softmax_layer.cpp
6.07 KB
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
// This file is part of OpenCV project.
// It is subject to the license terms in the LICENSE file found in the top-level directory
// of this distribution and at http://opencv.org/license.html.
#include "../precomp.hpp"
#include "layers_common.hpp"
#include <algorithm>
#include <stdlib.h>
namespace cv
{
namespace dnn
{
class SoftMaxLayerInt8Impl CV_FINAL : public SoftmaxLayerInt8
{
public:
SoftMaxLayerInt8Impl(const LayerParams& params)
{
axisRaw = params.get<int>("axis", 1);
logSoftMax = params.get<bool>("log_softmax", false);
output_sc = params.get<float>("scales");
output_zp = params.get<int>("zeropoints");
setParamsFrom(params);
}
bool getMemoryShapes(const std::vector<MatShape> &inputs,
const int requiredOutputs,
std::vector<MatShape> &outputs,
std::vector<MatShape> &internals) const CV_OVERRIDE
{
bool inplace = Layer::getMemoryShapes(inputs, requiredOutputs, outputs, internals);
MatShape shape = inputs[0];
int cAxis = normalize_axis(axisRaw, shape.size());
shape[cAxis] = 1;
internals.assign(1, shape);
return inplace;
}
virtual bool supportBackend(int backendId) CV_OVERRIDE
{
return backendId == DNN_BACKEND_OPENCV;
}
virtual bool tryFuse(Ptr<Layer>& top) CV_OVERRIDE
{
Ptr<DequantizeLayer> dequantize_layer = top.dynamicCast<DequantizeLayer>();
return !dequantize_layer.empty() && preferableTarget != DNN_TARGET_OPENCL_FP16;
}
void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr) CV_OVERRIDE
{
CV_TRACE_FUNCTION();
CV_TRACE_ARG_VALUE(name, "name", name.c_str());
std::vector<Mat> inputs, outputs, internals;
inputs_arr.getMatVector(inputs);
outputs_arr.getMatVector(outputs);
internals_arr.getMatVector(internals);
const Mat &src = inputs[0];
Mat &dst = outputs[0];
int axis = normalize_axis(axisRaw, src.dims);
size_t outerSize = src.total(0, axis), channels = src.size[axis],
innerSize = src.total(axis + 1);
CV_Assert(src.type() == CV_8S && (dst.type() == CV_8S || dst.type() == CV_32F));
CV_Assert(src.isContinuous() && dst.isContinuous());
size_t outerStep = src.total(axis);
size_t cnStep = src.total(axis + 1);
const int8_t *srcPtr = src.ptr<int8_t>();
const float *expPtr = blobs[0].ptr<float>();
if (dst.type() == CV_32F)
{
float *dstPtr = dst.ptr<float>();
for (size_t outerDim = 0; outerDim < outerSize; outerDim++)
{
size_t srcOffset = outerDim * outerStep;
std::vector<float> expSum(innerSize, 0.f);
// sum exp along axis
for (size_t cnDim = 0; cnDim < channels; cnDim++)
{
const int offset = srcOffset + cnDim * cnStep;
for (size_t i = 0; i < innerSize; i++)
expSum[i] += expPtr[srcPtr[offset + i] + 128];
}
// divide by computed sum
for (size_t cnDim = 0; cnDim < channels; cnDim++)
{
const int offset = srcOffset + cnDim * cnStep;
for (size_t i = 0; i < innerSize; i++)
dstPtr[offset + i] = expPtr[srcPtr[offset + i] + 128]/expSum[i];
}
if (logSoftMax)
{
for (size_t cnDim = 0; cnDim < channels; cnDim++)
{
const int offset = srcOffset + cnDim * cnStep;
for (size_t i = 0; i < innerSize; i++)
dstPtr[offset + i] = log(dstPtr[offset + i]);
}
}
}
}
else
{
const float inv_scale = 1.f/output_sc;
int8_t *dstPtr = dst.ptr<int8_t>();
for (size_t outerDim = 0; outerDim < outerSize; outerDim++)
{
size_t srcOffset = outerDim * outerStep;
std::vector<float> expSum(innerSize, 0.f);
// sum exp along axis
for (size_t cnDim = 0; cnDim < channels; cnDim++)
{
const int offset = srcOffset + cnDim * cnStep;
for (size_t i = 0; i < innerSize; i++)
expSum[i] += expPtr[srcPtr[offset + i] + 128];
}
// divide by computed sum and quantize to int8
if (logSoftMax)
{
for (size_t cnDim = 0; cnDim < channels; cnDim++)
{
const int offset = srcOffset + cnDim * cnStep;
for (size_t i = 0; i < innerSize; i++)
dstPtr[offset + i] = saturate_cast<int8_t>(output_zp + std::round(inv_scale*log(expPtr[srcPtr[offset + i] + 128]/expSum[i])));
}
}
else
{
for (size_t cnDim = 0; cnDim < channels; cnDim++)
{
const int offset = srcOffset + cnDim * cnStep;
for (size_t i = 0; i < innerSize; i++)
dstPtr[offset + i] = saturate_cast<int8_t>(output_zp + std::round(inv_scale*(expPtr[srcPtr[offset + i] + 128]/expSum[i])));
}
}
}
}
}
int64 getFLOPS(const std::vector<MatShape> &inputs,
const std::vector<MatShape> &outputs) const CV_OVERRIDE
{
CV_UNUSED(outputs); // suppress unused variable warning
int64 flops = 0;
for (int i = 0; i < inputs.size(); i++)
{
flops += 4*total(inputs[i]);
}
return flops;
}
int axisRaw;
};
Ptr<SoftmaxLayerInt8> SoftmaxLayerInt8::create(const LayerParams& params)
{
return Ptr<SoftmaxLayerInt8>(new SoftMaxLayerInt8Impl(params));
}
}
}