concision_core/nn/
dropout.rs

1/*
2    Appellation: dropout <module>
3    Contrib: FL03 <jo3mccain@icloud.com>
4*/
5#![allow(unused_imports)]
6use crate::Forward;
7use nd::prelude::*;
8use nd::{DataOwned, ScalarOperand};
9#[cfg(feature = "rand")]
10use ndrand::{rand_distr::Bernoulli, RandomExt};
11use num::traits::Num;
12
13#[cfg(feature = "rand")]
14pub fn dropout<A, S, D>(array: &ArrayBase<S, D>, p: f64) -> Array<A, D>
15where
16    A: Num + ScalarOperand,
17    D: Dimension,
18    S: DataOwned<Elem = A>,
19{
20    // Create a Bernoulli distribution for dropout
21    let distribution = Bernoulli::new(p).unwrap();
22
23    // Create a mask of the same shape as the input array
24    let mask: Array<bool, D> = Array::random(array.dim(), distribution);
25    let mask = mask.mapv(|x| if x { A::zero() } else { A::one() });
26
27    // Element-wise multiplication to apply dropout
28    array * mask
29}
30
31/// [Dropout] randomly zeroizes elements with a given probability (`p`).
32pub trait Dropout {
33    type Output;
34
35    fn dropout(&self, p: f64) -> Self::Output;
36}
37
38/// The [DropoutLayer] layer is randomly zeroizes inputs with a given probability (`p`).
39/// This regularization technique is often used to prevent overfitting.
40///
41///
42/// ### Config
43///
44/// - (p) Probability of dropping an element
45#[derive(Clone, Copy, Debug, PartialEq, PartialOrd)]
46#[cfg_attr(feature = "serde", derive(serde::Deserialize, serde::Serialize))]
47pub struct DropoutLayer {
48    pub(crate) p: f64,
49}
50
51/*
52 ************* Implementations *************
53*/
54#[cfg(feature = "rand")]
55impl<A, S, D> Dropout for ArrayBase<S, D>
56where
57    A: Num + ScalarOperand,
58    D: Dimension,
59    S: DataOwned<Elem = A>,
60{
61    type Output = Array<A, D>;
62
63    fn dropout(&self, p: f64) -> Self::Output {
64        dropout(self, p)
65    }
66}
67
68impl DropoutLayer {
69    pub fn new(p: f64) -> Self {
70        Self { p }
71    }
72
73    pub fn scale(&self) -> f64 {
74        (1f64 - self.p).recip()
75    }
76}
77
78impl Default for DropoutLayer {
79    fn default() -> Self {
80        Self::new(0.5)
81    }
82}
83
84#[cfg(feature = "rand")]
85impl<A, S, D> Forward<ArrayBase<S, D>> for DropoutLayer
86where
87    A: Num + ScalarOperand,
88    D: Dimension,
89    S: DataOwned<Elem = A>,
90{
91    type Output = Array<A, D>;
92
93    fn forward(&self, input: &ArrayBase<S, D>) -> Self::Output {
94        input.dropout(self.p)
95    }
96}
concision_core/nn/dropout.rs

concision_core/nn/
dropout.rs
