crfs 0.4.0

Pure Rust port of CRFsuite: a fast implementation of Conditional Random Fields (CRFs)
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
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189

use crfs::Attribute;
use crfs::train::Trainer;
use std::path::Path;

/// Test that AP algorithm can train and produce predictions
#[test]
fn test_ap_basic_training() {
    let xseq = vec![
        vec![Attribute::new("walk", 1.0), Attribute::new("shop", 0.5)],
        vec![Attribute::new("walk", 1.0)],
        vec![Attribute::new("walk", 1.0), Attribute::new("clean", 0.5)],
        vec![Attribute::new("shop", 0.5), Attribute::new("clean", 0.5)],
        vec![Attribute::new("walk", 0.5), Attribute::new("clean", 1.0)],
        vec![Attribute::new("clean", 1.0), Attribute::new("shop", 0.1)],
        vec![Attribute::new("walk", 1.0), Attribute::new("shop", 0.5)],
        vec![],
        vec![Attribute::new("clean", 1.0)],
    ];
    let yseq = [
        "sunny", "sunny", "sunny", "rainy", "rainy", "rainy", "sunny", "sunny", "rainy",
    ];

    let mut trainer = Trainer::averaged_perceptron();
    trainer.verbose(true);

    // Set parameters
    trainer.params_mut().set_max_iterations(50).unwrap();
    trainer.params_mut().set_epsilon(0.01).unwrap();
    trainer.params_mut().set_shuffle_seed(Some(1));

    // Add training data
    trainer.append(&xseq, &yseq).unwrap();

    // Train model
    let model_path = Path::new("/tmp/test_ap.crfsuite");
    trainer.train(model_path).unwrap();

    // Verify model file was created
    assert!(model_path.exists());

    // Load and test the model
    let model_data = std::fs::read(model_path).unwrap();
    let model = crfs::Model::new(&model_data).unwrap();
    let tagger = model.tagger().unwrap();
    let predicted = tagger.tag(&xseq).unwrap();

    // Check that predictions match training labels reasonably well
    let mut correct = 0;
    for (pred, true_label) in predicted.iter().zip(yseq.iter()) {
        if pred == true_label {
            correct += 1;
        }
    }
    let accuracy = correct as f64 / yseq.len() as f64;
    println!("AP Accuracy: {:.2}%", accuracy * 100.0);

    // Should achieve reasonable accuracy on training data
    assert!(
        accuracy > 0.7,
        "AP accuracy too low: {:.2}%",
        accuracy * 100.0
    );
}

/// Test AP with verbose output disabled
#[test]
fn test_ap_no_verbose() {
    let xseq = vec![
        vec![Attribute::new("walk", 1.0)],
        vec![Attribute::new("clean", 1.0)],
    ];
    let yseq = ["sunny", "rainy"];

    let mut trainer = Trainer::averaged_perceptron();
    trainer.verbose(false);
    trainer.params_mut().set_max_iterations(10).unwrap();
    trainer.params_mut().set_shuffle_seed(Some(1));
    trainer.append(&xseq, &yseq).unwrap();

    let model_path = Path::new("/tmp/test_ap_quiet.crfsuite");
    trainer.train(model_path).unwrap();
    assert!(model_path.exists());
}

/// Test AP convergence with epsilon
#[test]
fn test_ap_convergence() {
    // Simple data that should converge quickly
    let xseq = vec![
        vec![Attribute::new("a", 1.0)],
        vec![Attribute::new("b", 1.0)],
        vec![Attribute::new("a", 1.0)],
        vec![Attribute::new("b", 1.0)],
    ];
    let yseq = ["X", "Y", "X", "Y"];

    let mut trainer = Trainer::averaged_perceptron();
    trainer.verbose(true);
    trainer.params_mut().set_max_iterations(100).unwrap();
    trainer.params_mut().set_epsilon(0.000001).unwrap(); // Very low epsilon for convergence
    trainer.params_mut().set_shuffle_seed(Some(1));

    trainer.append(&xseq, &yseq).unwrap();

    let model_path = Path::new("/tmp/test_ap_converge.crfsuite");
    trainer.train(model_path).unwrap();

    // Verify perfect prediction on training data
    let model_data = std::fs::read(model_path).unwrap();
    let model = crfs::Model::new(&model_data).unwrap();
    let tagger = model.tagger().unwrap();
    let predicted = tagger.tag(&xseq).unwrap();

    assert_eq!(predicted, yseq);
}

/// Compare AP and LBFGS on the same dataset
#[test]
fn test_ap_vs_lbfgs() {
    let xseq = vec![
        vec![Attribute::new("walk", 1.0), Attribute::new("shop", 0.5)],
        vec![Attribute::new("walk", 1.0)],
        vec![Attribute::new("walk", 1.0), Attribute::new("clean", 0.5)],
        vec![Attribute::new("shop", 0.5), Attribute::new("clean", 0.5)],
        vec![Attribute::new("walk", 0.5), Attribute::new("clean", 1.0)],
        vec![Attribute::new("clean", 1.0), Attribute::new("shop", 0.1)],
    ];
    let yseq = ["sunny", "sunny", "sunny", "rainy", "rainy", "rainy"];

    // Train with AP
    let mut ap_trainer = Trainer::averaged_perceptron();
    ap_trainer.verbose(true);
    ap_trainer.params_mut().set_max_iterations(100).unwrap();
    ap_trainer.params_mut().set_epsilon(0.001).unwrap();
    ap_trainer.params_mut().set_shuffle_seed(Some(1));
    ap_trainer.append(&xseq, &yseq).unwrap();
    let ap_model_path = Path::new("/tmp/test_ap_compare.crfsuite");
    ap_trainer.train(ap_model_path).unwrap();

    // Train with LBFGS
    let mut lbfgs_trainer = Trainer::lbfgs();
    lbfgs_trainer.verbose(false);
    lbfgs_trainer.params_mut().set_max_iterations(50).unwrap();
    lbfgs_trainer.append(&xseq, &yseq).unwrap();
    let lbfgs_model_path = Path::new("/tmp/test_lbfgs_compare.crfsuite");
    lbfgs_trainer.train(lbfgs_model_path).unwrap();

    // Test both models
    let ap_model_data = std::fs::read(ap_model_path).unwrap();
    let ap_model = crfs::Model::new(&ap_model_data).unwrap();
    let ap_tagger = ap_model.tagger().unwrap();
    let ap_predicted = ap_tagger.tag(&xseq).unwrap();

    let lbfgs_model_data = std::fs::read(lbfgs_model_path).unwrap();
    let lbfgs_model = crfs::Model::new(&lbfgs_model_data).unwrap();
    let lbfgs_tagger = lbfgs_model.tagger().unwrap();
    let lbfgs_predicted = lbfgs_tagger.tag(&xseq).unwrap();

    // Calculate accuracies
    let ap_correct = ap_predicted
        .iter()
        .zip(yseq.iter())
        .filter(|(p, t)| p == t)
        .count();
    let lbfgs_correct = lbfgs_predicted
        .iter()
        .zip(yseq.iter())
        .filter(|(p, t)| p == t)
        .count();

    let ap_accuracy = ap_correct as f64 / yseq.len() as f64;
    let lbfgs_accuracy = lbfgs_correct as f64 / yseq.len() as f64;

    println!("AP Accuracy: {:.2}%", ap_accuracy * 100.0);
    println!("LBFGS Accuracy: {:.2}%", lbfgs_accuracy * 100.0);

    // Both should achieve reasonable accuracy
    // Note: AP may not always match LBFGS on small datasets
    assert!(
        ap_accuracy > 0.5,
        "AP accuracy too low: {:.2}%",
        ap_accuracy * 100.0
    );
    assert!(
        lbfgs_accuracy > 0.7,
        "LBFGS accuracy too low: {:.2}%",
        lbfgs_accuracy * 100.0
    );
}