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
|
#include "caffe2/core/init.h"
#include "caffe2/core/operator.h"
#include "caffe2/core/tensor.h"
#include "caffe2/utils/math.h"
#include "caffe2/utils/proto_utils.h"
#include "gtest/gtest.h"
#include <cmath>
#include <random>
namespace caffe2 {
void AddConstInput(const vector<int64_t>& shape,
const float value,
const string& name,
Workspace* ws) {
DeviceOption option;
CPUContext context(option);
Blob* blob = ws->CreateBlob(name);
auto* tensor = BlobGetMutableTensor(blob, CPU);
tensor->Resize(shape);
math::Set<float, CPUContext>(
tensor->numel(), value, tensor->template mutable_data<float>(), &context);
}
void AddNoiseInput(const vector<int64_t>& shape,
const string& name,
Workspace* ws) {
DeviceOption option;
CPUContext context(option);
Blob* blob = ws->CreateBlob(name);
auto* tensor = BlobGetMutableTensor(blob, CPU);
tensor->Resize(shape);
math::RandGaussian<float, CPUContext>(
tensor->numel(),
0.0f,
10.0f,
tensor->template mutable_data<float>(),
&context);
}
inline float relativeError(float a, float b) {
return std::abs(a - b) / (0.5f * (std::abs(a) + std::abs(b)));
}
void compare(int N, int inputC, int H, int W,
int outputC,
int kernelH, int kernelW, int strideH, int strideW,
int padT, int padL, int padB, int padR,
int adjH, int adjW,
float maxRelErr, float absErrForRelErrFailure) {
LOG(INFO) <<
"running N " << N << " inputC " << inputC << " H " << H << " W " << W <<
" outputC " << outputC <<
" kernelH " << kernelH << " kernelW " << kernelW <<
" strideH " << strideH << " strideW " << strideW <<
" padT " << padT << " padL " << padL <<
" padB " << padB << " padR " << padR <<
" adjH " << adjH << " adjW " << adjW;
Workspace ws;
OperatorDef def1;
def1.set_name("test");
def1.set_type("ConvTranspose");
def1.set_engine("MOBILE");
def1.add_input("X");
def1.add_input("W");
def1.add_input("B");
def1.add_output("Y1");
def1.add_arg()->CopyFrom(MakeArgument("kernel_h", kernelH));
def1.add_arg()->CopyFrom(MakeArgument("kernel_w", kernelW));
def1.add_arg()->CopyFrom(MakeArgument("stride_h", strideH));
def1.add_arg()->CopyFrom(MakeArgument("stride_w", strideW));
def1.add_arg()->CopyFrom(MakeArgument("pad_t", padT));
def1.add_arg()->CopyFrom(MakeArgument("pad_l", padL));
def1.add_arg()->CopyFrom(MakeArgument("pad_b", padB));
def1.add_arg()->CopyFrom(MakeArgument("pad_r", padR));
def1.add_arg()->CopyFrom(MakeArgument("adj_h", adjH));
def1.add_arg()->CopyFrom(MakeArgument("adj_w", adjW));
AddNoiseInput(vector<int64_t>{N, inputC, H, W}, "X", &ws);
AddNoiseInput(vector<int64_t>{inputC, outputC, kernelH, kernelW}, "W", &ws);
AddNoiseInput(vector<int64_t>{outputC}, "B", &ws);
unique_ptr<OperatorBase> op1(CreateOperator(def1, &ws));
EXPECT_NE(nullptr, op1.get());
OperatorDef def2;
def2.set_name("test");
def2.set_type("ConvTranspose");
def2.add_input("X");
def2.add_input("W");
def2.add_input("B");
def2.add_output("Y2");
def2.add_arg()->CopyFrom(MakeArgument("kernel_h", kernelH));
def2.add_arg()->CopyFrom(MakeArgument("kernel_w", kernelW));
def2.add_arg()->CopyFrom(MakeArgument("stride_h", strideH));
def2.add_arg()->CopyFrom(MakeArgument("stride_w", strideW));
def2.add_arg()->CopyFrom(MakeArgument("pad_t", padT));
def2.add_arg()->CopyFrom(MakeArgument("pad_l", padL));
def2.add_arg()->CopyFrom(MakeArgument("pad_b", padB));
def2.add_arg()->CopyFrom(MakeArgument("pad_r", padR));
def2.add_arg()->CopyFrom(MakeArgument("adj_h", adjH));
def2.add_arg()->CopyFrom(MakeArgument("adj_w", adjW));
unique_ptr<OperatorBase> op2(CreateOperator(def2, &ws));
EXPECT_NE(nullptr, op2.get());
EXPECT_TRUE(op1->Run());
Blob* Y1blob = ws.GetBlob("Y1");
EXPECT_NE(nullptr, Y1blob);
auto& Y1 = Y1blob->Get<TensorCPU>();
EXPECT_TRUE(op2->Run());
Blob* Y2blob = ws.GetBlob("Y2");
EXPECT_NE(nullptr, Y2blob);
auto& Y2 = Y2blob->Get<TensorCPU>();
// Compare all output points
for (int n = 0; n < Y1.dim32(0); ++n) {
for (int c = 0; c < Y1.dim32(1); ++c) {
for (int h = 0; h < Y1.dim32(2); ++h) {
for (int w = 0; w < Y1.dim32(3); ++w) {
int offset =
n * Y1.dim32(1) * Y1.dim32(2) * Y1.dim32(3) +
c * Y1.dim32(2) * Y1.dim32(3) +
h * Y1.dim32(3) +
w;
auto v1 = Y1.data<float>()[offset];
auto v2 = Y2.data<float>()[offset];
float relErr = relativeError(v1, v2);
float absErr = std::abs(v1 - v2);
// For small values / small difference, the relative error
// can be huge but the absolute error will be small
EXPECT_TRUE(relErr <= maxRelErr ||
(absErr <= absErrForRelErrFailure)) <<
v1 << " " << v2 << " (rel err " << relErr << ") " <<
"(" << n << " " << c << " " << h << " " << w << ") " <<
"running N " << N << " inputC " << inputC <<
" H " << H << " W " << W <<
" outputC " << outputC <<
" kernelH " << kernelH << " kernelW " << kernelW <<
" strideH " << strideH << " strideW " << strideW <<
" padT " << padT << " padL " << padL <<
" padB " << padB << " padR " << padR <<
" adjH " << adjH << " adjW " << adjW;
}
}
}
}
}
} // namespace caffe2
int randInt(int a, int b) {
static std::random_device rd;
static std::mt19937 gen(rd());
return std::uniform_int_distribution<int>(a, b)(gen);
}
// TODO(#14383029) cblas_sgemm not yet implemented on limited mobile cases.
#if (defined(__ARM_NEON__) || defined(__ARM_NEON))
TEST(ConvTransposeMobile, Test) {
for (int i = 0; i < 10; ++i) {
int n = randInt(1, 3);
int planesIn = randInt(1, 10);
int h = randInt(10, 200);
int w = randInt(10, 200);
int planesOut = randInt(1, 10);
int kernelH = randInt(2, 5);
int kernelW = randInt(2, 5);
int strideH = randInt(1, 4);
int strideW = randInt(1, 4);
int padT = randInt(0, 3);
int padB = randInt(0, 3);
int padL = 0;
int padR = 0;
int adjH = randInt(0, 3);
if (adjH >= strideH) { adjH = strideH - 1; }
int adjW = randInt(0, 3);
if (adjW >= strideW) { adjW = strideW - 1; }
caffe2::compare(n, planesIn, h, w,
planesOut,
kernelH, kernelW,
strideH, strideW,
padT, padL, padB, padR,
adjH, adjW, 0.002f, 0.001f);
}
}
#endif
|
