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

//! ResNet implementation.
//!
//! See "Deep Residual Learning for Image Recognition" He et al. 2015
//! <https://arxiv.org/abs/1512.03385>
use crate::{nn, nn::Conv2D, nn::FuncT, nn::ModuleT};

fn conv2d(p: nn::Path, c_in: i64, c_out: i64, ksize: i64, padding: i64, stride: i64) -> Conv2D {
    let conv2d_cfg = nn::ConvConfig { stride, padding, bias: false, ..Default::default() };
    nn::conv2d(p, c_in, c_out, ksize, conv2d_cfg)
}

fn downsample(p: nn::Path, c_in: i64, c_out: i64, stride: i64) -> impl ModuleT {
    if stride != 1 || c_in != c_out {
        nn::seq_t().add(conv2d(&p / "0", c_in, c_out, 1, 0, stride)).add(nn::batch_norm2d(
            &p / "1",
            c_out,
            Default::default(),
        ))
    } else {
        nn::seq_t()
    }
}

fn basic_block(p: nn::Path, c_in: i64, c_out: i64, stride: i64) -> impl ModuleT {
    let conv1 = conv2d(&p / "conv1", c_in, c_out, 3, 1, stride);
    let bn1 = nn::batch_norm2d(&p / "bn1", c_out, Default::default());
    let conv2 = conv2d(&p / "conv2", c_out, c_out, 3, 1, 1);
    let bn2 = nn::batch_norm2d(&p / "bn2", c_out, Default::default());
    let downsample = downsample(&p / "downsample", c_in, c_out, stride);
    nn::func_t(move |xs, train| {
        let ys = xs.apply(&conv1).apply_t(&bn1, train).relu().apply(&conv2).apply_t(&bn2, train);
        (xs.apply_t(&downsample, train) + ys).relu()
    })
}

fn basic_layer(p: nn::Path, c_in: i64, c_out: i64, stride: i64, cnt: i64) -> impl ModuleT {
    let mut layer = nn::seq_t().add(basic_block(&p / "0", c_in, c_out, stride));
    for block_index in 1..cnt {
        layer = layer.add(basic_block(&p / &block_index.to_string(), c_out, c_out, 1))
    }
    layer
}

fn resnet(
    p: &nn::Path,
    nclasses: Option<i64>,
    c1: i64,
    c2: i64,
    c3: i64,
    c4: i64,
) -> FuncT<'static> {
    let conv1 = conv2d(p / "conv1", 3, 64, 7, 3, 2);
    let bn1 = nn::batch_norm2d(p / "bn1", 64, Default::default());
    let layer1 = basic_layer(p / "layer1", 64, 64, 1, c1);
    let layer2 = basic_layer(p / "layer2", 64, 128, 2, c2);
    let layer3 = basic_layer(p / "layer3", 128, 256, 2, c3);
    let layer4 = basic_layer(p / "layer4", 256, 512, 2, c4);
    let fc = nclasses.map(|n| nn::linear(p / "fc", 512, n, Default::default()));
    nn::func_t(move |xs, train| {
        xs.apply(&conv1)
            .apply_t(&bn1, train)
            .relu()
            .max_pool2d([3, 3], [2, 2], [1, 1], [1, 1], false)
            .apply_t(&layer1, train)
            .apply_t(&layer2, train)
            .apply_t(&layer3, train)
            .apply_t(&layer4, train)
            .adaptive_avg_pool2d([1, 1])
            .flat_view()
            .apply_opt(&fc)
    })
}

/// Creates a ResNet-18 model.
///
/// Pre-trained weights can be downloaded at the following link:
/// <https://github.com/LaurentMazare/tch-rs/releases/download/untagged-eb220e5c19f9bb250bd1/resnet18.ot>
pub fn resnet18(p: &nn::Path, num_classes: i64) -> FuncT<'static> {
    resnet(p, Some(num_classes), 2, 2, 2, 2)
}

pub fn resnet18_no_final_layer(p: &nn::Path) -> FuncT<'static> {
    resnet(p, None, 2, 2, 2, 2)
}

/// Creates a ResNet-34 model.
///
/// Pre-trained weights can be downloaded at the following link:
/// <https://github.com/LaurentMazare/tch-rs/releases/download/untagged-eb220e5c19f9bb250bd1/resnet34.ot>
pub fn resnet34(p: &nn::Path, num_classes: i64) -> FuncT<'static> {
    resnet(p, Some(num_classes), 3, 4, 6, 3)
}

pub fn resnet34_no_final_layer(p: &nn::Path) -> FuncT<'static> {
    resnet(p, None, 3, 4, 6, 3)
}

// Bottleneck versions for ResNet 50, 101, and 152.

fn bottleneck_block(p: nn::Path, c_in: i64, c_out: i64, stride: i64, e: i64) -> impl ModuleT {
    let e_dim = e * c_out;
    let conv1 = conv2d(&p / "conv1", c_in, c_out, 1, 0, 1);
    let bn1 = nn::batch_norm2d(&p / "bn1", c_out, Default::default());
    let conv2 = conv2d(&p / "conv2", c_out, c_out, 3, 1, stride);
    let bn2 = nn::batch_norm2d(&p / "bn2", c_out, Default::default());
    let conv3 = conv2d(&p / "conv3", c_out, e_dim, 1, 0, 1);
    let bn3 = nn::batch_norm2d(&p / "bn3", e_dim, Default::default());
    let downsample = downsample(&p / "downsample", c_in, e_dim, stride);
    nn::func_t(move |xs, train| {
        let ys = xs
            .apply(&conv1)
            .apply_t(&bn1, train)
            .relu()
            .apply(&conv2)
            .apply_t(&bn2, train)
            .relu()
            .apply(&conv3)
            .apply_t(&bn3, train);
        (xs.apply_t(&downsample, train) + ys).relu()
    })
}

fn bottleneck_layer(p: nn::Path, c_in: i64, c_out: i64, stride: i64, cnt: i64) -> impl ModuleT {
    let mut layer = nn::seq_t().add(bottleneck_block(&p / "0", c_in, c_out, stride, 4));
    for block_index in 1..cnt {
        layer = layer.add(bottleneck_block(&p / &block_index.to_string(), 4 * c_out, c_out, 1, 4))
    }
    layer
}

fn bottleneck_resnet(
    p: &nn::Path,
    nclasses: Option<i64>,
    c1: i64,
    c2: i64,
    c3: i64,
    c4: i64,
) -> impl ModuleT {
    let conv1 = conv2d(p / "conv1", 3, 64, 7, 3, 2);
    let bn1 = nn::batch_norm2d(p / "bn1", 64, Default::default());
    let layer1 = bottleneck_layer(p / "layer1", 64, 64, 1, c1);
    let layer2 = bottleneck_layer(p / "layer2", 4 * 64, 128, 2, c2);
    let layer3 = bottleneck_layer(p / "layer3", 4 * 128, 256, 2, c3);
    let layer4 = bottleneck_layer(p / "layer4", 4 * 256, 512, 2, c4);
    let fc = nclasses.map(|n| nn::linear(p / "fc", 4 * 512, n, Default::default()));
    nn::func_t(move |xs, train| {
        xs.apply(&conv1)
            .apply_t(&bn1, train)
            .relu()
            .max_pool2d([3, 3], [2, 2], [1, 1], [1, 1], false)
            .apply_t(&layer1, train)
            .apply_t(&layer2, train)
            .apply_t(&layer3, train)
            .apply_t(&layer4, train)
            .adaptive_avg_pool2d([1, 1])
            .flat_view()
            .apply_opt(&fc)
    })
}

pub fn resnet50(p: &nn::Path, num_classes: i64) -> impl ModuleT {
    bottleneck_resnet(p, Some(num_classes), 3, 4, 6, 3)
}

pub fn resnet50_no_final_layer(p: &nn::Path) -> impl ModuleT {
    bottleneck_resnet(p, None, 3, 4, 6, 3)
}

pub fn resnet101(p: &nn::Path, num_classes: i64) -> impl ModuleT {
    bottleneck_resnet(p, Some(num_classes), 3, 4, 23, 3)
}

pub fn resnet101_no_final_layer(p: &nn::Path) -> impl ModuleT {
    bottleneck_resnet(p, None, 3, 4, 23, 3)
}

pub fn resnet152(p: &nn::Path, num_classes: i64) -> impl ModuleT {
    bottleneck_resnet(p, Some(num_classes), 3, 8, 36, 3)
}

pub fn resnet150_no_final_layer(p: &nn::Path) -> impl ModuleT {
    bottleneck_resnet(p, None, 3, 8, 36, 3)
}