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
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207

use crfs::{Attribute, Trainer};
use std::path::Path;

#[test]
fn test_l2sgd_basic_training() {
    // Create simple training data
    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 L2SGD
    let mut trainer = Trainer::l2sgd();
    trainer.verbose(true);
    trainer.params_mut().set_c2(1.0).unwrap();
    trainer.params_mut().set_max_iterations(50).unwrap();
    trainer.params_mut().set_period(10).unwrap();

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

    // Train model
    let model_path = Path::new("/tmp/test_l2sgd.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 (p, t) in predicted.iter().zip(yseq.iter()) {
        if p == t {
            correct += 1;
        }
    }
    let accuracy = correct as f64 / yseq.len() as f64;
    println!("L2SGD Accuracy: {:.2}%", accuracy * 100.0);
    assert!(accuracy > 0.5, "L2SGD accuracy too low");
}

#[test]
fn test_l2sgd_calibration() {
    // Test that calibration works
    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::l2sgd();
    trainer.verbose(true);
    trainer.params_mut().set_c2(1.0).unwrap();
    trainer.params_mut().set_max_iterations(20).unwrap();
    trainer.params_mut().set_calibration_samples(4).unwrap();
    trainer.params_mut().set_calibration_candidates(5).unwrap();
    trainer.append(&xseq, &yseq).unwrap();

    let model_path = Path::new("/tmp/test_l2sgd_calibration.crfsuite");
    trainer.train(model_path).unwrap();
    assert!(model_path.exists());
}

#[test]
fn test_l2sgd_vs_lbfgs() {
    // Compare L2SGD with LBFGS on the same dataset
    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 L2SGD
    let mut l2sgd_trainer = Trainer::l2sgd();
    l2sgd_trainer.verbose(false);
    l2sgd_trainer.params_mut().set_c2(1.0).unwrap();
    l2sgd_trainer.params_mut().set_max_iterations(100).unwrap();
    l2sgd_trainer.params_mut().set_period(10).unwrap();
    l2sgd_trainer.append(&xseq, &yseq).unwrap();
    let l2sgd_model_path = Path::new("/tmp/test_l2sgd_compare.crfsuite");
    l2sgd_trainer.train(l2sgd_model_path).unwrap();

    // Train with LBFGS
    let mut lbfgs_trainer = Trainer::lbfgs();
    lbfgs_trainer.verbose(false);
    lbfgs_trainer.params_mut().set_c1(0.0).unwrap();
    lbfgs_trainer.params_mut().set_c2(1.0).unwrap();
    lbfgs_trainer.params_mut().set_max_iterations(100).unwrap();
    lbfgs_trainer.append(&xseq, &yseq).unwrap();
    let lbfgs_model_path = Path::new("/tmp/test_lbfgs_compare_l2sgd.crfsuite");
    lbfgs_trainer.train(lbfgs_model_path).unwrap();

    // Test both models
    let l2sgd_model_data = std::fs::read(l2sgd_model_path).unwrap();
    let l2sgd_model = crfs::Model::new(&l2sgd_model_data).unwrap();
    let l2sgd_tagger = l2sgd_model.tagger().unwrap();
    let l2sgd_predicted = l2sgd_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 l2sgd_correct = l2sgd_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 l2sgd_accuracy = l2sgd_correct as f64 / yseq.len() as f64;
    let lbfgs_accuracy = lbfgs_correct as f64 / yseq.len() as f64;

    println!("L2SGD Accuracy: {:.2}%", l2sgd_accuracy * 100.0);
    println!("LBFGS Accuracy: {:.2}%", lbfgs_accuracy * 100.0);

    // Both should achieve reasonable accuracy
    assert!(
        l2sgd_accuracy > 0.5,
        "L2SGD accuracy too low: {:.2}%",
        l2sgd_accuracy * 100.0
    );
    assert!(
        lbfgs_accuracy > 0.7,
        "LBFGS accuracy too low: {:.2}%",
        lbfgs_accuracy * 100.0
    );
}

#[test]
fn test_l2sgd_parameter_validation() {
    let mut trainer = Trainer::l2sgd();

    // Valid parameters
    assert!(trainer.params_mut().set_c2(1.0).is_ok());
    assert!(trainer.params_mut().set_period(10).is_ok());
    assert!(trainer.params_mut().set_delta(1e-5).is_ok());
    assert!(trainer.params_mut().set_calibration_eta(0.1).is_ok());
    assert!(trainer.params_mut().set_calibration_rate(2.0).is_ok());

    // Invalid parameters
    assert!(trainer.params_mut().set_period(0).is_err()); // period must be positive
    assert!(trainer.params_mut().set_delta(0.0).is_err()); // delta must be positive
    assert!(trainer.params_mut().set_calibration_eta(0.0).is_err()); // eta must be positive
    assert!(trainer.params_mut().set_calibration_rate(1.0).is_err()); // rate must be > 1.0
}

#[test]
fn test_l2sgd_convergence() {
    // Test convergence behavior
    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)],
        vec![Attribute::new("a", 1.0)],
        vec![Attribute::new("b", 1.0)],
    ];
    let yseq = ["X", "Y", "X", "Y", "X", "Y"];

    let mut trainer = Trainer::l2sgd();
    trainer.verbose(true);
    trainer.params_mut().set_c2(1.0).unwrap();
    trainer.params_mut().set_max_iterations(100).unwrap();
    trainer.params_mut().set_period(5).unwrap();
    trainer.params_mut().set_delta(1e-4).unwrap();
    trainer.append(&xseq, &yseq).unwrap();

    let model_path = Path::new("/tmp/test_l2sgd_converge.crfsuite");
    trainer.train(model_path).unwrap();

    // Verify model can predict
    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();

    // Should achieve reasonable accuracy
    let correct = predicted
        .iter()
        .zip(yseq.iter())
        .filter(|(p, t)| p == t)
        .count();
    let accuracy = correct as f64 / yseq.len() as f64;
    assert!(accuracy > 0.5, "Accuracy too low: {:.2}%", accuracy * 100.0);
}