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
|
#ifndef CAFFE2_OPERATORS_NORMALIZE_OP_H_
#define CAFFE2_OPERATORS_NORMALIZE_OP_H_
#include "caffe2/core/context.h"
#include "caffe2/core/operator.h"
#include "caffe2/utils/eigen_utils.h"
#include "caffe2/utils/math.h"
#define KEPS 1e-12f
namespace caffe2 {
template <typename T, class Context>
class NormalizeOp final : public Operator<Context> {
public:
USE_OPERATOR_CONTEXT_FUNCTIONS;
template <class... Args>
explicit NormalizeOp(Args&&... args)
: Operator<Context>(std::forward<Args>(args)...) {}
bool RunOnDevice() override {
const auto& x = Input(0);
const auto* xData = x.template data<T>();
auto* y = Output(0, x.sizes(), at::dtype<T>());
auto* yData = y->template mutable_data<T>();
const auto canonical_axis = x.canonical_axis_index(
this->template GetSingleArgument<int>("axis", -1));
const int64_t m = x.dim(canonical_axis);
const size_t n = x.numel() / m;
const size_t sf = x.size_from_dim(canonical_axis + 1);
DoNormalize(xData, yData, m, n, sf);
return true;
}
private:
const T kEps_ = KEPS;
void DoNormalize(
const T* xData,
T* yData,
const int m,
const int n,
const int sf) {
using InnerStride = Eigen::InnerStride<Eigen::Dynamic>;
using StridedVec =
Eigen::Map<Eigen::Matrix<T, 1, Eigen::Dynamic>, 0, InnerStride>;
using ConstStridedVec =
Eigen::Map<const Eigen::Matrix<T, 1, Eigen::Dynamic>, 0, InnerStride>;
for (const auto i : c10::irange(n)) {
auto base = (i / sf) * sf * m + (i % sf);
ConstStridedVec xVec(xData + base, 1, m, InnerStride(sf));
auto norm = xVec.template lpNorm<2>();
norm = std::max(norm, kEps_);
StridedVec yVec(yData + base, 1, m, InnerStride(sf));
yVec = xVec / norm;
}
}
};
template <typename T, class Context>
class NormalizeGradientOp final : public Operator<Context> {
public:
USE_OPERATOR_CONTEXT_FUNCTIONS;
template <class... Args>
explicit NormalizeGradientOp(Args&&... args)
: Operator<Context>(std::forward<Args>(args)...) {}
bool RunOnDevice() override {
const auto& x = Input(0);
const auto& gOut = Input(GRAD_OUT);
auto* gIn = Output(GRAD_IN, gOut.sizes(), at::dtype<T>());
const auto* xData = x.template data<T>();
const auto* gOutData = gOut.template data<T>();
auto* gInData = gIn->template mutable_data<T>();
const auto canonical_axis = x.canonical_axis_index(
this->template GetSingleArgument<int>("axis", -1));
const int m = x.dim32(canonical_axis);
const int n = x.numel() / m;
const int sf = x.size_from_dim(canonical_axis + 1);
DoNormalize(xData, gOutData, gInData, m, n, sf);
return true;
}
private:
const T kEps_ = KEPS;
void DoNormalize(
const T* xData,
const T* gOutData,
T* gInData,
const int m,
const int n,
const int sf);
INPUT_TAGS(INPUT, GRAD_OUT);
OUTPUT_TAGS(GRAD_IN);
};
} // namespace caffe2
#endif // CAFFE2_OPERATORS_NORMALIZE_OP_H_
|
