rusty_machine/learning/toolkit/
cost_fn.rs

1//! Cost Functions.
2//!
3//! This module contains a number of structs implementing the `CostFunc` trait.
4//!
5//! These structs are used within Neural Networks and
6//! Generalized Linear Regression (not yet implemented).
7//!
8//! You can also create your own custom cost functions for use in your models.
9//! Just create a struct implementing the `CostFunc` trait.
10
11use linalg::{Matrix, BaseMatrix, BaseMatrixMut};
12use linalg::Vector;
13
14/// Trait for cost functions in models.
15pub trait CostFunc<T> {
16    /// The cost function.
17    fn cost(outputs: &T, targets: &T) -> f64;
18
19    /// The gradient of the cost function.
20    fn grad_cost(outputs: &T, targets: &T) -> T;
21}
22
23/// The mean squared error cost function.
24#[derive(Clone, Copy, Debug)]
25pub struct MeanSqError;
26
27// For generics we need a trait for "Hadamard product" here
28// Which is "Elementwise multiplication".
29impl CostFunc<Matrix<f64>> for MeanSqError {
30    fn cost(outputs: &Matrix<f64>, targets: &Matrix<f64>) -> f64 {
31        let diff = outputs - targets;
32        let sq_diff = &diff.elemul(&diff);
33
34        let n = diff.rows();
35
36        sq_diff.sum() / (2f64 * (n as f64))
37    }
38
39    fn grad_cost(outputs: &Matrix<f64>, targets: &Matrix<f64>) -> Matrix<f64> {
40        outputs - targets
41    }
42}
43
44impl CostFunc<Vector<f64>> for MeanSqError {
45    fn cost(outputs: &Vector<f64>, targets: &Vector<f64>) -> f64 {
46        let diff = outputs - targets;
47        let sq_diff = &diff.elemul(&diff);
48
49        let n = diff.size();
50
51        sq_diff.sum() / (2f64 * (n as f64))
52    }
53
54    fn grad_cost(outputs: &Vector<f64>, targets: &Vector<f64>) -> Vector<f64> {
55        outputs - targets
56    }
57}
58
59/// The cross entropy error cost function.
60#[derive(Clone, Copy, Debug)]
61pub struct CrossEntropyError;
62
63impl CostFunc<Matrix<f64>> for CrossEntropyError {
64    fn cost(outputs: &Matrix<f64>, targets: &Matrix<f64>) -> f64 {
65        // The cost for a single
66        let log_inv_output = (-outputs + 1f64).apply(&ln);
67        let log_output = outputs.clone().apply(&ln);
68
69        let mat_cost = targets.elemul(&log_output) + (-targets + 1f64).elemul(&log_inv_output);
70
71        let n = outputs.rows();
72
73        -(mat_cost.sum()) / (n as f64)
74    }
75
76    fn grad_cost(outputs: &Matrix<f64>, targets: &Matrix<f64>) -> Matrix<f64> {
77        (outputs - targets).elediv(&(outputs.elemul(&(-outputs + 1f64))))
78    }
79}
80
81impl CostFunc<Vector<f64>> for CrossEntropyError {
82    fn cost(outputs: &Vector<f64>, targets: &Vector<f64>) -> f64 {
83        // The cost for a single
84        let log_inv_output = (-outputs + 1f64).apply(&ln);
85        let log_output = outputs.clone().apply(&ln);
86
87        let mat_cost = targets.elemul(&log_output) + (-targets + 1f64).elemul(&log_inv_output);
88
89        let n = outputs.size();
90
91        -(mat_cost.sum()) / (n as f64)
92    }
93
94    fn grad_cost(outputs: &Vector<f64>, targets: &Vector<f64>) -> Vector<f64> {
95        (outputs - targets).elediv(&(outputs.elemul(&(-outputs + 1f64))))
96    }
97}
98
99/// Logarithm for applying within cost function.
100fn ln(x: f64) -> f64 {
101    x.ln()
102}
rusty_machine/learning/toolkit/cost_fn.rs

rusty_machine/learning/toolkit/
cost_fn.rs
