1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
|
#include "tensorflow/core/framework/op.h"
namespace tensorflow {
REGISTER_OP("ApplyGradientDescent")
.Input("var: Ref(T)")
.Input("alpha: T")
.Input("delta: T")
.Output("out: Ref(T)")
.Attr("T: numbertype")
.Attr("use_locking: bool = false")
.Doc(R"doc(
Update '*var' by subtracting 'alpha' * 'delta' from it.
var: Should be from a Variable().
alpha: Scaling factor. Must be a scalar.
delta: The change.
out: Same as "var".
use_locking: If True, the subtraction will be protected by a lock;
otherwise the behavior is undefined, but may exhibit less contention.
)doc");
REGISTER_OP("ApplyAdagrad")
.Input("var: Ref(T)")
.Input("accum: Ref(T)")
.Input("lr: T")
.Input("grad: T")
.Output("out: Ref(T)")
.Attr("T: numbertype")
.Attr("use_locking: bool = false")
.Doc(R"doc(
Update '*var' according to the adagrad scheme.
accum += grad * grad
var -= lr * grad * (1 / sqrt(accum))
var: Should be from a Variable().
accum: Should be from a Variable().
lr: Scaling factor. Must be a scalar.
grad: The gradient.
out: Same as "var".
use_locking: If True, updating of the var and accum tensors will be protected by
a lock; otherwise the behavior is undefined, but may exhibit less contention.
)doc");
REGISTER_OP("SparseApplyAdagrad")
.Input("var: Ref(T)")
.Input("accum: Ref(T)")
.Input("lr: T")
.Input("grad: T")
.Input("indices: Tindices")
.Output("out: Ref(T)")
.Attr("T: numbertype")
.Attr("Tindices: {int32, int64}")
.Attr("use_locking: bool = false")
.Doc(R"doc(
Update relevant entries in '*var' and '*accum' according to the adagrad scheme.
That is for rows we have grad for, we update var and accum as follows:
accum += grad * grad
var -= lr * grad * (1 / sqrt(accum))
var: Should be from a Variable().
accum: Should be from a Variable().
lr: Learning rate. Must be a scalar.
grad: The gradient.
indices: A vector of indices into the first dimension of var and accum.
out: Same as "var".
use_locking: If True, updating of the var and accum tensors will be protected by
a lock; otherwise the behavior is undefined, but may exhibit less contention.
)doc");
REGISTER_OP("ApplyMomentum")
.Input("var: Ref(T)")
.Input("accum: Ref(T)")
.Input("lr: T")
.Input("grad: T")
.Input("momentum: T")
.Output("out: Ref(T)")
.Attr("T: numbertype")
.Attr("use_locking: bool = false")
.Doc(R"doc(
Update '*var' according to the momentum scheme.
accum = accum * momentum + grad
var -= lr * accum
var: Should be from a Variable().
accum: Should be from a Variable().
lr: Scaling factor. Must be a scalar.
grad: The gradient.
momentum: Momentum. Must be a scalar.
out: Same as "var".
use_locking: If True, updating of the var and accum tensors will be protected by
a lock; otherwise the behavior is undefined, but may exhibit less contention.
)doc");
REGISTER_OP("SparseApplyMomentum")
.Input("var: Ref(T)")
.Input("accum: Ref(T)")
.Input("lr: T")
.Input("grad: T")
.Input("indices: Tindices")
.Input("momentum: T")
.Output("out: Ref(T)")
.Attr("T: numbertype")
.Attr("Tindices: {int32, int64}")
.Attr("use_locking: bool = false")
.Doc(R"doc(
Update relevant entries in '*var' and '*accum' according to the momentum scheme.
That is for rows we have grad for, we update var and accum as follows:
accum = accum * momentum + grad
var -= lr * accum
var: Should be from a Variable().
accum: Should be from a Variable().
lr: Learning rate. Must be a scalar.
grad: The gradient.
indices: A vector of indices into the first dimension of var and accum.
momentum: Momentum. Must be a scalar.
out: Same as "var".
use_locking: If True, updating of the var and accum tensors will be protected by
a lock; otherwise the behavior is undefined, but may exhibit less contention.
)doc");
REGISTER_OP("ApplyAdam")
.Input("var: Ref(T)")
.Input("m: Ref(T)")
.Input("v: Ref(T)")
.Input("beta1_power: T")
.Input("beta2_power: T")
.Input("lr: T")
.Input("beta1: T")
.Input("beta2: T")
.Input("epsilon: T")
.Input("grad: T")
.Output("out: Ref(T)")
.Attr("T: numbertype")
.Attr("use_locking: bool = false")
.Doc(R"doc(
Update '*var' according to the Adam algorithm.
lr_t <- learning_rate * sqrt(1 - beta2^t) / (1 - beta1^t)
m_t <- beta1 * m_{t-1} + (1 - beta1) * g_t
v_t <- beta2 * v_{t-1} + (1 - beta2) * g_t * g_t
variable <- variable - lr_t * m_t / (sqrt(v_t) + epsilon)
var: Should be from a Variable().
m: Should be from a Variable().
v: Should be from a Variable().
beta1_power: Must be a scalar.
beta2_power: Must be a scalar.
lr: Scaling factor. Must be a scalar.
beta1: Momentum factor. Must be a scalar.
beta2: Momentum factor. Must be a scalar.
epsilon: Ridge term. Must be a scalar.
grad: The gradient.
out: Same as "var".
use_locking: If True, updating of the var, m, and v tensors will be protected by
a lock; otherwise the behavior is undefined, but may exhibit less contention.
)doc");
REGISTER_OP("ApplyRMSProp")
.Input("var: Ref(T)")
.Input("ms: Ref(T)")
.Input("mom: Ref(T)")
.Input("lr: T")
.Input("rho: T")
.Input("momentum: T")
.Input("epsilon: T")
.Input("grad: T")
.Output("out: Ref(T)")
.Attr("T: numbertype")
.Attr("use_locking: bool = false")
.Doc(R"doc(
Update '*var' according to the RMSProp algorithm.
mean_square = decay * mean_square + (1-decay) * gradient ** 2
Delta = learning_rate * gradient / sqrt(mean_square + epsilon)
ms <- rho * ms_{t-1} + (1-rho) * grad * grad
mom <- momentum * mom_{t-1} + lr * grad / sqrt(ms + epsilon)
var <- var - mom
var: Should be from a Variable().
ms: Should be from a Variable().
mom: Should be from a Variable().
lr: Scaling factor. Must be a scalar.
epsilon: Ridge term. Must be a scalar.
rho: Decay rate. Must be a scalar.
grad: The gradient.
out: Same as "var".
use_locking: If True, updating of the var, m, and v tensors will be protected by
a lock; otherwise the behavior is undefined, but may exhibit less contention.
)doc");
} // namespace tensorflow
|
