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

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::celu;

/// CELU (Continuously Differentiable Exponential Linear Unit) layer.
///
/// Applies the CELU function element-wise:
/// `celu(x) = max(0, x) + min(0, alpha * (exp(x / alpha) - 1))`
///
/// Should be created with [CeluConfig](CeluConfig).
#[derive(Module, Debug)]
#[module(custom_display)]
pub struct Celu {
    /// The alpha value for the CELU formulation.
    pub alpha: f64,
}

/// Configuration to create a [Celu](Celu) layer using the [init function](CeluConfig::init).
#[derive(Config, Debug)]
pub struct CeluConfig {
    /// The finite, non-zero alpha value for the CELU formulation. Default is 1.0.
    #[config(default = "1.0")]
    pub alpha: f64,
}

impl CeluConfig {
    /// Initialize a new [Celu](Celu) Layer
    pub fn init(&self) -> Celu {
        if !self.alpha.is_finite() || self.alpha == 0.0 {
            panic!(
                "CELU alpha must be finite and non-zero, but got {}",
                self.alpha
            );
        }

        Celu { alpha: self.alpha }
    }
}

impl ModuleDisplay for Celu {
    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("alpha", &self.alpha).optional()
    }
}

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

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

    #[test]
    fn test_celu_forward() {
        let device = Default::default();
        let model = CeluConfig::new().init();
        let input = Tensor::<2>::from_data(TensorData::from([[0.5, -0.5, -1.0]]), &device);
        let out = model.forward(input);
        // celu(0.5, 1) = 0.5
        // celu(-0.5, 1) = 1 * (exp(-0.5) - 1) = -0.393469
        // celu(-1.0, 1) = 1 * (exp(-1) - 1) = -0.632121
        let expected = TensorData::from([[0.5, -0.393469, -0.632121]]);
        out.to_data()
            .assert_approx_eq::<FT>(&expected, Tolerance::default());
    }

    #[test]
    fn test_celu_with_alpha() {
        let device = Default::default();
        let model = CeluConfig::new().with_alpha(2.0).init();
        let input = Tensor::<2>::from_data(TensorData::from([[0.0, -2.0]]), &device);
        let out = model.forward(input);
        // celu(0, 2) = 0
        // celu(-2, 2) = 2 * (exp(-1) - 1) = -1.264241
        let expected = TensorData::from([[0.0, -1.264241]]);
        out.to_data()
            .assert_approx_eq::<FT>(&expected, Tolerance::default());
    }

    #[test]
    fn display() {
        let config = CeluConfig::new().init();
        assert_eq!(alloc::format!("{config}"), "Celu {alpha: 1}");
    }

    #[test]
    #[should_panic(expected = "CELU alpha must be finite and non-zero")]
    fn zero_alpha_should_panic() {
        CeluConfig::new().with_alpha(0.0).init();
    }

    #[test]
    fn negative_alpha_is_valid() {
        let device = Default::default();
        let model = CeluConfig::new().with_alpha(-1.0).init();
        let input = Tensor::<2>::from_data(TensorData::from([[-1.0, 0.0, 1.0]]), &device);
        let expected = TensorData::from([[-1.7182819, 0.0, 1.0]]);

        model
            .forward(input)
            .to_data()
            .assert_approx_eq::<FT>(&expected, Tolerance::default());
    }

    #[test]
    #[should_panic(expected = "CELU alpha must be finite and non-zero")]
    fn nan_alpha_should_panic() {
        CeluConfig::new().with_alpha(f64::NAN).init();
    }

    #[test]
    #[should_panic(expected = "CELU alpha must be finite and non-zero")]
    fn infinite_alpha_should_panic() {
        CeluConfig::new().with_alpha(f64::INFINITY).init();
    }
}