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
|
#include <cmath>
#include <vector>
#include "gtest/gtest.h"
#include "caffe/blob.hpp"
#include "caffe/common.hpp"
#include "caffe/filler.hpp"
#include "caffe/layers/sigmoid_cross_entropy_loss_layer.hpp"
#include "caffe/test/test_caffe_main.hpp"
#include "caffe/test/test_gradient_check_util.hpp"
namespace caffe {
template <typename TypeParam>
class SigmoidCrossEntropyLossLayerTest : public MultiDeviceTest<TypeParam> {
typedef typename TypeParam::Dtype Dtype;
protected:
SigmoidCrossEntropyLossLayerTest()
: blob_bottom_data_(new Blob<Dtype>(10, 5, 1, 1)),
blob_bottom_targets_(new Blob<Dtype>(10, 5, 1, 1)),
blob_top_loss_(new Blob<Dtype>()) {
// Fill the data vector
FillerParameter data_filler_param;
data_filler_param.set_std(1);
GaussianFiller<Dtype> data_filler(data_filler_param);
data_filler.Fill(blob_bottom_data_);
blob_bottom_vec_.push_back(blob_bottom_data_);
// Fill the targets vector
FillerParameter targets_filler_param;
targets_filler_param.set_min(0);
targets_filler_param.set_max(1);
UniformFiller<Dtype> targets_filler(targets_filler_param);
targets_filler.Fill(blob_bottom_targets_);
blob_bottom_vec_.push_back(blob_bottom_targets_);
blob_top_vec_.push_back(blob_top_loss_);
}
virtual ~SigmoidCrossEntropyLossLayerTest() {
delete blob_bottom_data_;
delete blob_bottom_targets_;
delete blob_top_loss_;
}
Dtype SigmoidCrossEntropyLossReference(const int count, const int num,
const Dtype* input,
const Dtype* target) {
Dtype loss = 0;
for (int i = 0; i < count; ++i) {
const Dtype prediction = 1 / (1 + exp(-input[i]));
EXPECT_LE(prediction, 1);
EXPECT_GE(prediction, 0);
EXPECT_LE(target[i], 1);
EXPECT_GE(target[i], 0);
loss -= target[i] * log(prediction + (target[i] == Dtype(0)));
loss -= (1 - target[i]) * log(1 - prediction + (target[i] == Dtype(1)));
}
return loss / num;
}
void TestForward() {
LayerParameter layer_param;
const Dtype kLossWeight = 3.7;
layer_param.add_loss_weight(kLossWeight);
FillerParameter data_filler_param;
data_filler_param.set_std(1);
GaussianFiller<Dtype> data_filler(data_filler_param);
FillerParameter targets_filler_param;
targets_filler_param.set_min(0.0);
targets_filler_param.set_max(1.0);
UniformFiller<Dtype> targets_filler(targets_filler_param);
Dtype eps = 2e-2;
for (int i = 0; i < 100; ++i) {
// Fill the data vector
data_filler.Fill(this->blob_bottom_data_);
// Fill the targets vector
targets_filler.Fill(this->blob_bottom_targets_);
SigmoidCrossEntropyLossLayer<Dtype> layer(layer_param);
layer.SetUp(this->blob_bottom_vec_, this->blob_top_vec_);
Dtype layer_loss =
layer.Forward(this->blob_bottom_vec_, this->blob_top_vec_);
const int count = this->blob_bottom_data_->count();
const int num = this->blob_bottom_data_->num();
const Dtype* blob_bottom_data = this->blob_bottom_data_->cpu_data();
const Dtype* blob_bottom_targets =
this->blob_bottom_targets_->cpu_data();
Dtype reference_loss = kLossWeight * SigmoidCrossEntropyLossReference(
count, num, blob_bottom_data, blob_bottom_targets);
EXPECT_NEAR(reference_loss, layer_loss, eps) << "debug: trial #" << i;
}
}
Blob<Dtype>* const blob_bottom_data_;
Blob<Dtype>* const blob_bottom_targets_;
Blob<Dtype>* const blob_top_loss_;
vector<Blob<Dtype>*> blob_bottom_vec_;
vector<Blob<Dtype>*> blob_top_vec_;
};
TYPED_TEST_CASE(SigmoidCrossEntropyLossLayerTest, TestDtypesAndDevices);
TYPED_TEST(SigmoidCrossEntropyLossLayerTest, TestSigmoidCrossEntropyLoss) {
this->TestForward();
}
TYPED_TEST(SigmoidCrossEntropyLossLayerTest, TestGradient) {
typedef typename TypeParam::Dtype Dtype;
LayerParameter layer_param;
const Dtype kLossWeight = 3.7;
layer_param.add_loss_weight(kLossWeight);
SigmoidCrossEntropyLossLayer<Dtype> layer(layer_param);
layer.SetUp(this->blob_bottom_vec_, this->blob_top_vec_);
GradientChecker<Dtype> checker(1e-2, 1e-2, 1701);
checker.CheckGradientExhaustive(&layer, this->blob_bottom_vec_,
this->blob_top_vec_, 0);
}
TYPED_TEST(SigmoidCrossEntropyLossLayerTest, TestIgnoreGradient) {
typedef typename TypeParam::Dtype Dtype;
FillerParameter data_filler_param;
data_filler_param.set_std(1);
GaussianFiller<Dtype> data_filler(data_filler_param);
data_filler.Fill(this->blob_bottom_data_);
LayerParameter layer_param;
LossParameter* loss_param = layer_param.mutable_loss_param();
loss_param->set_ignore_label(-1);
Dtype* target = this->blob_bottom_targets_->mutable_cpu_data();
const int count = this->blob_bottom_targets_->count();
// Ignore half of targets, then check that diff of this half is zero,
// while the other half is nonzero.
caffe_set(count / 2, Dtype(-1), target);
SigmoidCrossEntropyLossLayer<Dtype> layer(layer_param);
layer.SetUp(this->blob_bottom_vec_, this->blob_top_vec_);
layer.Forward(this->blob_bottom_vec_, this->blob_top_vec_);
vector<bool> propagate_down(2);
propagate_down[0] = true;
propagate_down[1] = false;
layer.Backward(this->blob_top_vec_, propagate_down, this->blob_bottom_vec_);
const Dtype* diff = this->blob_bottom_data_->cpu_diff();
for (int i = 0; i < count / 2; ++i) {
EXPECT_FLOAT_EQ(diff[i], 0.);
EXPECT_NE(diff[i + count / 2], 0.);
}
}
} // namespace caffe
|
