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
|
#include <functional>
#include <memory>
#include <vector>
#include <gtest/gtest.h>
#include "tensorflow/core/framework/allocator.h"
#include "tensorflow/core/framework/fake_input.h"
#include "tensorflow/core/framework/graph.pb.h"
#include "tensorflow/core/framework/node_def_builder.h"
#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/framework/tensor_testutil.h"
#include "tensorflow/core/framework/types.h"
#include "tensorflow/core/kernels/ops_testutil.h"
#include "tensorflow/core/kernels/ops_util.h"
#include "tensorflow/core/lib/core/status_test_util.h"
#include "tensorflow/core/lib/random/simple_philox.h"
#include "tensorflow/core/public/tensor.h"
namespace tensorflow {
static const float tol_ = 1e-4;
class LRNFloatTest : public OpsTestBase {
protected:
LRNFloatTest() : philox_(123, 17), rand_(&philox_) { RequireDefaultOps(); }
int GetIntAttr(const string& name) {
int value;
TF_CHECK_OK(GetNodeAttr(*node_def(), name, &value));
return value;
}
float GetFloatAttr(const string& name) {
float value;
TF_CHECK_OK(GetNodeAttr(*node_def(), name, &value));
return value;
}
bool Compare() {
const auto& input = GetInput(0);
const int64 batch_size = input.dim_size(0);
const int64 rows = input.dim_size(1);
const int64 cols = input.dim_size(2);
const int64 depth = input.dim_size(3);
const int64 rest = cols * rows * batch_size;
const int64 depth_radius = GetIntAttr("depth_radius");
const float bias = GetFloatAttr("bias");
const float alpha = GetFloatAttr("alpha");
const float beta = GetFloatAttr("beta");
Eigen::Tensor<float, 4, Eigen::RowMajor> expected(batch_size, rows, cols,
depth);
auto out = expected.reshape(Eigen::DSizes<int64, 2>{rest, depth});
auto in = input.shaped<float, 2>({rest, depth});
for (int64 i = 0; i < rest; ++i) {
Eigen::Tensor<float, 1, Eigen::RowMajor> out_col(depth);
for (int64 d = 0; d < depth; ++d) {
float denom = 0.0f;
for (int64 r = std::max(0ll, d - depth_radius);
r < std::min(depth, d + depth_radius + 1); ++r) {
denom += in(i, r) * in(i, r);
}
denom = std::pow(denom * alpha + bias, beta);
out_col(d) = in(i, d) / denom;
}
out.chip<0>(i) = out_col;
}
auto actual = GetOutput(0)->tensor<float, 4>();
Eigen::Tensor<float, 0, Eigen::RowMajor> sum =
((expected - actual).abs() > actual.constant(tol_))
.select(actual.constant(1), actual.constant(0))
.sum();
return sum() == 0;
}
random::PhiloxRandom philox_;
random::SimplePhilox rand_;
};
TEST_F(LRNFloatTest, Depth96) {
ASSERT_OK(NodeDefBuilder("lrn_op", "LRN")
.Input(FakeInput())
.Attr("depth_radius", 5)
.Attr("bias", 1.0f)
.Attr("alpha", 0.1f)
.Attr("beta", 2.0f)
.Finalize(node_def()));
ASSERT_OK(InitOp());
AddInput<float>(TensorShape({1, 1, 1, 96}),
[this](int i) -> float { return i + 1; });
ASSERT_OK(RunOpKernel());
auto actual = GetOutput(0)->tensor<float, 4>();
// Output for Node 0 with Value 1:
// 1 / (1 + 0.1*(1^2 + 2^2 + 3^2 + 4^2 + 5^2 + 6^2))^2
EXPECT_NEAR(1. / (10.1 * 10.1), actual(0, 0, 0, 0), tol_);
// Output for Node 5 with Value 6:
// 6 / (1 + 0.1*(1^2 + 2^2 + 3^2 + 4^2 + 5^2 + 6^2 ... + 11^2))^2
EXPECT_NEAR(6. / (51.6 * 51.6), actual(0, 0, 0, 5), tol_);
// Output for Node 63 with value 64:
// 64 / (1 + 0.1*(59^2 + 60^2 + 61^2 + 62^2 + 63^2 + 64^2))^2
EXPECT_NEAR(64. / (2272.1 * 2272.1), actual(0, 0, 0, 63), tol_);
// Output for Node 64 with value 65:
// 65 / (1 + 0.1*(65^2 + 66^2 + 67^2 + 68^2 + 69^2 + 70^2))^2
EXPECT_NEAR(65. / (2736.5 * 2736.5), actual(0, 0, 0, 64), tol_);
// Output for Node 95 with value 96:
// 96 / (1 + 0.1*(91^2 + 92^2 + 93^2 + 94^2 + 95^2 + 96^2))^2
EXPECT_NEAR(96. / (5248.1 * 5248.1), actual(0, 0, 0, 95), tol_);
EXPECT_TRUE(Compare());
}
TEST_F(LRNFloatTest, Depth16) {
ASSERT_OK(NodeDefBuilder("lrn_op", "LRN")
.Input(FakeInput())
.Attr("depth_radius", 5)
.Attr("bias", 1.0f)
.Attr("alpha", 0.1f)
.Attr("beta", 2.0f)
.Finalize(node_def()));
ASSERT_OK(InitOp());
AddInput<float>(TensorShape({1, 1, 1, 16}),
[this](int i) -> float { return i + 1; });
ASSERT_OK(RunOpKernel());
auto actual = GetOutput(0)->tensor<float, 4>();
// Output for Node 0 with Value 1:
// 1 / (1 + 0.1*(1^2 + 2^2 + 3^2 + 4^2 + 5^2 + 6^2))^2
EXPECT_NEAR(1. / (10.1 * 10.1), actual(0, 0, 0, 0), tol_);
// Output for Node 5 with Value 6:
// 6 / (1 + 0.1*(1^2 + 2^2 + 3^2 + 4^2 + 5^2 + 6^2 ... + 11^2))^2
EXPECT_NEAR(6. / (51.6 * 51.6), actual(0, 0, 0, 5), tol_);
// Output for Node 15 with value 16:
// 16 / (1 + 0.1*(11^2 + 12^2 + 13^2 + 14^2 + 15^2 + 16^2))^2
EXPECT_NEAR(16. / (112.1 * 112.1), actual(0, 0, 0, 15), tol_);
EXPECT_TRUE(Compare());
}
static double RndGaussian(random::SimplePhilox* rnd) {
// Box-Muller transformation.
// See, for example, http://www.taygeta.com/random/gaussian.html
double x1, x2;
double r;
do {
x1 = 2 * rnd->RandDouble() - 1;
x2 = 2 * rnd->RandDouble() - 1;
r = x1 * x1 + x2 * x2;
} while (r == 0 || r >= 1.0);
double w = sqrt(-2.0 * log(r) / r);
return x1 * w;
}
#define TCASE(NAME, DEPTH, BATCH, DEPTH_RADIUS, BIAS, ALPHA, BETA) \
TEST_F(LRNFloatTest, NAME) { \
ASSERT_OK(NodeDefBuilder("lrn_op", "LRN") \
.Input(FakeInput()) \
.Attr("depth_radius", (DEPTH_RADIUS)) \
.Attr("bias", (BIAS)) \
.Attr("alpha", ((ALPHA) / 10)) \
.Attr("beta", (BETA)) \
.Finalize(node_def())); \
ASSERT_OK(InitOp()); \
AddInput<float>(TensorShape({BATCH, 1, 1, DEPTH}), \
[this](int i) -> float { return RndGaussian(&rand_); }); \
ASSERT_OK(RunOpKernel()); \
EXPECT_TRUE(Compare()); \
}
// clang-format off
// DEPTH BATCH DEPTH_RADIUS BIAS ALPHA BETA
TCASE(T0, 4, 2, 2, 1.0f, 1.0f, 2.0f)
TCASE(T1, 16, 1, 5, 1.0f, 1.0f, 2.0f)
TCASE(T2, 16, 32, 2, 1.0f, 2.0f, 1.0f)
TCASE(T3, 128, 4, 3, 2.0f, 1.0f, 1.0f)
// clang-format on
#undef TCASE
} // namespace tensorflow
|
