burn_core/nn/
dropout.rs

1use crate as burn;
2
3use crate::config::Config;
4use crate::module::{Content, DisplaySettings, Module, ModuleDisplay};
5use crate::tensor::backend::Backend;
6use crate::tensor::{Distribution, Tensor};
7
8/// Configuration to create a [Dropout](Dropout) layer using the [init function](DropoutConfig::init).
9#[derive(Config, Debug)]
10pub struct DropoutConfig {
11    /// The probability of randomly zeroes some elements of the input tensor during training.
12    pub prob: f64,
13}
14
15/// Set at random some elements of the input tensor to zero during training.
16///
17/// This is an effective regularization technique as describe in the paper
18/// [Improving neural networks by preventing co-adaptation of feature detectors](https://arxiv.org/abs/1207.0580).
19///
20/// The input is also scaled during training to `1 / (1 - prob_keep)`.
21///
22/// Should be created with [DropoutConfig].
23#[derive(Module, Clone, Debug)]
24#[module(custom_display)]
25pub struct Dropout {
26    /// The probability of randomly zeroes some elements of the input tensor during training.
27    pub prob: f64,
28}
29
30impl DropoutConfig {
31    /// Initialize a new [dropout](Dropout) module.
32    pub fn init(&self) -> Dropout {
33        if self.prob < 0.0 || self.prob > 1.0 {
34            panic!(
35                "Dropout probability should be between 0 and 1, but got {}",
36                self.prob
37            );
38        }
39        Dropout { prob: self.prob }
40    }
41}
42
43impl Dropout {
44    /// Applies the forward pass on the input tensor.
45    ///
46    /// See [Dropout](Dropout) for more information.
47    ///
48    /// # Shapes
49    ///
50    /// - input: `[..., any]`
51    /// - output: `[..., any]`
52    pub fn forward<B: Backend, const D: usize>(&self, input: Tensor<B, D>) -> Tensor<B, D> {
53        if !B::ad_enabled() || self.prob == 0.0 {
54            return input;
55        }
56
57        let prob_keep = 1.0 - self.prob;
58        let random = input.random_like(Distribution::Bernoulli(prob_keep));
59        let x = input * random;
60
61        x * (1.0 / prob_keep)
62    }
63}
64
65impl ModuleDisplay for Dropout {
66    fn custom_settings(&self) -> Option<DisplaySettings> {
67        DisplaySettings::new()
68            .with_new_line_after_attribute(false)
69            .optional()
70    }
71
72    fn custom_content(&self, content: Content) -> Option<Content> {
73        content.add("prob", &self.prob).optional()
74    }
75}
76
77#[cfg(test)]
78mod tests {
79    use super::*;
80    use crate::tensor::Shape;
81
82    #[cfg(feature = "std")]
83    use crate::{TestAutodiffBackend, TestBackend};
84
85    #[cfg(not(feature = "std"))]
86    use crate::TestBackend;
87
88    #[cfg(feature = "std")]
89    #[test]
90    fn with_ad_backend_should_mark_input() {
91        let tensor =
92            Tensor::<TestAutodiffBackend, 2>::ones(Shape::new([100, 100]), &Default::default());
93        let dropout = DropoutConfig::new(0.5).init();
94
95        let output = dropout.forward(tensor.clone());
96
97        assert_ne!(tensor.to_data(), output.to_data());
98    }
99
100    #[test]
101    fn without_ad_backend_should_not_change_input() {
102        let tensor = Tensor::<TestBackend, 2>::ones(Shape::new([100, 100]), &Default::default());
103        let dropout = DropoutConfig::new(0.5).init();
104
105        let output = dropout.forward(tensor.clone());
106
107        assert_eq!(tensor.to_data(), output.to_data());
108    }
109
110    #[test]
111    fn display() {
112        let config = DropoutConfig::new(0.5);
113        let layer = config.init();
114
115        assert_eq!(alloc::format!("{layer}"), "Dropout {prob: 0.5}");
116    }
117
118    #[test]
119    #[should_panic = "Dropout probability should be between 0 and 1,"]
120    fn dropout_prob_invalid() {
121        let config = DropoutConfig::new(-10.);
122        let _layer = config.init();
123    }
124}
burn_core/nn/dropout.rs

burn_core/nn/
dropout.rs
