**:
70	- Couple the input and forget gates if 1.
69	+ Couple the input and forget gates if 1, default 0.
70	+ * **output_sequence**:
71	+ The sequence output for the hidden is optional if 0. Default 0.
71	- * **layout**:
72	- The shape format of inputs X, initial_h, initial_c and outputs Y,
73	- Y_h, Y_c. If 0, the following shapes are expected: X.shape =
74	- [seq_length, batch_size, input_size], Y.shape = [seq_length,
75	- num_directions, batch_size, hidden_size], initial_h.shape =
76	- Y_h.shape = initial_c.shape = Y_c.shape = [num_directions,
77	- batch_size, hidden_size]. If 1, the following shapes are expected:
78	- X.shape = [batch_size, seq_length, input_size], Y.shape =
79	- [batch_size, seq_length, num_directions, hidden_size],
80	- initial_h.shape = Y_h.shape = initial_c.shape = Y_c.shape =
81	- [batch_size, num_directions, hidden_size].
82 72	**Inputs**
83 73	Between 3 and 8 inputs.
84 74	* **X** (heterogeneous) - **T**:
85 75	The input sequences packed (and potentially padded) into one 3-D
86 76	tensor with the shape of [seq_length, batch_size, input_size].
87 77	* **W** (heterogeneous) - **T**:
88 78	The weight tensor for the gates. Concatenation of W[iofc] and
89 79	WB[iofc] (if bidirectional) along dimension 0. The tensor has
90 80	shape [num_directions, 4*hidden_size, input_size].
91 81	* **R** (heterogeneous) - **T**:
92 82	The recurrence weight tensor. Concatenation of R[iofc] and
93 83	RB[iofc] (if bidirectional) along dimension 0. This tensor has
94 84	shape [num_directions, 4*hidden_size, hidden_size].
95 85	* **B** (optional, heterogeneous) - **T**:
96 86	The bias tensor for input gate. Concatenation of [Wb[iofc],
97 87	Rb[iofc]], and [WBb[iofc], RBb[iofc]] (if bidirectional) along
98 88	dimension 0. This tensor has shape [num_directions,
99 89	8*hidden_size]. Optional: If not specified - assumed to be 0.
100 90	* **sequence_lens** (optional, heterogeneous) - **T1**:
101 91	Optional tensor specifying lengths of the sequences in a batch. If
102 92	not specified - assumed all sequences in the batch to have length
103 93	seq_length. It has shape [batch_size].
104 94	* **initial_h** (optional, heterogeneous) - **T**:
105 95	Optional initial value of the hidden. If not specified - assumed to
106 96	be 0. It has shape [num_directions, batch_size, hidden_size].
107 97	* **initial_c** (optional, heterogeneous) - **T**:
108 98	Optional initial value of the cell. If not specified - assumed to be
109 99	0. It has shape [num_directions, batch_size, hidden_size].
110 100	* **P** (optional, heterogeneous) - **T**:
111 101	The weight tensor for peepholes. Concatenation of P[iof] and
112 102	PB[iof] (if bidirectional) along dimension 0. It has shape
113 103	[num_directions, 3*hidde_size]. Optional: If not specified -
114 104	assumed to be 0.
115 105	**Outputs**
116 106	Between 0 and 3 outputs.
117 107	* **Y** (optional, heterogeneous) - **T**:
118 108	A tensor that concats all the intermediate output values of the
119 109	hidden. It has shape [seq_length, num_directions, batch_size,
120	- hidden_size].
110	+ hidden_size]. It is optional if output_sequence is 0.
121 111	* **Y_h** (optional, heterogeneous) - **T**:
122 112	The last output value of the hidden. It has shape [num_directions,
123 113	batch_size, hidden_size].
124 114	* **Y_c** (optional, heterogeneous) - **T**:
125 115	The last output value of the cell. It has shape [num_directions,
126 116	batch_size, hidden_size].
127 117	**Type Constraints**
128 118	* **T** in (
129 119	tensor(double),
130 120	tensor(float),
131 121	tensor(float16)
132 122	):
133 123	Constrain input and output types to float tensors.
134 124	* **T1** in (
135 125	tensor(int32)
136 126	):
137 127	Constrain seq_lens to integer tensor.