Skip to content

GitLab

Switch branch/tag
Find file Normal viewHistoryPermalink
resnet.cpp 4.02 KB
Newer Older
Shahriar's avatar
C++ Models (#728)  
Shahriar committed
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 
#include "resnet.h"

namespace vision {
namespace models {
namespace _resnetimpl {
torch::nn::Conv2d conv3x3(
    int64_t in,
    int64_t out,
    int64_t stride,
    int64_t groups) {
  torch::nn::Conv2dOptions O(in, out, 3);
  O.padding(1).stride(stride).groups(groups).with_bias(false);
  return torch::nn::Conv2d(O);
}

torch::nn::Conv2d conv1x1(int64_t in, int64_t out, int64_t stride) {
  torch::nn::Conv2dOptions O(in, out, 1);
  O.stride(stride).with_bias(false);
  return torch::nn::Conv2d(O);
}

int BasicBlock::expansion = 1;
int Bottleneck::expansion = 4;

BasicBlock::BasicBlock(
    int64_t inplanes,
    int64_t planes,
    int64_t stride,
    torch::nn::Sequential downsample,
    int64_t groups,
    int64_t base_width)
    : stride(stride), downsample(downsample) {
  if (groups != 1 or base_width != 64) {
    std::cerr << "BasicBlock only supports groups=1 and base_width=64"
              << std::endl;
    assert(false);
  }

  // Both conv1 and downsample layers downsample the input when stride != 1
  conv1 = conv3x3(inplanes, planes, stride);
  conv2 = conv3x3(planes, planes);

  bn1 = torch::nn::BatchNorm(planes);
  bn2 = torch::nn::BatchNorm(planes);

  register_module("conv1", conv1);
  register_module("conv2", conv2);

  register_module("bn1", bn1);
  register_module("bn2", bn2);

  if (!downsample.is_empty())
    register_module("downsample", this->downsample);
}

Bottleneck::Bottleneck(
    int64_t inplanes,
    int64_t planes,
    int64_t stride,
    torch::nn::Sequential downsample,
    int64_t groups,
    int64_t base_width)
    : stride(stride), downsample(downsample) {
  auto width = int64_t(planes * (base_width / 64.)) * groups;

  // Both conv2 and downsample layers downsample the input when stride != 1
  conv1 = conv1x1(inplanes, width);
  conv2 = conv3x3(width, width, stride, groups);
  conv3 = conv1x1(width, planes * expansion);

  bn1 = torch::nn::BatchNorm(width);
  bn2 = torch::nn::BatchNorm(width);
  bn3 = torch::nn::BatchNorm(planes * expansion);

  register_module("conv1", conv1);
  register_module("conv2", conv2);
  register_module("conv3", conv3);

  register_module("bn1", bn1);
  register_module("bn2", bn2);
  register_module("bn3", bn3);

  if (!downsample.is_empty())
    register_module("downsample", this->downsample);
}

torch::Tensor Bottleneck::forward(torch::Tensor X) {
  auto identity = X;

  auto out = conv1->forward(X);
  out = bn1->forward(out).relu_();

  out = conv2->forward(out);
  out = bn2->forward(out).relu_();

  out = conv3->forward(out);
  out = bn3->forward(out);

  if (!downsample.is_empty())
    identity = downsample->forward(X);

  out += identity;
  return out.relu_();
}

torch::Tensor BasicBlock::forward(torch::Tensor x) {
  auto identity = x;

  auto out = conv1->forward(x);
  out = bn1->forward(out).relu_();

  out = conv2->forward(out);
  out = bn2->forward(out);

  if (!downsample.is_empty())
    identity = downsample->forward(x);

  out += identity;
  return out.relu_();
}
} // namespace _resnetimpl

ResNet18Impl::ResNet18Impl(int64_t num_classes, bool zero_init_residual)
    : ResNetImpl({2, 2, 2, 2}, num_classes, zero_init_residual) {}

ResNet34Impl::ResNet34Impl(int64_t num_classes, bool zero_init_residual)
    : ResNetImpl({3, 4, 6, 3}, num_classes, zero_init_residual) {}

ResNet50Impl::ResNet50Impl(int64_t num_classes, bool zero_init_residual)
    : ResNetImpl({3, 4, 6, 3}, num_classes, zero_init_residual) {}

ResNet101Impl::ResNet101Impl(int64_t num_classes, bool zero_init_residual)
    : ResNetImpl({3, 4, 23, 3}, num_classes, zero_init_residual) {}

ResNet152Impl::ResNet152Impl(int64_t num_classes, bool zero_init_residual)
    : ResNetImpl({3, 8, 36, 3}, num_classes, zero_init_residual) {}

ResNext50_32x4dImpl::ResNext50_32x4dImpl(
    int64_t num_classes,
    bool zero_init_residual)
    : ResNetImpl({3, 4, 6, 3}, num_classes, zero_init_residual, 32, 4) {}

ResNext101_32x8dImpl::ResNext101_32x8dImpl(
    int64_t num_classes,
    bool zero_init_residual)
    : ResNetImpl({3, 4, 23, 3}, num_classes, zero_init_residual, 32, 8) {}

} // namespace models
} // namespace vision