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
|
#include "caffe2/operators/softsign_op.h"
#include "caffe2/utils/eigen_utils.h"
#include <algorithm>
#include <functional>
namespace caffe2 {
template <>
template <typename T>
bool SoftsignFunctor<CPUContext>::
operator()(const int N, const T* X, T* Y, CPUContext* /* context */) const {
ConstEigenVectorArrayMap<T> X_arr(X, N);
EigenVectorMap<T>(Y, N) = (T(1) + X_arr.abs()).inverse() * X_arr;
return true;
}
template <>
template <typename T>
bool SoftsignGradientFunctor<CPUContext>::Forward(
const std::vector<int>& X_dims,
const std::vector<int>& /* dY_dims */,
const T* X,
const T* dY,
T* dX,
CPUContext* /* context */) const {
const int size = std::accumulate(
X_dims.cbegin(), X_dims.cend(), 1, std::multiplies<int>());
ConstEigenVectorArrayMap<T> dY_arr(dY, size);
ConstEigenVectorArrayMap<T> X_arr(X, size);
EigenVectorMap<T>(dX, size) =
dY_arr * (T(1) + X_arr.abs()).square().inverse();
return true;
}
REGISTER_CPU_OPERATOR(
Softsign,
UnaryElementwiseOp<
TensorTypes<float>,
CPUContext,
SoftsignFunctor<CPUContext>>);
REGISTER_CPU_GRADIENT_OPERATOR(
SoftsignGradient,
BinaryElementwiseOp<
TensorTypes<float>,
CPUContext,
SoftsignGradientFunctor<CPUContext>>);
OPERATOR_SCHEMA(Softsign)
.NumInputs(1)
.NumOutputs(1)
.AllowInplace({{0, 0}})
.IdenticalTypeAndShape()
.SetDoc(R"DOC(
*Softsign* takes one input data tensor $X$ and produces one output data $Y,$ where the softsign function, $y = \frac{x}{1+ |x|}$, is applied to $X$ elementwise. This operation can be done in an in-place fashion too, by providing the same input and output blobs.
Github Links:
- https://github.com/pytorch/pytorch/blob/master/caffe2/operators/softsign_op.cc
<details>
<summary> <b>Example</b> </summary>
**Code**
```
workspace.ResetWorkspace()
op = core.CreateOperator(
"Softsign",
["X"],
["Y"],
)
workspace.FeedBlob("X", np.random.randn(3, 3).astype(np.float32))
print("X:\n", workspace.FetchBlob("X"), "\n")
workspace.RunOperatorOnce(op)
print("Y:\n", workspace.FetchBlob("Y"))
```
**Result**
```
X:
[[-1.3060539 0.7242748 -1.9907674 ]
[-0.64802396 -0.03244735 0.7455406 ]
[-0.298492 -0.5774271 2.8364444 ]]
Y:
[[-0.5663588 0.420046 -0.6656376 ]
[-0.39321268 -0.03142761 0.4271116 ]
[-0.2298759 -0.36605626 0.739342 ]]
```
</details>
)DOC")
.Input(0, "input", "Input data blob to be operated on.")
.Output(0, "output", "Output data blob with same shape as input")
.InheritOnnxSchema();
GRADIENT_OPERATOR_SCHEMA(SoftsignGradient)
.NumInputs(2)
.NumOutputs(1)
.AllowInplace({{1, 0}})
.SetDoc(R"DOC(
Calculates the softsign gradient (sgn(x)/(1+|x|)^2) of the given input tensor
element-wise.
)DOC")
.Input(0, "input", "1-D input tensor")
.Input(1, "input", "1-D input tensor")
.Output(
0,
"output",
"The softsign gradient (sgn(x)/(1+|x|)^2) values of the input tensor "
"computed element-wise");
namespace {
class GetSoftsignGradient : public GradientMakerBase {
using GradientMakerBase::GradientMakerBase;
std::vector<OperatorDef> GetGradientDefs() override {
CAFFE_ENFORCE(
I(0) != O(0),
"Cannot compute softsign gradient "
"if you choose to do an in-place calculation.");
return SingleGradientDef(
"SoftsignGradient",
"",
std::vector<std::string>{I(0), GO(0)},
std::vector<std::string>{GI(0)});
}
};
} // namespace
REGISTER_GRADIENT(Softsign, GetSoftsignGradient);
} // namespace caffe2
|
