entrenar 0.7.9

Training & Optimization library with autograd, LoRA, quantization, and model merging
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
297
298
299
300
301
302
303
304
305
306
307
308

//! Property-based tests for distillation

#![allow(clippy::field_reassign_with_default)]

use super::*;
use ndarray::{array, Array2};
use proptest::prelude::*;

// Helper to generate random logits
fn logits_strategy(batch_size: usize, num_classes: usize) -> impl Strategy<Value = Array2<f32>> {
    prop::collection::vec(prop::collection::vec(-10.0f32..10.0, num_classes), batch_size).prop_map(
        move |data| {
            let flat: Vec<f32> = data.into_iter().flatten().collect();
            Array2::from_shape_vec((batch_size, num_classes), flat)
                .expect("operation should succeed")
        },
    )
}

// Helper to generate random hidden states
fn hiddens_strategy(batch_size: usize, hidden_dim: usize) -> impl Strategy<Value = Array2<f32>> {
    prop::collection::vec(prop::collection::vec(-5.0f32..5.0, hidden_dim), batch_size).prop_map(
        move |data| {
            let flat: Vec<f32> = data.into_iter().flatten().collect();
            Array2::from_shape_vec((batch_size, hidden_dim), flat)
                .expect("operation should succeed")
        },
    )
}

proptest! {
    /// Distillation loss should always be non-negative
    #[test]
    fn prop_distillation_loss_non_negative(
        temperature in 0.5f32..10.0,
        alpha in 0.0f32..=1.0,
    ) {
        let loss_fn = DistillationLoss::new(temperature, alpha);

        let student = array![[2.0, 1.0, 0.5], [1.5, 2.5, 0.8]];
        let teacher = array![[1.8, 1.1, 0.6], [1.6, 2.3, 0.9]];
        let labels = vec![0, 1];

        let loss = loss_fn.forward(&student, &teacher, &labels);

        prop_assert!(loss >= 0.0);
        prop_assert!(loss.is_finite());
    }

    /// Higher temperature should smooth distributions more
    #[test]
    fn prop_temperature_smooths_distributions(
        temp_low in 0.5f32..2.0,
        temp_high in 5.0f32..10.0,
    ) {
        let student = array![[10.0, 1.0, 0.1]]; // Sharp distribution
        let teacher = array![[5.0, 4.0, 3.0]];  // More balanced
        let labels = vec![0];

        let loss_low = DistillationLoss::new(temp_low, 1.0);
        let loss_high = DistillationLoss::new(temp_high, 1.0);

        let kl_low = loss_low.forward(&student, &teacher, &labels);
        let kl_high = loss_high.forward(&student, &teacher, &labels);

        // Both should be finite and non-negative
        prop_assert!(kl_low.is_finite() && kl_low >= 0.0);
        prop_assert!(kl_high.is_finite() && kl_high >= 0.0);
    }

    /// When alpha=0, should only use hard loss (independent of teacher)
    #[test]
    fn prop_alpha_zero_ignores_teacher(
        student in logits_strategy(2, 3),
    ) {
        let teacher1 = array![[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]];
        let teacher2 = array![[6.0, 5.0, 4.0], [3.0, 2.0, 1.0]];
        let labels = vec![0, 1];

        let loss_fn = DistillationLoss::new(2.0, 0.0); // alpha=0

        let loss1 = loss_fn.forward(&student, &teacher1, &labels);
        let loss2 = loss_fn.forward(&student, &teacher2, &labels);

        // Losses should be identical since teacher is ignored
        prop_assert!((loss1 - loss2).abs() < 1e-4);
    }

    /// When alpha=1, hard loss should not affect result
    #[test]
    fn prop_alpha_one_uses_only_soft_targets(
        student in logits_strategy(2, 3),
        teacher in logits_strategy(2, 3),
    ) {
        let labels1 = vec![0, 0];
        let labels2 = vec![1, 2];

        let loss_fn = DistillationLoss::new(2.0, 1.0); // alpha=1

        let loss1 = loss_fn.forward(&student, &teacher, &labels1);
        let loss2 = loss_fn.forward(&student, &teacher, &labels2);

        // Losses should be identical since labels are ignored
        prop_assert!((loss1 - loss2).abs() < 1e-4);
    }

    /// Ensemble with uniform weights should average teachers
    #[test]
    fn prop_ensemble_averages_uniformly(
        t1 in logits_strategy(2, 3),
        t2 in logits_strategy(2, 3),
    ) {
        let distiller = EnsembleDistiller::uniform(2, 2.0);
        let teachers = vec![t1.clone(), t2.clone()];

        let ensemble = distiller.combine_teachers(&teachers);

        // Check that ensemble is average of teachers
        for i in 0..t1.nrows() {
            for j in 0..t1.ncols() {
                let expected = f32::midpoint(t1[[i, j]], t2[[i, j]]);
                prop_assert!((ensemble[[i, j]] - expected).abs() < 1e-4);
            }
        }
    }

    /// MSE loss should be zero for identical inputs
    #[test]
    fn prop_mse_zero_for_identical(
        hidden in hiddens_strategy(2, 4),
    ) {
        use super::progressive::ProgressiveDistiller;

        let distiller = ProgressiveDistiller::uniform(1, 2.0);
        let hiddens_vec = vec![hidden.clone()];

        let loss = distiller.layer_wise_mse_loss(&hiddens_vec, &hiddens_vec);

        prop_assert!(loss.abs() < 1e-5);
    }

    /// Cosine similarity should be 1 for identical inputs
    #[test]
    fn prop_cosine_one_for_identical(
        hidden in hiddens_strategy(2, 4),
    ) {
        use super::progressive::ProgressiveDistiller;

        let distiller = ProgressiveDistiller::uniform(1, 2.0);
        let hiddens_vec = vec![hidden.clone()];

        let loss = distiller.layer_wise_cosine_loss(&hiddens_vec, &hiddens_vec);

        // Cosine loss = 1 - cosine_similarity
        // For identical inputs, cosine_similarity = 1, so loss = 0
        prop_assert!(loss.abs() < 1e-5);
    }

    /// MSE loss should be symmetric
    #[test]
    fn prop_mse_symmetric(
        h1 in hiddens_strategy(2, 4),
        h2 in hiddens_strategy(2, 4),
    ) {
        use super::progressive::ProgressiveDistiller;

        let distiller = ProgressiveDistiller::uniform(1, 2.0);

        let loss1 = distiller.layer_wise_mse_loss(&[h1.clone()], &[h2.clone()]);
        let loss2 = distiller.layer_wise_mse_loss(&[h2.clone()], &[h1.clone()]);

        prop_assert!((loss1 - loss2).abs() < 1e-5);
    }

    /// Cosine similarity should be symmetric
    #[test]
    fn prop_cosine_symmetric(
        h1 in hiddens_strategy(2, 4),
        h2 in hiddens_strategy(2, 4),
    ) {
        use super::progressive::ProgressiveDistiller;

        let distiller = ProgressiveDistiller::uniform(1, 2.0);

        let loss1 = distiller.layer_wise_cosine_loss(&[h1.clone()], &[h2.clone()]);
        let loss2 = distiller.layer_wise_cosine_loss(&[h2.clone()], &[h1.clone()]);

        prop_assert!((loss1 - loss2).abs() < 1e-5);
    }

    /// Progressive distillation loss should be finite and non-negative
    #[test]
    fn prop_progressive_loss_valid(
        temperature in 0.5f32..10.0,
        alpha in 0.0f32..=1.0,
        beta in 0.0f32..=1.0,
    ) {
        use super::progressive::ProgressiveDistiller;

        let distiller = ProgressiveDistiller::uniform(2, temperature);

        let student_logits = array![[2.0, 1.0, 0.5]];
        let teacher_logits = array![[1.8, 1.1, 0.6]];
        let student_hiddens = vec![array![[1.0, 2.0]], array![[3.0, 4.0]]];
        let teacher_hiddens = vec![array![[1.1, 2.1]], array![[3.1, 4.1]]];
        let labels = vec![0];

        let loss = distiller.combined_loss(
            &student_logits,
            &teacher_logits,
            &student_hiddens,
            &teacher_hiddens,
            &labels,
            alpha,
            beta,
        );

        prop_assert!(loss >= 0.0);
        prop_assert!(loss.is_finite());
    }

    /// Weighted ensemble should respect weights
    #[test]
    fn prop_ensemble_respects_weights(
        w1 in 0.1f32..10.0,
        w2 in 0.1f32..10.0,
    ) {
        let distiller = EnsembleDistiller::new(vec![w1, w2], 2.0);

        // Weights should be normalized
        let sum: f32 = distiller.weights.iter().sum();
        prop_assert!((sum - 1.0).abs() < 1e-5);

        // Individual weights should be proportional to input
        let expected_w1 = w1 / (w1 + w2);
        let expected_w2 = w2 / (w1 + w2);
        prop_assert!((distiller.weights[0] - expected_w1).abs() < 1e-5);
        prop_assert!((distiller.weights[1] - expected_w2).abs() < 1e-5);
    }
}

// Integration tests
#[cfg(test)]
mod integration_tests {
    use super::*;

    #[test]
    fn test_end_to_end_distillation() {
        // Simulate a simple distillation scenario
        let loss_fn = DistillationLoss::new(3.0, 0.7);

        // Teacher has confident predictions
        let teacher = array![[10.0, 2.0, 1.0], [1.0, 12.0, 2.0], [2.0, 1.0, 11.0]];

        // Student initially random
        let student = array![[5.0, 5.0, 5.0], [5.0, 5.0, 5.0], [5.0, 5.0, 5.0]];

        let labels = vec![0, 1, 2];

        let initial_loss = loss_fn.forward(&student, &teacher, &labels);

        // After "training", student should have improved predictions
        let student_improved = array![[9.0, 3.0, 2.0], [2.0, 10.0, 3.0], [3.0, 2.0, 9.0]];

        let final_loss = loss_fn.forward(&student_improved, &teacher, &labels);

        // Loss should decrease after improvement
        assert!(final_loss < initial_loss);
    }

    #[test]
    fn test_multi_teacher_ensemble() {
        let distiller = EnsembleDistiller::new(vec![2.0, 1.0, 1.0], 2.0);

        let t1 = array![[10.0, 1.0, 1.0]]; // Strong on class 0
        let t2 = array![[1.0, 10.0, 1.0]]; // Strong on class 1
        let t3 = array![[1.0, 1.0, 10.0]]; // Strong on class 2
        let teachers = vec![t1, t2, t3];

        let ensemble = distiller.combine_teachers(&teachers);

        // Ensemble should be influenced most by t1 (weight=2.0)
        assert!(ensemble[[0, 0]] > ensemble[[0, 1]]);
        assert!(ensemble[[0, 0]] > ensemble[[0, 2]]);
    }

    #[test]
    fn test_progressive_multi_layer() {
        use super::progressive::ProgressiveDistiller;

        // More weight on later layers
        let distiller = ProgressiveDistiller::new(vec![0.5, 1.0, 2.0], 2.0);

        let student_hiddens = vec![array![[1.0, 2.0]], array![[3.0, 4.0]], array![[5.0, 6.0]]];

        let teacher_hiddens = vec![
            array![[1.1, 2.1]],
            array![[3.1, 4.1]],
            array![[5.5, 6.5]], // Larger diff in last layer
        ];

        let loss = distiller.layer_wise_mse_loss(&student_hiddens, &teacher_hiddens);

        // Loss should be dominated by last layer due to higher weight
        assert!(loss > 0.0);
        assert!(loss.is_finite());
    }
}