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
|
#include "caffe2/operators/softmax_op.h"
#include "caffe2/operators/softmax_shared.h"
namespace caffe2 {
// Implementation for the CPU context.
template <>
bool SoftmaxOp<float, CPUContext>::RunOnDevice() {
auto& X = Input(0);
const auto canonical_axis = X.canonical_axis_index(axis_);
const int N = X.size_to_dim(canonical_axis);
const int D = X.size_from_dim(canonical_axis);
auto* Y = Output(0, X.sizes(), at::dtype<float>());
float* Ydata = Y->template mutable_data<float>();
// First, get scales
if (!scale_.defined()) {
scale_ = caffe2::empty({N}, at::dtype<float>().device(CPU));
} else if (scale_.numel() != N) {
scale_.Resize(N);
}
if (!rowmax_.defined()) {
rowmax_ = caffe2::empty({N}, at::dtype<float>().device(CPU));
} else if (rowmax_.numel() != N) {
rowmax_.Resize(N);
}
if (!sum_multiplier_.defined()) {
sum_multiplier_ = caffe2::empty({D}, at::dtype<float>().device(CPU));
math::Set<float, CPUContext>(D, 1.f, sum_multiplier_.mutable_data<float>(), &context_);
} else if (sum_multiplier_.numel() != D) {
sum_multiplier_.Resize(D);
math::Set<float, CPUContext>(D, 1.f, sum_multiplier_.mutable_data<float>(), &context_);
}
SoftmaxCPU(
context_,
N,
D,
X.data<float>(),
Ydata,
scale_.mutable_data<float>(),
sum_multiplier_.data<float>(),
false,
rowmax_.mutable_data<float>());
return true;
}
// Implementation for the CPU context.
template <>
bool SoftmaxGradientOp<float, CPUContext>::RunOnDevice() {
auto& Y = Input(0);
auto& dY = Input(1);
const auto canonical_axis = Y.canonical_axis_index(axis_);
const int64_t N = Y.size_to_dim(canonical_axis);
const int64_t D = Y.size_from_dim(canonical_axis);
// First, get scales
if (!scale_.defined()) {
scale_ = caffe2::empty({N}, at::dtype<float>().device(CPU));
} else if (scale_.numel() != N) {
scale_.Resize(N);
}
if (!sum_multiplier_.defined()) {
sum_multiplier_ = caffe2::empty({D}, at::dtype<float>().device(CPU));
math::Set<float, CPUContext>(D, 1.f, sum_multiplier_.mutable_data<float>(), &context_);
} else if (sum_multiplier_.numel() != D) {
sum_multiplier_.Resize(D);
math::Set<float, CPUContext>(D, 1.f, sum_multiplier_.mutable_data<float>(), &context_);
}
auto* dX = Output(0, Y.sizes(), at::dtype<float>());
const float* Ydata = Y.data<float>();
const float* dYdata = dY.data<float>();
float* dXdata = dX->mutable_data<float>();
if (N == 0) {
return true;
}
context_.CopySameDevice<float>(Y.numel(), dYdata, dXdata);
float* scaledata = scale_.mutable_data<float>();
for (int i = 0; i < N; ++i) {
math::Dot<float, CPUContext>(D, Ydata + i * D, dYdata + i * D,
scaledata + i, &context_);
}
math::Gemm<float, CPUContext>(CblasNoTrans, CblasNoTrans, N, D, 1, -1,
scaledata, sum_multiplier_.data<float>(), 1,
dXdata, &context_);
math::Mul<float, CPUContext>(Y.numel(), dXdata, Ydata, dXdata, &context_);
return true;
}
REGISTER_CPU_OPERATOR(Softmax, SoftmaxOp<float, CPUContext>);
REGISTER_CPU_GRADIENT_OPERATOR(
SoftmaxGradient,
SoftmaxGradientOp<float, CPUContext>);
OPERATOR_SCHEMA(Softmax)
.NumInputs(1)
.NumOutputs(1)
.IdenticalTypeAndShape()
.SetDoc(R"DOC(
Applies the Softmax function to an n-dimensional input Tensor rescaling them so
that the elements of the n-dimensional output Tensor lie in the range (0,1) and
sum to 1. The softmax operator is typically the last layer in a classifier network,
as its output can be interpreted as confidence probabilities of an input belonging
to each class. The input is a 2-D tensor (Tensor) of size (batch_size x
input_feature_dimensions). The output tensor has the same shape and contains the
softmax normalized values of the corresponding input. The softmax function is
defined as follows:
$$softmax(x_i) = \frac{\exp(x_i)}{\sum_{j} \exp(x_j)}$$
The input does not need to explicitly be a 2D vector; rather, it will be coerced
into one. For an arbitrary n-dimensional tensor `X` in
$[a_0, a_1, ..., a_{k-1}, a_k, ..., a_{n-1}]$, where k is the `axis` provided,
then `X` will be coerced into a 2-dimensional tensor with dimensions
$[(a_0 * ... * a_{k-1}), (a_k * ... * a_{n-1})]$. For the default case where
`axis`=1, the `X` tensor will be coerced into a 2D tensor of dimensions
$[a_0, (a_1 * ... * a_{n-1})]$, where $a_0$ is often the batch size. In this
situation, we must have $a_0 = N$ and $a_1 * ... * a_{n-1} = D$. Each of these
dimensions must be matched correctly, or else the operator will throw errors.
Github Links:
- https://github.com/pytorch/pytorch/blob/master/caffe2/operators/softmax_op.h
- https://github.com/pytorch/pytorch/blob/master/caffe2/operators/softmax_op.cc
<details>
<summary> <b>Example</b> </summary>
**Code**
```
workspace.ResetWorkspace()
op = core.CreateOperator(
"Softmax",
["X"],
["Y"]
)
workspace.FeedBlob("X", np.random.randn(1, 5).astype(np.float32))
print("input:", workspace.FetchBlob("X"))
workspace.RunOperatorOnce(op)
print("softmax:", workspace.FetchBlob("Y"))
```
**Result**
```
input: [[ 0.0417839 0.61960053 -0.23150268 -0.64389366 -3.0000346 ]]
softmax: [[0.24422921 0.43525138 0.18582782 0.12303016 0.01166145]]
```
</details>
)DOC")
.Arg(
"axis",
"*(type: int; default: 1)* Axis of the inputs when coerced to 2D matrix.")
.Input(
0,
"X",
"*(type: Tensor`<float>`)* Input tensor that's coerced into a 2D matrix of size (NxD) as described above.")
.Output(
0,
"Y",
"*(type: Tensor`<float>`)* The softmax normalized output tensor with the same shape as input tensor.")
.InheritOnnxSchema();
// Input: Y, dY. Output: dX
GRADIENT_OPERATOR_SCHEMA(SoftmaxGradient).NumInputs(2).NumOutputs(1);
class GetSoftmaxGradient : public GradientMakerBase {
using GradientMakerBase::GradientMakerBase;
vector<OperatorDef> GetGradientDefs() override {
return SingleGradientDef(
def_.type() + "Gradient", "",
vector<string>{O(0), GO(0)},
vector<string>{GI(0)});
}
};
REGISTER_GRADIENT(Softmax, GetSoftmaxGradient);
REGISTER_GRADIENT(SoftmaxFp16, GetSoftmaxGradient);
} // namespace caffe2
|
