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
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
|
#include "caffe2/sgd/adagrad_fused.h"
namespace caffe2 {
struct adagrad_update_prefetch_inlined {
void operator()(
int N,
const float* w,
const float* w_n, // prefetch ptr
const float* g,
const float* h,
const float* h_n, // prefetch ptr
float* nw,
float* nw_n, // prefetch ptr
float* nh,
float* nh_n, // prefetch ptr
float epsilon,
float lr,
float weight_decay) {
return internal::adagrad_update_prefetch_inlined(
N, w, w_n, g, h, h_n, nw, nw_n, nh, nh_n, epsilon, lr, weight_decay);
}
};
OPERATOR_SCHEMA(SparseAdagradFusedWithSparseLengthsSumGradient)
.NumInputs(6)
.NumOutputs(2)
.EnforceOneToOneInplace()
.SetDoc(R"DOC(
Fused operator of
SparseLengthsIndicesInGradientSumGradient (gradient of SparseLengthsSum) +
SparseAdagrad.
Given inputs (param, moment, indices, grad, lr), runs the sparse AdaGrad
update on (param, grad, moment[indices], lr), and returns (new_param,
new_moment) as in the dense case. Additional input (lengths) is for fused
SparseLengthsIndicesInGradientSumGradient operator.
)DOC")
.Input(0, "param", "Parameters to be updated")
.Input(1, "moment", "Moment history")
.Input(
2,
"indices",
"Integer vector containing indices of the first dimension of param for the slices that are being updated")
.Input(3, "grad", "Gradient computed")
.Input(4, "lr", "learning rate")
.Input(
5,
"lengths",
"Non negative vector with sum of elements equal to indices length")
.Output(0, "output_param", "Updated parameters")
.Output(1, "output_moment", "Updated moment")
.Arg("epsilon", "Default 1e-5");
REGISTER_CPU_OPERATOR(
SparseAdagradFusedWithSparseLengthsSumGradient,
SparseAdagradFusedWithSparseLengthsSumGradientOp<
float,
float,
int,
adagrad_update_prefetch_inlined,
/*is_mean=*/false>);
// Match the GPU Approx op, here Approx and Exact are the same for
// SparseAdagradFusedWithSparseLengthsSumGradient op
OPERATOR_SCHEMA(SparseAdagradFusedWithSparseLengthsSumGradientApprox)
.NumInputs(6)
.NumOutputs(2)
.EnforceOneToOneInplace()
.SetDoc(R"DOC(
Fused operator of
SparseLengthsIndicesInGradientSumGradient (gradient of SparseLengthsSum) +
SparseAdagrad.
Given inputs (param, moment, indices, grad, lr), runs the sparse AdaGrad
update on (param, grad, moment[indices], lr), and returns (new_param,
new_moment) as in the dense case. Additional input (lengths) is for fused
SparseLengthsIndicesInGradientSumGradient operator.
)DOC")
.Input(0, "param", "Parameters to be updated")
.Input(1, "moment", "Moment history")
.Input(
2,
"indices",
"Integer vector containing indices of the first dimension of param for the slices that are being updated")
.Input(3, "grad", "Gradient computed")
.Input(4, "lr", "learning rate")
.Input(
5,
"lengths",
"Non negative vector with sum of elements equal to indices length")
.Output(0, "output_param", "Updated parameters")
.Output(1, "output_moment", "Updated moment")
.Arg("epsilon", "Default 1e-5");
REGISTER_CPU_OPERATOR(
SparseAdagradFusedWithSparseLengthsSumGradientApprox,
SparseAdagradFusedWithSparseLengthsSumGradientOp<
float,
float,
int,
adagrad_update_prefetch_inlined,
/*is_mean=*/false>);
OPERATOR_SCHEMA(SparseAdagradFusedWithSparseLengthsMeanGradient)
.NumInputs(6)
.NumOutputs(2)
.EnforceOneToOneInplace()
.SetDoc(R"DOC(
Fused operator of
SparseLengthsIndicesInGradientMeanGradient (gradient of SparseLengthsMean) +
SparseAdagrad.
Given inputs (param, moment, indices, grad, lr), runs the sparse AdaGrad
update on (param, grad, moment[indices], lr), and returns (new_param,
new_moment) as in the dense case. Additional input (lengths) is for fused
SparseLengthsIndicesInGradientMeanGradient operator.
)DOC")
.Input(0, "param", "Parameters to be updated")
.Input(1, "moment", "Moment history")
.Input(
2,
"indices",
"Integer vector containing indices of the first dimension of param for the slices that are being updated")
.Input(3, "grad", "Gradient computed")
.Input(4, "lr", "learning rate")
.Input(
5,
"lengths",
"Non negative vector with sum of elements equal to indices length")
.Output(0, "output_param", "Updated parameters")
.Output(1, "output_moment", "Updated moment")
.Arg("epsilon", "Default 1e-5");
REGISTER_CPU_OPERATOR(
SparseAdagradFusedWithSparseLengthsMeanGradient,
SparseAdagradFusedWithSparseLengthsSumGradientOp<
float,
float,
int,
adagrad_update_prefetch_inlined,
/*is_mean=*/true>);
// Match the GPU Approx op, here Approx and Exact are the same for
// SparseAdagradFusedWithSparseLengthsMeanGradient op
OPERATOR_SCHEMA(SparseAdagradFusedWithSparseLengthsMeanGradientApprox)
.NumInputs(6)
.NumOutputs(2)
.EnforceOneToOneInplace()
.SetDoc(R"DOC(
Fused operator of
SparseLengthsIndicesInGradientMeanGradient (gradient of SparseLengthsMean) +
SparseAdagrad.
Given inputs (param, moment, indices, grad, lr), runs the sparse AdaGrad
update on (param, grad, moment[indices], lr), and returns (new_param,
new_moment) as in the dense case. Additional input (lengths) is for fused
SparseLengthsIndicesInGradientMeanGradient operator.
)DOC")
.Input(0, "param", "Parameters to be updated")
.Input(1, "moment", "Moment history")
.Input(
2,
"indices",
"Integer vector containing indices of the first dimension of param for the slices that are being updated")
.Input(3, "grad", "Gradient computed")
.Input(4, "lr", "learning rate")
.Input(
5,
"lengths",
"Non negative vector with sum of elements equal to indices length")
.Output(0, "output_param", "Updated parameters")
.Output(1, "output_moment", "Updated moment")
.Arg("epsilon", "Default 1e-5");
REGISTER_CPU_OPERATOR(
SparseAdagradFusedWithSparseLengthsMeanGradientApprox,
SparseAdagradFusedWithSparseLengthsSumGradientOp<
float,
float,
int,
adagrad_update_prefetch_inlined,
/*is_mean=*/true>);
OPERATOR_SCHEMA(SparseAdagradFusedWithSparseLengthsWeightedSumGradient)
.NumInputs(7)
.NumOutputs(3)
.EnforceInplace({{0, 0}, {1, 1}})
.SetDoc(R"DOC(
Fused operator of SparseLengthsIndicesInGradientWeightedSumWithMainInputGradient
(gradient of SparseLengthsWeightedSum) + SparseAdagrad, where weights are
positional weights computed with LengthsRangeFill + Gather pattern.
Given inputs (param, moment, indices, grad, lr), runs the sparse AdaGrad
update on (param, grad, moment[indices], lr), and returns (new_param,
new_moment) as in the dense case.
There're auxiliary inputs (aux_param) for which gradient is computed
and returns (aux_grad).
Yet additional input (lengths) is for fused
SparseLengthsIndicesInGradientWeightedSumWithMainInputGradient operator.
)DOC")
.Input(0, "param", "Parameters to be updated")
.Input(1, "moment", "Moment history")
.Input(2, "aux_param", "Auxiliary parameters to be updated")
.Input(
3,
"indices",
"Integer vector containing indices of the first dimension of param for the slices that are being updated")
.Input(4, "grad", "Gradient computed")
.Input(5, "lr", "learning rate")
.Input(
6,
"lengths",
"Non negative vector with sum of elements equal to indices length")
.Output(0, "output_param", "Updated parameters")
.Output(1, "output_moment", "Updated moment")
.Output(2, "aux_grad", "Auxiliary gradient")
.Arg("epsilon", "Default 1e-5");
REGISTER_CPU_OPERATOR(
SparseAdagradFusedWithSparseLengthsWeightedSumGradient,
SparseAdagradFusedWithSparseLengthsWeightedSumGradientOp<
float,
float,
int,
adagrad_update_prefetch_inlined>);
OPERATOR_SCHEMA(SparseAdagradFusedWithSparseLengthsWeightedSumGradientApprox)
.NumInputs(7)
.NumOutputs(3)
.EnforceInplace({{0, 0}, {1, 1}})
.SetDoc(R"DOC(
Approximately fused operator of
SparseLengthsIndicesInGradientWeightedSumWithMainInputGradient
(gradient of SparseLengthsWeightedSum) + SparseAdagrad, where weights are
positional weights computed with LengthsRangeFill + Gather pattern.
Given inputs (param, moment, indices, grad, lr), runs the sparse AdaGrad
update on (param, grad, moment[indices], lr), and returns (new_param,
new_moment) as in the dense case.
There's race condition w.r.t. ordering between reading params and writing to
param, hence the name Approx.
There're auxiliary inputs (aux_param) for which gradient is computed and
returns (aux_grad).
Yet additional input (lengths) is for fused
SparseLengthsIndicesInGradientWeightedSumWithMainInputGradient operator.
)DOC")
.Input(0, "param", "Parameters to be updated")
.Input(1, "moment", "Moment history")
.Input(2, "aux_param", "Auxiliary parameters to be updated")
.Input(
3,
"indices",
"Integer vector containing indices of the first dimension of param for the slices that are being updated")
.Input(4, "grad", "Gradient computed")
.Input(5, "lr", "learning rate")
.Input(
6,
"lengths",
"Non negative vector with sum of elements equal to indices length")
.Output(0, "output_param", "Updated parameters")
.Output(1, "output_moment", "Updated moment")
.Output(2, "aux_grad", "Auxiliary gradients")
.Arg("epsilon", "Default 1e-5");
REGISTER_CPU_OPERATOR(
SparseAdagradFusedWithSparseLengthsWeightedSumGradientApprox,
SparseAdagradFusedWithSparseLengthsWeightedSumGradientApproxOp<
float,
float,
int,
adagrad_update_prefetch_inlined>);
} // namespace caffe2
|
