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
|
#include "caffe2/operators/rowmul_op.h"
namespace caffe2 {
namespace {
REGISTER_CPU_OPERATOR(ReduceTailSum, ReduceTailSumOp<float, CPUContext>);
REGISTER_CPU_OPERATOR(RowMul, RowMulOp<float, CPUContext>);
OPERATOR_SCHEMA(ReduceTailSum)
.NumInputs(1, 1)
.NumOutputs(1)
.SetDoc(R"DOC(
Reduce the tailing dimensions
)DOC")
.Input(0, "mat", "The matrix")
.Output(0, "output", "Output");
OPERATOR_SCHEMA(RowMul)
.NumInputs(2, 2)
.NumOutputs(1)
.SetDoc(R"DOC(
Given a matrix A and column vector w, the output is the multiplication of row i
of A and element i of w, e.g. C[i][j] = A[i][j] * w[i]. This operator should be
deprecated when the gradient operator of Mul with broadcast is implemented.
)DOC")
.Input(0, "mat", "The matrix")
.Input(1, "w", "The column vector")
.Output(0, "output", "Output");
class GetRowMulGradient : public GradientMakerBase {
using GradientMakerBase::GradientMakerBase;
vector<OperatorDef> GetGradientDefs() override {
return vector<OperatorDef>{
CreateOperatorDef(
"RowMul", "", vector<string>{GO(0), I(1)}, vector<string>{GI(0)}),
CreateOperatorDef(
"Mul",
"",
vector<string>{GO(0), I(0)},
vector<string>{GI(1) + "before_aggregate"}),
CreateOperatorDef(
"ReduceTailSum",
"",
vector<string>{GI(1) + "before_aggregate"},
vector<string>{GI(1)})};
}
};
REGISTER_GRADIENT(RowMul, GetRowMulGradient);
} // namespace
} // namespace caffe2
|
