millwright 0.2.1

A unified ML framework for Rust — proven Rust crates, assembled into one machine.
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

#![cfg(feature = "onnx")]
use millwright::prelude::*;

// A low-cardinality integer column `g` (one-hot inferred) + a numeric `x`.
fn categorical_data() -> (Dataset, Frame) {
    let mut rows = Vec::new();
    let mut y = Vec::new();
    for i in 0..30 {
        let g = (i % 3) as f64;
        let x = i as f64 * 0.1;
        rows.push(vec![x, g]);
        y.push(f64::from(u8::from(x + g > 2.0)));
    }
    let cols = vec!["x".to_string(), "g".to_string()];
    let ds = Dataset::new(Frame::from_rows(rows, cols.clone()).unwrap(), y).unwrap();
    let probe =
        Frame::from_rows(vec![vec![0.5, 1.0], vec![2.5, 2.0], vec![1.0, 0.0]], cols).unwrap();
    (ds, probe)
}

#[test]
fn onehot_scale_rf_serves_natively() {
    let (ds, probe) = categorical_data();
    let mut pipe = Pipeline::new()
        .step("encode", OneHotEncoder::infer())
        .step("scale", StandardScaler::new())
        .estimator("rf", RandomForest::new().n_trees(15).max_depth(4));
    pipe.fit(&ds).unwrap();
    let native = pipe.predict(&probe).unwrap();
    let path = std::env::temp_dir().join("mw_onehot_rf.onnx");
    pipe.export_onnx(&path).unwrap();
    let loaded = InferenceModel::load(&path).unwrap();
    assert_eq!(
        native,
        loaded.predict(&probe).unwrap(),
        "onehot+scale+rf via ONNX"
    );
    let _ = std::fs::remove_file(&path);
}

#[test]
fn onehot_scale_linear_serves_via_tract() {
    let (ds, probe) = categorical_data();
    let mut pipe = Pipeline::new()
        .step("encode", OneHotEncoder::infer())
        .step("scale", StandardScaler::new())
        .estimator("lr", LinearRegression::new());
    pipe.fit(&ds).unwrap();
    let native = pipe.predict(&probe).unwrap();
    let path = std::env::temp_dir().join("mw_onehot_lr.onnx");
    pipe.export_onnx(&path).unwrap();
    let via = InferenceModel::load(&path)
        .unwrap()
        .predict(&probe)
        .unwrap();
    for (a, b) in native.iter().zip(&via) {
        assert!((a - b).abs() < 1e-3, "native {a} vs onnx {b}");
    }
    let _ = std::fs::remove_file(&path);
}

#[test]
fn full_pipeline_impute_encode_scale_rf_round_trips() {
    // The realistic lifecycle pipeline: impute -> one-hot -> scale -> forest.
    let mut rows = Vec::new();
    let mut y = Vec::new();
    for i in 0..40 {
        let g = if i % 7 == 0 { f64::NAN } else { (i % 3) as f64 };
        let x = i as f64 * 0.1;
        rows.push(vec![x, g]);
        y.push(f64::from(u8::from(x > 2.0)));
    }
    let cols = vec!["x".to_string(), "g".to_string()];
    let ds = Dataset::new(Frame::from_rows(rows, cols.clone()).unwrap(), y).unwrap();
    let probe = Frame::from_rows(vec![vec![0.5, 1.0], vec![2.5, 2.0]], cols).unwrap();

    let mut pipe = Pipeline::new()
        .step("impute", SimpleImputer::median())
        .step("encode", OneHotEncoder::infer())
        .step("scale", StandardScaler::new())
        .estimator("rf", RandomForest::new().n_trees(20).max_depth(4));
    pipe.fit(&ds).unwrap();
    let native = pipe.predict(&probe).unwrap();

    let path = std::env::temp_dir().join("mw_full_pipe.onnx");
    pipe.export_onnx(&path).unwrap();
    let served = InferenceModel::load(&path)
        .unwrap()
        .predict(&probe)
        .unwrap();
    assert_eq!(
        native, served,
        "impute+encode+scale+rf must serve identically"
    );
    let _ = std::fs::remove_file(&path);
}

fn data_with_nans() -> (Dataset, Frame) {
    let mut rows = Vec::new();
    let mut y = Vec::new();
    for i in 0..20 {
        let a = if i % 5 == 0 { f64::NAN } else { i as f64 * 0.1 };
        rows.push(vec![a, i as f64 * 0.1]);
        y.push(0.0);
        rows.push(vec![9.0 + i as f64 * 0.1, 9.0 + i as f64 * 0.1]);
        y.push(1.0);
    }
    let cols = vec!["a".into(), "b".into()];
    let ds = Dataset::new(Frame::from_rows(rows, cols.clone()).unwrap(), y).unwrap();
    let probe = Frame::from_rows(vec![vec![f64::NAN, 0.2], vec![9.2, 9.3]], cols).unwrap();
    (ds, probe)
}

#[test]
fn impute_scale_rf_serves_natively() {
    let (ds, probe) = data_with_nans();
    let mut pipe = Pipeline::new()
        .step("impute", SimpleImputer::median())
        .step("scale", StandardScaler::new())
        .estimator("rf", RandomForest::new().n_trees(15).max_depth(4));
    pipe.fit(&ds).unwrap();
    let native = pipe.predict(&probe).unwrap();
    let path = std::env::temp_dir().join("mw_impute_rf.onnx");
    pipe.export_onnx(&path).unwrap();
    let loaded = InferenceModel::load(&path).unwrap();
    assert_eq!(
        native,
        loaded.predict(&probe).unwrap(),
        "impute+scale+rf via ONNX"
    );
    let _ = std::fs::remove_file(&path);
}

#[test]
fn impute_scale_linear_serves_via_tract() {
    let (ds, probe) = data_with_nans();
    let mut pipe = Pipeline::new()
        .step("impute", SimpleImputer::mean())
        .step("scale", StandardScaler::new())
        .estimator("lr", LinearRegression::new());
    pipe.fit(&ds).unwrap();
    let native = pipe.predict(&probe).unwrap();
    let path = std::env::temp_dir().join("mw_impute_lr.onnx");
    pipe.export_onnx(&path).unwrap();
    let loaded = InferenceModel::load(&path).unwrap();
    let via = loaded.predict(&probe).unwrap();
    for (a, b) in native.iter().zip(&via) {
        assert!((a - b).abs() < 1e-3, "native {a} vs onnx {b}");
    }
    let _ = std::fs::remove_file(&path);
}