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
|
#include <gtest/gtest.h>
#include <c10/util/irange.h>
#include <torch/torch.h>
#include <test/cpp/api/init_baseline.h>
#include <test/cpp/api/support.h>
#include <functional>
#include <vector>
void check_exact_values(
const std::vector<torch::Tensor>& parameters,
const std::vector<std::vector<torch::Tensor>>& expected_parameters) {
ASSERT_EQ(parameters.size(), expected_parameters.size());
for (const auto i : c10::irange(parameters.size())) {
auto layerParameters = parameters[i];
auto expectedLayerParameters = expected_parameters[i];
if (static_cast<size_t>(layerParameters.size(0)) !=
expectedLayerParameters.size()) {
std::cout << "layer #" << i
<< " layerParameters size: " << layerParameters.size(0)
<< " != "
<< " expectedLayerParameters size: "
<< expectedLayerParameters.size() << std::endl;
ASSERT_TRUE(false);
}
for (const auto p : c10::irange(layerParameters.size(0))) {
// Always compare using double dtype, regardless of the original dtype of
// the tensors
auto tensor = layerParameters[p].to(torch::kFloat64);
auto expectedTensor = expectedLayerParameters[p].to(torch::kFloat64);
if (!tensor.allclose(expectedTensor, /*rtol=*/1e-3, /*atol=*/5e-4)) {
std::cout << "layer " << i << ": " << tensor << " != " << expectedTensor
<< " (parameter " << p << ")" << std::endl;
ASSERT_TRUE(false);
}
}
}
}
void check_initializer_against_baseline(
std::function<void(torch::Tensor)> initializer,
std::vector<std::vector<torch::Tensor>> expected) {
torch::manual_seed(0);
auto layer1 = torch::nn::Linear(7, 15);
initializer(layer1->weight);
layer1->to(torch::kFloat64);
auto layer2 = torch::nn::Linear(15, 15);
initializer(layer2->weight);
layer2->to(torch::kFloat64);
auto layer3 = torch::nn::Linear(15, 2);
initializer(layer3->weight);
layer3->to(torch::kFloat64);
auto parameters = std::vector<torch::Tensor>{
layer1->weight,
layer2->weight,
layer3->weight,
};
check_exact_values(parameters, expected);
}
TEST(InitTest, ProducesPyTorchValues_XavierUniform) {
auto expected = expected_parameters::Xavier_Uniform();
auto initializer = [](torch::Tensor tensor) {
torch::nn::init::xavier_uniform_(tensor);
};
check_initializer_against_baseline(initializer, expected);
}
TEST(InitTest, ProducesPyTorchValues_XavierNormal) {
auto expected = expected_parameters::Xavier_Normal();
auto initializer = [](torch::Tensor tensor) {
torch::nn::init::xavier_normal_(tensor);
};
check_initializer_against_baseline(initializer, expected);
}
TEST(InitTest, ProducesPyTorchValues_KaimingNormal) {
auto expected = expected_parameters::Kaiming_Normal();
auto initializer = [](torch::Tensor tensor) {
torch::nn::init::kaiming_normal_(tensor);
};
check_initializer_against_baseline(initializer, expected);
}
TEST(InitTest, ProducesPyTorchValues_KaimingUniform) {
auto expected = expected_parameters::Kaiming_Uniform();
auto initializer = [](torch::Tensor tensor) {
torch::nn::init::kaiming_uniform_(tensor);
};
check_initializer_against_baseline(initializer, expected);
}
TEST(InitTest, CanInitializeTensorThatRequiresGrad) {
auto tensor = torch::empty({3, 4}, torch::requires_grad());
ASSERT_THROWS_WITH(
tensor.fill_(1),
"a leaf Variable that requires grad "
"is being used in an in-place operation");
ASSERT_EQ(torch::nn::init::ones_(tensor).sum().item<int32_t>(), 12);
}
TEST(InitTest, CalculateGainWithTanh) {
double gain = torch::nn::init::calculate_gain(torch::kTanh);
ASSERT_DOUBLE_EQ(gain, 5.0 / 3.0);
}
TEST(InitTest, CalculateGainWithRelu) {
double gain = torch::nn::init::calculate_gain(torch::kReLU);
ASSERT_DOUBLE_EQ(gain, std::sqrt(2.0));
}
TEST(InitTest, CalculateGainWithLeakyRelu) {
double gain = torch::nn::init::calculate_gain(torch::kLeakyReLU);
ASSERT_DOUBLE_EQ(gain, std::sqrt(2.0 / (1 + pow(0.01, 2))));
}
TEST(InitTest, CanInitializeCnnWithOrthogonal) {
torch::nn::Conv2d conv_layer(torch::nn::Conv2dOptions(3, 2, 3).stride(2));
torch::nn::init::orthogonal_(conv_layer->named_parameters()["weight"]);
}
|
