rustyml 0.13.0

A high-performance machine learning & deep learning library in pure Rust, offering ML algorithms and neural network support
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

//! Binary cross-entropy loss for binary classification

use crate::error::Error;
use crate::neural_network::Tensor;
use crate::neural_network::losses::{clip_probabilities, validate_same_shape};
use crate::neural_network::traits::Loss;

/// Binary cross-entropy loss for binary classification
///
/// # Examples
///
/// ```rust
/// use rustyml::neural_network::losses::*;
/// use rustyml::neural_network::Tensor;
/// use rustyml::neural_network::traits::Loss;
/// use ndarray::array;
///
/// // Create a binary cross entropy loss function
/// let bce_loss = BinaryCrossEntropy::new();
///
/// // True binary labels (0 or 1)
/// let y_true = Tensor::from(array![[0.0f32, 1.0f32, 1.0f32, 0.0f32]].into_dyn());
///
/// // Predicted probabilities (between 0 and 1)
/// let y_pred = Tensor::from(array![[0.1f32, 0.9f32, 0.8f32, 0.2f32]].into_dyn());
///
/// // Compute the loss
/// let loss = bce_loss.compute_loss(&y_true, &y_pred).unwrap();
/// println!("Binary Cross Entropy Loss: {:.4}", loss);
///
/// // Compute gradients for backpropagation
/// let gradients = bce_loss.compute_grad(&y_true, &y_pred).unwrap();
/// println!("Gradients: {:?}", gradients);
/// ```
#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
pub struct BinaryCrossEntropy;

impl BinaryCrossEntropy {
    /// Creates a new `BinaryCrossEntropy`
    ///
    /// # Returns
    ///
    /// - `BinaryCrossEntropy` - a unit-like struct carrying no state
    pub fn new() -> Self {
        Self {}
    }
}

impl Loss for BinaryCrossEntropy {
    fn compute_loss(&self, y_true: &Tensor, y_pred: &Tensor) -> Result<f32, Error> {
        validate_same_shape(y_true, y_pred)?;

        // Clip into (0, 1) to avoid log(0)
        let y_pred_clipped = clip_probabilities(y_pred);

        // -1/n * sum[y_true * log(y_pred) + (1 - y_true) * log(1 - y_pred)]
        let log_pred = y_pred_clipped.mapv(|y_p| y_p.ln());
        let log_one_minus_pred = (1.0 - &y_pred_clipped).mapv(|y_p| y_p.ln());
        let losses = y_true * &log_pred + (1.0 - y_true) * &log_one_minus_pred;

        let n = losses.len() as f32;
        Ok(-losses.sum() / n)
    }

    fn compute_grad(&self, y_true: &Tensor, y_pred: &Tensor) -> Result<Tensor, Error> {
        validate_same_shape(y_true, y_pred)?;

        // Clip into (0, 1) to avoid division by zero
        let y_pred_clipped = clip_probabilities(y_pred);

        // gradient: -y_true/y_pred + (1 - y_true)/(1 - y_pred)
        let grad = -y_true / &y_pred_clipped + (1.0 - y_true) / (1.0 - &y_pred_clipped);

        // Average over element count
        let n = grad.len() as f32;
        Ok(grad / n)
    }
}