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
|
#include "adam_op.h"
namespace caffe2 {
REGISTER_CPU_OPERATOR(Adam, AdamOp<float, CPUContext>);
OPERATOR_SCHEMA(Adam)
.NumInputs(6)
.NumOutputs(3, 4)
.AllowInplace({{0, 0}, {1, 1}, {2, 2}})
.DeviceInferenceFunction([](const OperatorDef& def) {
auto op_device =
def.has_device_option() ? def.device_option() : DeviceOption();
vector<DeviceOption> in_dev(def.input_size(), op_device);
vector<DeviceOption> out_dev(def.output_size(), op_device);
// ITER input lives on CPU
in_dev[5] = DeviceOption();
return std::make_pair(in_dev, out_dev);
})
.SetDoc(R"DOC(
Computes the Adam update (https://arxiv.org/abs/1412.6980) for an
input gradient and momentum parameters. Concretely, given inputs
(param, m1, m2, grad, lr, iters),
t = iters + 1
correction_multiplier = sqrt(1 - power(beta2, t)) /
(1 - power(beta1, t))
m1_o = (beta1 * m1) + (1 - beta1) * grad
m2_o = (beta2 * m2) + (1 - beta2) * np.square(grad)
grad_o = correction_multiplier * m1_o / \
(sqrt(m2_o) + epsilon)
param_o = param + lr * grad_o
and returns (param_o, m1_o, m2_o, grad_o), in which grad_o is an optional output
)DOC")
.Input(0, "param", "Parameters to be updated")
.Input(1, "moment_1", "First moment history")
.Input(2, "moment_2", "Second moment history")
.Input(3, "grad", "Gradient computed")
.Input(4, "lr", "learning rate")
.Input(5, "iter", "iteration number")
.Output(0, "output_param", "Updated parameters")
.Output(1, "output_moment_1", "Updated first moment")
.Output(2, "output_moment_2", "Updated second moment")
.Output(3, "output_grad", "Optional Effective gradient")
.Arg("beta1", "Default 0.9")
.Arg("beta2", "Default 0.999")
.Arg("epsilon", "Default 1e-5");
REGISTER_CPU_OPERATOR(SparseAdam, SparseAdamOp<float, CPUContext>);
OPERATOR_SCHEMA(SparseAdam)
.NumInputs(7)
.NumOutputs(3, 4)
.EnforceInplace({{0, 0}, {1, 1}, {2, 2}})
.DeviceInferenceFunction([](const OperatorDef& def) {
auto op_device =
def.has_device_option() ? def.device_option() : DeviceOption();
vector<DeviceOption> in_dev(def.input_size(), op_device);
vector<DeviceOption> out_dev(def.output_size(), op_device);
// ITER input lives on CPU
in_dev[6] = DeviceOption();
return std::make_pair(in_dev, out_dev);
})
.SetDoc(R"DOC(
Computes the Adam Update for the sparse case.
Given inputs (param, moment1, moment2, indices, grad, lr, iter), runs the dense
Adam on (param, moment1[indices], momemnt2[indices], lr, iter) and returns
(new_param, new_moment1, new_moment2) as in dense case.
Adam can be customized as Rectified Adam (RAdam) by setting enableRAdam = true.
)DOC")
.Input(0, "param", "Parameters to be updated")
.Input(1, "moment_1", "First moment history")
.Input(2, "moment_2", "Second moment history")
.Input(3, "indices", "Sparse indices")
.Input(4, "grad", "Gradient computed")
.Input(5, "lr", "learning rate")
.Input(6, "iter", "iteration number")
.Output(0, "output_param", "Updated parameters")
.Output(1, "output_moment_1", "Updated first moment")
.Output(2, "output_moment_2", "Updated second moment")
.Output(3, "output_grad", "Optional Effective gradient")
.Arg("beta1", "Default 0.9")
.Arg("beta2", "Default 0.999")
.Arg("epsilon", "Default 1e-5")
.Arg("enableRAdam", "Default false");
REGISTER_CPU_OPERATOR(SmartDecaySparseAdam, SmartDecaySparseAdamOp<float, CPUContext>);
OPERATOR_SCHEMA(SmartDecaySparseAdam)
.NumInputs(8)
.NumOutputs(4)
.EnforceInplace({{0, 0}, {1, 1}, {2, 2}, {3, 3}})
.DeviceInferenceFunction([](const OperatorDef& def) {
auto op_device =
def.has_device_option() ? def.device_option() : DeviceOption();
vector<DeviceOption> in_dev(def.input_size(), op_device);
vector<DeviceOption> out_dev(def.output_size(), op_device);
// ITER input lives on CPU
in_dev[7] = DeviceOption();
return std::make_pair(in_dev, out_dev);
})
.SetDoc(R"DOC(
Computes the Adam Update for the sparse case.
Given inputs (param, moment1, moment2, indices, grad, lr, iter), runs the dense
Adam on (param, moment1[indices], momemnt2[indices], lr, iter) and returns
(new_param, new_moment1, new_moment2) as in dense case.
Adam can be customized as Rectified Adam (RAdam) by setting enableRAdam = true.
)DOC")
.Input(0, "param", "Parameters to be updated")
.Input(1, "moment_1", "First moment history")
.Input(2, "moment_2", "Second moment history")
.Input(3, "last_seen", "Minibatch index when each weight was last seen")
.Input(4, "indices", "Sparse indices")
.Input(5, "grad", "Gradient computed")
.Input(6, "lr", "learning rate")
.Input(7, "iter", "iteration number")
.Output(0, "output_param", "Updated parameters")
.Output(1, "output_moment_1", "Updated first moment")
.Output(2, "output_moment_2", "Updated second moment")
.Output(3, "output_last_seen", "Updated minibatch index when each weight was last seen")
.Arg("beta1", "Default 0.9")
.Arg("beta2", "Default 0.999")
.Arg("epsilon", "Default 1e-5");
REGISTER_CPU_OPERATOR(
RowWiseSparseAdam,
RowWiseSparseAdamOp<float, CPUContext>);
OPERATOR_SCHEMA(RowWiseSparseAdam)
.NumInputs(7)
.NumOutputs(3, 4)
.EnforceInplace({{0, 0}, {1, 1}, {2, 2}})
.DeviceInferenceFunction([](const OperatorDef& def) {
auto op_device =
def.has_device_option() ? def.device_option() : DeviceOption();
vector<DeviceOption> in_dev(def.input_size(), op_device);
vector<DeviceOption> out_dev(def.output_size(), op_device);
// ITER input lives on CPU
in_dev[6] = DeviceOption();
return std::make_pair(in_dev, out_dev);
})
.SetDoc(R"DOC(
Computes a modified Adam Update for the sparse case.
Given inputs (param, moment1, moment2, indices, grad, lr, iter), runs the
Adam update on (param, moment1[indices], moment2[indices], lr, iter) and returns
(new_param, new_moment1, new_moment2), where moment2 is a 1D tensor
with length equal to the number of rows in param:
shape(moment2) == shape(param)[0]. Each element of moment2 is
applied to an entire row of param, and the new moment2 values are
calculated by averaging across the row.
)DOC")
.Input(0, "param", "Parameters to be updated")
.Input(1, "moment_1", "First moment history")
.Input(2, "moment_2", "Second moment history")
.Input(3, "indices", "Sparse indices")
.Input(4, "grad", "Gradient computed")
.Input(5, "lr", "learning rate")
.Input(6, "iter", "iteration number")
.Output(0, "output_param", "Updated parameters")
.Output(1, "output_moment_1", "Updated first moment")
.Output(2, "output_moment_2", "Updated second moment")
.Output(3, "output_grad", "Optional Effective gradient")
.Arg("beta1", "Default 0.9")
.Arg("beta2", "Default 0.999")
.Arg("epsilon", "Default 1e-5");
SHOULD_NOT_DO_GRADIENT(Adam);
SHOULD_NOT_DO_GRADIENT(SparseAdam);
SHOULD_NOT_DO_GRADIENT(RowWiseSparseAdam);
} // namespace caffe2
|
