scale_layer.cpp
6.78 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
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
// 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 <opencv2/imgproc.hpp>
#include <opencv2/dnn/shape_utils.hpp>
namespace cv
{
namespace dnn
{
class ScaleLayerInt8Impl CV_FINAL : public ScaleLayerInt8
{
public:
Mat weights, bias;
ScaleLayerInt8Impl(const LayerParams& params)
{
setParamsFrom(params);
hasBias = params.get<bool>("bias_term", false);
axis = params.get<int>("axis", 1);
hasWeights = false;
output_sc = params.get<float>("scales");
output_zp = params.get<int>("zeropoints");
DictValue inpSc = params.get("input_scales");
DictValue inpZp = params.get("input_zeropoints");
for (int i = 0; i < inpSc.size(); i++)
{
inp_sc.push_back(inpSc.get<float>(i));
inp_zp.push_back(inpZp.get<int>(i));
}
}
bool getMemoryShapes(const std::vector<MatShape> &inputs,
const int requiredOutputs,
std::vector<MatShape> &outputs,
std::vector<MatShape> &internals) const CV_OVERRIDE
{
outputs.assign(1, inputs[0]);
return true;
}
virtual void finalize(InputArrayOfArrays inputs_arr, OutputArrayOfArrays) CV_OVERRIDE
{
std::vector<Mat> inputs;
inputs_arr.getMatVector(inputs);
hasWeights = blobs.size() == 2 || (blobs.size() <= 1 && !hasBias);
CV_Assert((inputs.size() == 2 && blobs.empty()) || blobs.size() == (int)hasWeights + (int)hasBias);
if (!blobs.empty())
{
Mat w = hasWeights ? blobs[0] : Mat::ones(blobs[0].size(), CV_32F);
Mat b = hasBias ? blobs.back() : Mat::zeros(blobs.back().size(), CV_32F);
w = w.reshape(1, 1);
b = b.reshape(1, 1);
w.convertTo(weights, CV_32F, inp_sc[0]/output_sc);
addWeighted(b, 1.0/output_sc, weights, -inp_zp[0], output_zp, bias, CV_32F);
}
else
{
// initialized during forward()
weights = Mat(); bias = Mat();
}
}
virtual bool supportBackend(int backendId) CV_OVERRIDE
{
return backendId == DNN_BACKEND_OPENCV;
}
bool setActivation(const Ptr<ActivationLayer>& layer) CV_OVERRIDE
{
Ptr<ActivationLayerInt8> activ_int8 = layer.dynamicCast<ActivationLayerInt8>();
if (!activ_int8.empty())
{
return activ_int8->blobs.empty();
}
return false;
}
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;
inputs_arr.getMatVector(inputs);
outputs_arr.getMatVector(outputs);
Mat &inpBlob = inputs[0];
Mat &outBlob = outputs[0];
if (blobs.empty())
{
CV_Assert(inp_sc.size() == 2 && inp_zp.size() == 2);
Mat inp_dequantized, w, b;
inputs[1].reshape(1, 1).convertTo(inp_dequantized, CV_32F, inp_sc[1], -(inp_sc[1]*inp_zp[1]));
w = hasWeights ? inp_dequantized : Mat::ones(inp_dequantized.size(), CV_32F);
b = hasBias ? inp_dequantized : Mat::zeros(inp_dequantized.size(), CV_32F);
w.convertTo(weights, CV_32F, inp_sc[0]/output_sc);
addWeighted(b, 1.0/output_sc, weights, -inp_zp[0], output_zp, bias, CV_32F);
}
MatShape inpShape = shape(inpBlob);
const int numWeights = weights.total();
CV_Assert(numWeights != 0);
CV_CheckEQ(weights.total(), bias.total(), "Incompatible weights/bias blobs");
int endAxis;
for (endAxis = axis + 1; endAxis <= inpBlob.dims; ++endAxis)
{
if (total(inpShape, axis, endAxis) == numWeights)
break;
}
CV_Assert(total(inpShape, axis, endAxis) == numWeights);
CV_CheckTypeEQ(inpBlob.type(), CV_8SC1, ""); CV_CheckTypeEQ(outBlob.type(), CV_8SC1, "");
int numSlices = total(inpShape, 0, axis);
int8_t* inpData = (int8_t*)inpBlob.data;
int8_t* outData = (int8_t*)outBlob.data;
if (endAxis != inpBlob.dims)
{
float* weightsData = (float*)weights.data;
float* biasesData = (float*)bias.data;
int spatialSize = total(inpShape, endAxis); // spatialSize != 1
for (int i = 0; i < numSlices; ++i)
{
for (int j = 0; j < numWeights; ++j)
{
float w = weightsData[j];
float b = biasesData[j];
Mat inpSlice(1, spatialSize, CV_8S, inpData);
Mat outSlice(1, spatialSize, CV_8S, outData);
inpSlice.convertTo(outSlice, CV_8S, w, b);
inpData += spatialSize;
outData += spatialSize;
}
}
}
else
{
for (int i = 0; i < numSlices; ++i)
{
Mat inpSlice(1, numWeights, CV_8S, inpData);
Mat outSlice(1, numWeights, CV_8S, outData);
multiply(inpSlice, weights, outSlice, 1.0, CV_8S);
add(outSlice, bias, outSlice, Mat(), CV_8S);
inpData += numWeights;
outData += numWeights;
}
}
}
void getScaleShift(Mat& scale, Mat& shift) const CV_OVERRIDE
{
scale = (hasWeights && !blobs.empty()) ? blobs[0] : Mat();
shift = (hasBias && !blobs.empty()) ? blobs.back() : Mat();
}
void getScaleZeropoint(float& scale, int& zeropoint) const CV_OVERRIDE
{
scale = output_sc;
zeropoint = output_zp;
}
virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
const std::vector<MatShape> &outputs) const CV_OVERRIDE
{
CV_UNUSED(outputs); // suppress unused variable warning
long flops = 0;
for(int i = 0; i < inputs.size(); i++)
{
flops += 2*total(inputs[i]);
}
return flops;
}
private:
bool hasWeights;
std::vector<float> inp_sc;
std::vector<int> inp_zp;
};
Ptr<ScaleLayerInt8> ScaleLayerInt8::create(const LayerParams& params)
{
return Ptr<ScaleLayerInt8>(new ScaleLayerInt8Impl(params));
}
Ptr<Layer> ShiftLayerInt8::create(const LayerParams& params)
{
LayerParams scaleParams = params;
scaleParams.type = "ScaleInt8";
scaleParams.set("bias_term", true);
scaleParams.set("axis", 0);
return Ptr<ScaleLayerInt8>(new ScaleLayerInt8Impl(scaleParams));
}
} // namespace dnn
} // namespace cv