bunsen 0.29.0

bunsen is a batteries included common library for burn
Documentation 
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

//! Windowing operations for Swin Transformer v2

use burn::{
    prelude::{
        Backend,
        Tensor,
    },
    tensor::BasicOps,
};

use crate::contracts::unpack_shape_contract;

/// Window Partition
///
/// # Arguments
///
/// - `tensor`: Input tensor of `[batch, height, width, channels]`.
/// - `window_size`: Window size.
///
/// # Returns
///
/// Output tensor of `[batch * h_windows * w_windows, window_size, window_size,
/// channels]`.
///
/// # Panics
///
/// On shape contract failure.
#[inline]
#[must_use]
pub fn window_partition<B: Backend, K>(
    tensor: Tensor<B, 4, K>,
    window_size: usize,
) -> Tensor<B, 4, K>
where
    K: BasicOps<B>,
{
    let [b, h_wins, w_wins, c] = unpack_shape_contract!(
        [
            "batch",
            "h_wins" * "window_size",
            "w_wins" * "window_size",
            "channels"
        ],
        &tensor.dims(),
        &["batch", "h_wins", "w_wins", "channels"],
        &[("window_size", window_size)]
    );

    tensor
        .reshape([b, h_wins, window_size, w_wins, window_size, c])
        .swap_dims(2, 3)
        .reshape([b * h_wins * w_wins, window_size, window_size, c])
}

/// Window Reverse
///
/// # Arguments
///
/// - `windows`: Input tensor of `[batch * h_windows * w_windows, window_size,
///   window_size, channels]`.
/// - `window_size`: Window size.
/// - `height`: Height of the original image.
/// - `width`: Width of the original image.
///
/// # Returns
///
/// `[batch, height, width, channels]` output tensor.
///
/// # Panics
///
/// On shape contract failure.
#[inline]
#[must_use]
pub fn window_reverse<B: Backend, K>(
    windows: Tensor<B, 4, K>,
    window_size: usize,
    height: usize,
    width: usize,
) -> Tensor<B, 4, K>
where
    K: BasicOps<B>,
{
    let h_wins = height / window_size;
    let w_wins = width / window_size;

    let [b, c] = unpack_shape_contract!(
        [
            "batch" * "h_wins" * "w_wins",
            "window_size",
            "window_size",
            "channels"
        ],
        &windows.dims(),
        &["batch", "channels"],
        &[
            ("h_wins", h_wins),
            ("w_wins", w_wins),
            ("window_size", window_size),
        ],
    );

    windows
        .reshape([b, h_wins, w_wins, window_size, window_size, c])
        .swap_dims(2, 3)
        .reshape([b, height, width, c])
}

#[cfg(test)]
mod tests {
    use burn::{
        prelude::Tensor,
        tensor::{
            Distribution,
            Tolerance,
        },
    };
    use serial_test::serial;

    use super::*;
    use crate::support::testing::PerformanceBackend;

    #[test]
    #[serial]
    fn test_window_partition() {
        type B = PerformanceBackend;
        let device = Default::default();

        let b = 3;
        let window_size = 4;
        let channels = 3;

        let h_wins = 2;
        let w_wins = 3;
        let h = h_wins * window_size;
        let w = w_wins * window_size;

        let distribution = Distribution::Uniform(0.0, 1.0);
        let input = Tensor::<B, 4>::random([b, h, w, channels], distribution, &device);

        let windows = window_partition(input.clone(), window_size);

        assert_eq!(
            &windows.dims(),
            &[b * h_wins * w_wins, window_size, window_size, channels]
        );

        let reverse = window_reverse(windows, window_size, h, w);
        assert_eq!(&reverse.dims(), &[b, h, w, channels]);

        reverse
            .to_data()
            .assert_approx_eq(&input.to_data(), Tolerance::<f64>::default());
    }
}