oxirouter 0.1.0

Autonomous Semantic Federation Engine for the Edge - Learned source selection for SPARQL federated queries with context-awareness
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
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296

//! Tests that NeuralNetwork optimizer state survives save/load (Block U).
//!
//! After a save/load roundtrip the Adam moments, LR schedule epoch counter, and
//! early-stopping patience state must be fully restored.  Predictions must be
//! bit-for-bit identical (within float tolerance) to the pre-save network.

#[cfg(feature = "ml")]
mod tests {
    use oxirouter::ml::feature::FeatureVector;
    use oxirouter::ml::model::{Model, ModelPersistence, TrainingSample};
    use oxirouter::ml::neural::NeuralNetwork;
    use oxirouter::ml::optimizer::{AdamConfig, OptimizerType};
    use oxirouter::ml::schedule::EarlyStoppingConfig;

    /// Helper: serialize a NeuralNetwork to bytes via ModelPersistence trait.
    fn to_bytes(net: &NeuralNetwork) -> Vec<u8> {
        ModelPersistence::to_bytes(net)
    }

    /// Helper: deserialize a NeuralNetwork from bytes via ModelPersistence trait.
    fn from_bytes(bytes: &[u8]) -> oxirouter::Result<NeuralNetwork> {
        NeuralNetwork::from_bytes(bytes)
    }

    /// Build a FeatureVector of uniform value with the given dimension.
    fn uniform_fv(dim: usize, value: f32) -> FeatureVector {
        let mut fv = FeatureVector::new();
        for i in 0..dim {
            fv.add(format!("f{i}"), value);
        }
        fv
    }

    /// Build a TrainingSample directed at a specific source index within `num_sources`.
    fn make_sample(
        source_idx: usize,
        num_sources: usize,
        success: bool,
        latency_ms: u32,
    ) -> TrainingSample {
        let feature_value = source_idx as f32 / num_sources.max(1) as f32;
        let fv = uniform_fv(48, feature_value);
        let source_name = format!("source_{source_idx}");
        TrainingSample::new(
            fv,
            source_name,
            success,
            latency_ms,
            if success { 10 } else { 0 },
        )
    }

    // -------------------------------------------------------------------------
    // Test 1: Adam moments survive save/load — predictions identical
    // -------------------------------------------------------------------------
    #[test]
    fn test_adam_moments_survive_save_load() {
        let mut net = NeuralNetwork::new(48, &[32, 16], 3)
            .with_optimizer(OptimizerType::Adam(AdamConfig::default()));
        net.set_source_ids(vec![
            "source_0".to_string(),
            "source_1".to_string(),
            "source_2".to_string(),
        ]);

        // Train 50 steps to build up Adam moments
        for i in 0..50_usize {
            let s = make_sample(i % 3, 3, i % 2 == 0, 100);
            net.train(&[s]).expect("train should succeed");
        }

        // Save → load
        let bytes = to_bytes(&net);
        let restored = from_bytes(&bytes).expect("from_bytes failed");

        // Predictions must be identical
        let fv = uniform_fv(48, 0.5);
        let sources = vec![
            "source_0".to_string(),
            "source_1".to_string(),
            "source_2".to_string(),
        ];
        let source_refs: Vec<&String> = sources.iter().collect();

        let orig_pred = net
            .predict(&fv, &source_refs)
            .expect("predict (orig) failed");
        let rest_pred = restored
            .predict(&fv, &source_refs)
            .expect("predict (restored) failed");

        assert_eq!(orig_pred.len(), rest_pred.len());
        for (o, r) in orig_pred.iter().zip(rest_pred.iter()) {
            assert!(
                (o.1 - r.1).abs() < 1e-5,
                "prediction mismatch after save/load: orig={:.8}, restored={:.8}",
                o.1,
                r.1
            );
        }
    }

    // -------------------------------------------------------------------------
    // Test 2: optimizer_state is non-None after training; restored net can
    //         continue training immediately (no re-initialisation needed)
    // -------------------------------------------------------------------------
    #[test]
    fn test_optimizer_state_is_not_none_after_training() {
        let mut net = NeuralNetwork::new(48, &[32, 16], 2)
            .with_optimizer(OptimizerType::Adam(AdamConfig::default()));
        net.set_source_ids(vec!["source_0".to_string(), "source_1".to_string()]);

        let s = make_sample(0, 2, true, 50);
        net.train(&[s]).expect("initial train");

        // Save/load
        let bytes = to_bytes(&net);
        let mut restored = from_bytes(&bytes).expect("from_bytes");

        // Restored net must be able to train immediately without error
        let s2 = make_sample(1, 2, false, 200);
        restored
            .train(&[s2])
            .expect("training after restore should succeed");
    }

    // -------------------------------------------------------------------------
    // Test 3: early_stopping_state is preserved across save/load
    // -------------------------------------------------------------------------
    #[test]
    fn test_early_stopping_state_preserved() {
        let mut net = NeuralNetwork::new(48, &[16], 2).with_early_stopping(EarlyStoppingConfig {
            patience: 100,
            min_delta: 0.0,
        });
        net.set_source_ids(vec!["source_0".to_string(), "source_1".to_string()]);

        // Train a few steps
        for _ in 0..10_usize {
            let s = make_sample(0, 2, true, 100);
            net.train(&[s]).expect("train");
        }

        let bytes = to_bytes(&net);
        let mut restored = from_bytes(&bytes).expect("from_bytes");

        // Must still be able to train
        let s = make_sample(1, 2, true, 100);
        restored
            .train(&[s])
            .expect("train after restore should succeed");
    }

    // -------------------------------------------------------------------------
    // Test 4: SGD model (optimizer_state = None) loads cleanly (backward compat)
    // -------------------------------------------------------------------------
    #[test]
    fn test_sgd_model_loads_cleanly() {
        let mut net = NeuralNetwork::new(48, &[16], 2).with_optimizer(OptimizerType::SGD);
        net.set_source_ids(vec!["source_0".to_string(), "source_1".to_string()]);

        let s = make_sample(0, 2, true, 50);
        net.train(&[s]).expect("train");

        let bytes = to_bytes(&net);
        let restored = from_bytes(&bytes);
        assert!(
            restored.is_ok(),
            "SGD model should load cleanly: {restored:?}"
        );
    }

    // -------------------------------------------------------------------------
    // Test 5: weights are identical after roundtrip
    // -------------------------------------------------------------------------
    #[test]
    fn test_weights_identical_after_roundtrip() {
        let mut net = NeuralNetwork::new(48, &[32], 3)
            .with_optimizer(OptimizerType::Adam(AdamConfig::default()));
        net.set_source_ids(vec![
            "source_0".to_string(),
            "source_1".to_string(),
            "source_2".to_string(),
        ]);

        for _ in 0..20_usize {
            let s = make_sample(0, 3, true, 100);
            net.train(&[s]).expect("train");
        }

        let bytes = to_bytes(&net);
        let restored = from_bytes(&bytes).expect("deserialize");

        let fv = uniform_fv(48, 0.1);
        let sources = vec![
            "source_0".to_string(),
            "source_1".to_string(),
            "source_2".to_string(),
        ];
        let source_refs: Vec<&String> = sources.iter().collect();

        let p1 = net.predict(&fv, &source_refs).expect("predict1");
        let p2 = restored.predict(&fv, &source_refs).expect("predict2");

        assert_eq!(p1.len(), p2.len());
        for (a, b) in p1.iter().zip(p2.iter()) {
            assert!(
                (a.1 - b.1).abs() < 1e-5,
                "weight mismatch: {} vs {}",
                a.1,
                b.1
            );
        }
    }

    // -------------------------------------------------------------------------
    // Test 6: iterations counter is preserved
    // -------------------------------------------------------------------------
    #[test]
    fn test_iterations_preserved_after_roundtrip() {
        let mut net = NeuralNetwork::new(48, &[16], 2)
            .with_optimizer(OptimizerType::Adam(AdamConfig::default()));
        net.set_source_ids(vec!["source_0".to_string(), "source_1".to_string()]);

        for i in 0..30_usize {
            let s = make_sample(i % 2, 2, true, 100);
            net.train(&[s]).expect("train");
        }

        let pre_iterations = net.iterations();
        let bytes = to_bytes(&net);
        let restored = from_bytes(&bytes).expect("from_bytes");

        assert_eq!(
            pre_iterations,
            restored.iterations(),
            "iterations counter must survive save/load"
        );
    }

    // -------------------------------------------------------------------------
    // Test 7: differential — training after restore diverges from training a
    //         fresh network (proving moments are actually applied)
    // -------------------------------------------------------------------------
    #[test]
    fn test_continued_training_uses_restored_moments() {
        let mut net = NeuralNetwork::new(48, &[32, 16], 3)
            .with_optimizer(OptimizerType::Adam(AdamConfig::default()));
        net.set_source_ids(vec![
            "source_0".to_string(),
            "source_1".to_string(),
            "source_2".to_string(),
        ]);

        // Phase 1: 20 steps to build Adam moments
        for i in 0..20_usize {
            let s = make_sample(i % 3, 3, true, 100);
            net.train(&[s]).expect("phase1 train");
        }

        // Save/load → restored net has the same moments
        let bytes = to_bytes(&net);
        let mut restored = from_bytes(&bytes).expect("from_bytes");

        // Phase 2: 20 more steps on both
        for i in 0..20_usize {
            let s1 = make_sample(i % 3, 3, i % 2 == 0, 100);
            let s2 = make_sample(i % 3, 3, i % 2 == 0, 100);
            net.train(&[s1]).expect("phase2 net train");
            restored.train(&[s2]).expect("phase2 restored train");
        }

        // Both should produce identical predictions
        let fv = uniform_fv(48, 0.5);
        let sources = vec![
            "source_0".to_string(),
            "source_1".to_string(),
            "source_2".to_string(),
        ];
        let source_refs: Vec<&String> = sources.iter().collect();

        let p_net = net.predict(&fv, &source_refs).expect("predict net");
        let p_restored = restored
            .predict(&fv, &source_refs)
            .expect("predict restored");

        for (a, b) in p_net.iter().zip(p_restored.iter()) {
            assert!(
                (a.1 - b.1).abs() < 1e-4,
                "continued training diverged after restore: net={:.8}, restored={:.8}",
                a.1,
                b.1
            );
        }
    }
}