Struct SupportVectorMachine

Source

pub struct SupportVectorMachine { /* private fields */ }

Implementations§

Source §

impl SupportVectorMachine

Source

pub fn new( class_num: usize, feature_num: usize, lr: f32, bsz: Option<usize>, lambda: f32, ) -> Self

Support Vector Machine for binary classification

only supports linear kernel, using SGD to optimize hinge loss

update = lambda * (regularization(w) or marginal width or so called structure loss) + 1/batch * Hinge Loss

bsz: default will be len(dataset)

Examples found in repository ?

examples/05_svm.rs (line 22)

6fn main() -> std::io::Result<()> {
7    let datas = vec![
8        vec![1.0, 3.0],
9        vec![2.0, 3.0],
10        vec![1.0, 2.0],
11        vec![4.0, 0.0],
12        vec![3.0, 0.0],
13        vec![3.0, -1.0],
14        vec![3.0, 0.5],
15    ];
16    let labels = vec![0usize, 0, 0, 1, 1, 1, 1];
17    let mut label_map = HashMap::new();
18    label_map.insert(0, "up left".to_string());
19    label_map.insert(1, "down right".to_string());
20    let dataset = Dataset::new(datas, labels, Some(label_map));
21
22    let mut model = SupportVectorMachine::new(2, dataset.feature_len(), 1e-1, None, 1.0);
23    model.train(dataset.clone(), 1000, SVMLoss::Hinge, true, true)?;
24
25    let (correct, acc) = evaluate(&dataset, &model);
26    println!("correct {correct}/{} acc {acc}", dataset.len());
27
28    Ok(())
29}

Source

pub fn train( &mut self, dataset: Dataset<usize>, epoch: usize, loss: SVMLoss, early_stop: bool, verbose: bool, ) -> Result<()>

train the SVM model

verbose: whether show the training info

Examples found in repository ?

examples/05_svm.rs (line 23)

6fn main() -> std::io::Result<()> {
7    let datas = vec![
8        vec![1.0, 3.0],
9        vec![2.0, 3.0],
10        vec![1.0, 2.0],
11        vec![4.0, 0.0],
12        vec![3.0, 0.0],
13        vec![3.0, -1.0],
14        vec![3.0, 0.5],
15    ];
16    let labels = vec![0usize, 0, 0, 1, 1, 1, 1];
17    let mut label_map = HashMap::new();
18    label_map.insert(0, "up left".to_string());
19    label_map.insert(1, "down right".to_string());
20    let dataset = Dataset::new(datas, labels, Some(label_map));
21
22    let mut model = SupportVectorMachine::new(2, dataset.feature_len(), 1e-1, None, 1.0);
23    model.train(dataset.clone(), 1000, SVMLoss::Hinge, true, true)?;
24
25    let (correct, acc) = evaluate(&dataset, &model);
26    println!("correct {correct}/{} acc {acc}", dataset.len());
27
28    Ok(())
29}