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

//! InceptionV3.
use crate::{nn, nn::ModuleT, Tensor};

fn conv_bn(p: nn::Path, c_in: i64, c_out: i64, ksize: i64, pad: i64, stride: i64) -> impl ModuleT {
    let conv2d_cfg = nn::ConvConfig { stride, padding: pad, bias: false, ..Default::default() };
    let bn_cfg = nn::BatchNormConfig { eps: 0.001, ..Default::default() };
    nn::seq_t()
        .add(nn::conv2d(&p / "conv", c_in, c_out, ksize, conv2d_cfg))
        .add(nn::batch_norm2d(&p / "bn", c_out, bn_cfg))
        .add_fn(|xs| xs.relu())
}

fn conv_bn2(p: nn::Path, c_in: i64, c_out: i64, ksize: [i64; 2], pad: [i64; 2]) -> impl ModuleT {
    let conv2d_cfg =
        nn::ConvConfigND::<[i64; 2]> { padding: pad, bias: false, ..Default::default() };
    let bn_cfg = nn::BatchNormConfig { eps: 0.001, ..Default::default() };
    nn::seq_t()
        .add(nn::conv(&p / "conv", c_in, c_out, ksize, conv2d_cfg))
        .add(nn::batch_norm2d(&p / "bn", c_out, bn_cfg))
        .add_fn(|xs| xs.relu())
}

fn max_pool2d(xs: &Tensor, ksize: i64, stride: i64) -> Tensor {
    xs.max_pool2d([ksize, ksize], [stride, stride], [0, 0], [1, 1], false)
}

fn inception_a(p: nn::Path, c_in: i64, c_pool: i64) -> impl ModuleT {
    let b1 = conv_bn(&p / "branch1x1", c_in, 64, 1, 0, 1);
    let b2_1 = conv_bn(&p / "branch5x5_1", c_in, 48, 1, 0, 1);
    let b2_2 = conv_bn(&p / "branch5x5_2", 48, 64, 5, 2, 1);
    let b3_1 = conv_bn(&p / "branch3x3dbl_1", c_in, 64, 1, 0, 1);
    let b3_2 = conv_bn(&p / "branch3x3dbl_2", 64, 96, 3, 1, 1);
    let b3_3 = conv_bn(&p / "branch3x3dbl_3", 96, 96, 3, 1, 1);
    let bpool = conv_bn(&p / "branch_pool", c_in, c_pool, 1, 0, 1);
    nn::func_t(move |xs, tr| {
        let b1 = xs.apply_t(&b1, tr);
        let b2 = xs.apply_t(&b2_1, tr).apply_t(&b2_2, tr);
        let b3 = xs.apply_t(&b3_1, tr).apply_t(&b3_2, tr).apply_t(&b3_3, tr);
        let bpool = xs.avg_pool2d([3, 3], [1, 1], [1, 1], false, true, 9).apply_t(&bpool, tr);
        Tensor::cat(&[b1, b2, b3, bpool], 1)
    })
}

fn inception_b(p: nn::Path, c_in: i64) -> impl ModuleT {
    let b1 = conv_bn(&p / "branch3x3", c_in, 384, 3, 0, 2);
    let b2_1 = conv_bn(&p / "branch3x3dbl_1", c_in, 64, 1, 0, 1);
    let b2_2 = conv_bn(&p / "branch3x3dbl_2", 64, 96, 3, 1, 1);
    let b2_3 = conv_bn(&p / "branch3x3dbl_3", 96, 96, 3, 0, 2);
    nn::func_t(move |xs, tr| {
        let b1 = xs.apply_t(&b1, tr);
        let b2 = xs.apply_t(&b2_1, tr).apply_t(&b2_2, tr).apply_t(&b2_3, tr);
        let bpool = max_pool2d(xs, 3, 2);
        Tensor::cat(&[b1, b2, bpool], 1)
    })
}

fn inception_c(p: nn::Path, c_in: i64, c7: i64) -> impl ModuleT {
    let b1 = conv_bn(&p / "branch1x1", c_in, 192, 1, 0, 1);

    let b2_1 = conv_bn(&p / "branch7x7_1", c_in, c7, 1, 0, 1);
    let b2_2 = conv_bn2(&p / "branch7x7_2", c7, c7, [1, 7], [0, 3]);
    let b2_3 = conv_bn2(&p / "branch7x7_3", c7, 192, [7, 1], [3, 0]);

    let b3_1 = conv_bn(&p / "branch7x7dbl_1", c_in, c7, 1, 0, 1);
    let b3_2 = conv_bn2(&p / "branch7x7dbl_2", c7, c7, [7, 1], [3, 0]);
    let b3_3 = conv_bn2(&p / "branch7x7dbl_3", c7, c7, [1, 7], [0, 3]);
    let b3_4 = conv_bn2(&p / "branch7x7dbl_4", c7, c7, [7, 1], [3, 0]);
    let b3_5 = conv_bn2(&p / "branch7x7dbl_5", c7, 192, [1, 7], [0, 3]);

    let bpool = conv_bn(&p / "branch_pool", c_in, 192, 1, 0, 1);

    nn::func_t(move |xs, tr| {
        let b1 = xs.apply_t(&b1, tr);
        let b2 = xs.apply_t(&b2_1, tr).apply_t(&b2_2, tr).apply_t(&b2_3, tr);
        let b3 = xs
            .apply_t(&b3_1, tr)
            .apply_t(&b3_2, tr)
            .apply_t(&b3_3, tr)
            .apply_t(&b3_4, tr)
            .apply_t(&b3_5, tr);
        let bpool = xs.avg_pool2d([3, 3], [1, 1], [1, 1], false, true, 9).apply_t(&bpool, tr);
        Tensor::cat(&[b1, b2, b3, bpool], 1)
    })
}

fn inception_d(p: nn::Path, c_in: i64) -> impl ModuleT {
    let b1_1 = conv_bn(&p / "branch3x3_1", c_in, 192, 1, 0, 1);
    let b1_2 = conv_bn(&p / "branch3x3_2", 192, 320, 3, 0, 2);

    let b2_1 = conv_bn(&p / "branch7x7x3_1", c_in, 192, 1, 0, 1);
    let b2_2 = conv_bn2(&p / "branch7x7x3_2", 192, 192, [1, 7], [0, 3]);
    let b2_3 = conv_bn2(&p / "branch7x7x3_3", 192, 192, [7, 1], [3, 0]);
    let b2_4 = conv_bn(&p / "branch7x7x3_4", 192, 192, 3, 0, 2);

    nn::func_t(move |xs, tr| {
        let b1 = xs.apply_t(&b1_1, tr).apply_t(&b1_2, tr);
        let b2 = xs.apply_t(&b2_1, tr).apply_t(&b2_2, tr).apply_t(&b2_3, tr).apply_t(&b2_4, tr);
        let bpool = max_pool2d(xs, 3, 2);
        Tensor::cat(&[b1, b2, bpool], 1)
    })
}

fn inception_e(p: nn::Path, c_in: i64) -> impl ModuleT {
    let b1 = conv_bn(&p / "branch1x1", c_in, 320, 1, 0, 1);

    let b2_1 = conv_bn(&p / "branch3x3_1", c_in, 384, 1, 0, 1);
    let b2_2a = conv_bn2(&p / "branch3x3_2a", 384, 384, [1, 3], [0, 1]);
    let b2_2b = conv_bn2(&p / "branch3x3_2b", 384, 384, [3, 1], [1, 0]);

    let b3_1 = conv_bn(&p / "branch3x3dbl_1", c_in, 448, 1, 0, 1);
    let b3_2 = conv_bn(&p / "branch3x3dbl_2", 448, 384, 3, 1, 1);
    let b3_3a = conv_bn2(&p / "branch3x3dbl_3a", 384, 384, [1, 3], [0, 1]);
    let b3_3b = conv_bn2(&p / "branch3x3dbl_3b", 384, 384, [3, 1], [1, 0]);

    let bpool = conv_bn(&p / "branch_pool", c_in, 192, 1, 0, 1);

    nn::func_t(move |xs, tr| {
        let b1 = xs.apply_t(&b1, tr);

        let b2 = xs.apply_t(&b2_1, tr);
        let b2 = Tensor::cat(&[b2.apply_t(&b2_2a, tr), b2.apply_t(&b2_2b, tr)], 1);

        let b3 = xs.apply_t(&b3_1, tr).apply_t(&b3_2, tr);
        let b3 = Tensor::cat(&[b3.apply_t(&b3_3a, tr), b3.apply_t(&b3_3b, tr)], 1);

        let bpool = xs.avg_pool2d([3, 3], [1, 1], [1, 1], false, true, 9).apply_t(&bpool, tr);

        Tensor::cat(&[b1, b2, b3, bpool], 1)
    })
}

pub fn v3(p: &nn::Path, nclasses: i64) -> impl ModuleT {
    nn::seq_t()
        .add(conv_bn(p / "Conv2d_1a_3x3", 3, 32, 3, 0, 2))
        .add(conv_bn(p / "Conv2d_2a_3x3", 32, 32, 3, 0, 1))
        .add(conv_bn(p / "Conv2d_2b_3x3", 32, 64, 3, 1, 1))
        .add_fn(|xs| max_pool2d(&xs.relu(), 3, 2))
        .add(conv_bn(p / "Conv2d_3b_1x1", 64, 80, 1, 0, 1))
        .add(conv_bn(p / "Conv2d_4a_3x3", 80, 192, 3, 0, 1))
        .add_fn(|xs| max_pool2d(&xs.relu(), 3, 2))
        .add(inception_a(p / "Mixed_5b", 192, 32))
        .add(inception_a(p / "Mixed_5c", 256, 64))
        .add(inception_a(p / "Mixed_5d", 288, 64))
        .add(inception_b(p / "Mixed_6a", 288))
        .add(inception_c(p / "Mixed_6b", 768, 128))
        .add(inception_c(p / "Mixed_6c", 768, 160))
        .add(inception_c(p / "Mixed_6d", 768, 160))
        .add(inception_c(p / "Mixed_6e", 768, 192))
        .add(inception_d(p / "Mixed_7a", 768))
        .add(inception_e(p / "Mixed_7b", 1280))
        .add(inception_e(p / "Mixed_7c", 2048))
        .add_fn_t(|xs, train| xs.adaptive_avg_pool2d([1, 1]).dropout(0.5, train).flat_view())
        .add(nn::linear(p / "fc", 2048, nclasses, Default::default()))
}