burn-nn 0.22.0

Neural network building blocks for the Burn deep learning framework
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

use burn_core as burn;

use burn::config::Config;
use burn::module::Module;
use burn::module::{Content, DisplaySettings, ModuleDisplay};
use burn::tensor::Tensor;
use burn::tensor::activation::softplus_with_threshold;

/// Softplus layer.
///
/// Applies the softplus function element-wise:
/// `softplus(x) = (1/beta) * log(1 + exp(beta * x))`
///
/// Should be created with [SoftplusConfig](SoftplusConfig).
#[derive(Module, Debug)]
#[module(custom_display)]
pub struct Softplus {
    /// The beta value.
    pub beta: f64,
    /// The stability threshold.
    pub threshold: f64,
}

/// Configuration to create a [Softplus](Softplus) layer using the [init function](SoftplusConfig::init).
#[derive(Config, Debug)]
pub struct SoftplusConfig {
    /// The beta value. Default is 1.0
    #[config(default = "1.0")]
    pub beta: f64,
    /// The value of `beta * x` above which the linear approximation is used, which is what
    /// keeps `exp` from overflowing. Default is 20.0
    ///
    /// See [softplus_with_threshold](burn::tensor::activation::softplus_with_threshold).
    #[config(default = "20.0")]
    pub threshold: f64,
}

impl SoftplusConfig {
    /// Initialize a new [Softplus](Softplus) Layer
    pub fn init(&self) -> Softplus {
        Softplus {
            beta: self.beta,
            threshold: self.threshold,
        }
    }
}

impl ModuleDisplay for Softplus {
    fn custom_settings(&self) -> Option<DisplaySettings> {
        DisplaySettings::new()
            .with_new_line_after_attribute(false)
            .optional()
    }

    fn custom_content(&self, content: Content) -> Option<Content> {
        content
            .add("beta", &self.beta)
            .add("threshold", &self.threshold)
            .optional()
    }
}

impl Softplus {
    /// Forward pass for the Softplus layer.
    ///
    /// See [softplus_with_threshold](burn::tensor::activation::softplus_with_threshold) for
    /// more information.
    ///
    /// # Shapes
    /// - input: `[..., any]`
    /// - output: `[..., any]`
    pub fn forward<const D: usize>(&self, input: Tensor<D>) -> Tensor<D> {
        softplus_with_threshold(input, self.beta, self.threshold)
    }
}

#[cfg(test)]
#[allow(clippy::approx_constant)]
mod tests {
    use super::*;
    use burn::tensor::TensorData;
    use burn::tensor::Tolerance;
    type FT = f32;

    #[test]
    fn test_softplus_forward() {
        let device = Default::default();
        let model = SoftplusConfig::new().init();
        let input = Tensor::<2>::from_data(TensorData::from([[0.0, 1.0, -1.0]]), &device);
        let out = model.forward(input);
        // softplus(0) = log(2) ≈ 0.6931
        // softplus(1) = log(1 + e) ≈ 1.3133
        // softplus(-1) = log(1 + e^-1) ≈ 0.3133
        let expected = TensorData::from([[0.6931, 1.3133, 0.3133]]);
        out.to_data()
            .assert_approx_eq::<FT>(&expected, Tolerance::default());
    }

    #[test]
    fn test_softplus_with_beta() {
        let device = Default::default();
        let model = SoftplusConfig::new().with_beta(2.0).init();
        let input = Tensor::<2>::from_data(TensorData::from([[0.0, 1.0]]), &device);
        let out = model.forward(input);
        // softplus(0, beta=2) = (1/2) * log(1 + exp(0)) = 0.5 * log(2) ≈ 0.3466
        // softplus(1, beta=2) = (1/2) * log(1 + exp(2)) = 0.5 * log(8.389) ≈ 1.0635
        let expected = TensorData::from([[0.3466, 1.0635]]);
        out.to_data()
            .assert_approx_eq::<FT>(&expected, Tolerance::default());
    }

    #[test]
    fn test_softplus_default_threshold() {
        let device = Default::default();
        let model = SoftplusConfig::new().init();
        assert_eq!(model.threshold, 20.0);

        // The default threshold keeps saturated inputs finite.
        let input = Tensor::<2>::from_data(TensorData::from([[100.0, 1000.0]]), &device);
        let out = model.forward(input);
        let expected = TensorData::from([[100.0, 1000.0]]);
        out.to_data()
            .assert_approx_eq::<FT>(&expected, Tolerance::default());
    }

    #[test]
    fn test_softplus_with_threshold() {
        let device = Default::default();
        let input = Tensor::<2>::from_data(TensorData::from([[5.0]]), &device);

        // Below the threshold the curve is evaluated: softplus(5) = log(1 + e^5) ≈ 5.0067.
        let out = SoftplusConfig::new()
            .with_threshold(10.0)
            .init()
            .forward(input.clone());
        out.to_data()
            .assert_approx_eq::<FT>(&TensorData::from([[5.0067153]]), Tolerance::default());

        // Above it, the identity is substituted instead.
        let out = SoftplusConfig::new()
            .with_threshold(1.0)
            .init()
            .forward(input);
        out.to_data()
            .assert_approx_eq::<FT>(&TensorData::from([[5.0]]), Tolerance::default());
    }

    #[test]
    fn display() {
        let config = SoftplusConfig::new().init();
        assert_eq!(
            alloc::format!("{config}"),
            "Softplus {beta: 1, threshold: 20}"
        );
    }
}