brainharmony 0.1.0

Brain-Harmony multimodal brain foundation model — inference in Rust with Burn ML
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

use burn::backend::NdArray;
use burn::prelude::*;

type B = NdArray;

fn device() -> burn::backend::ndarray::NdArrayDevice {
    burn::backend::ndarray::NdArrayDevice::Cpu
}

// -- full_context_mask ------------------------------------------------------------

#[test]
fn full_context_mask_shape() {
    let n_rois = 10;
    let n_time = 5;
    let mask = brainharmony::full_context_mask::<B>(n_rois, n_time, &device());
    assert_eq!(mask.dims(), [1, n_rois * n_time]);
}

#[test]
fn full_context_mask_indices_sequential() {
    let n_rois = 8;
    let n_time = 4;
    let n = n_rois * n_time;
    let mask = brainharmony::full_context_mask::<B>(n_rois, n_time, &device());
    let data = mask.squeeze::<1>().into_data();
    let indices: Vec<i64> = data.to_vec::<i64>().unwrap();
    let expected: Vec<i64> = (0..n as i64).collect();
    assert_eq!(indices, expected);
}

#[test]
fn full_context_mask_single_patch() {
    let mask = brainharmony::full_context_mask::<B>(1, 1, &device());
    assert_eq!(mask.dims(), [1, 1]);
    let data = mask.into_data();
    let indices: Vec<i64> = data.to_vec::<i64>().unwrap();
    assert_eq!(indices, vec![0]);
}

// -- random_block_mask ------------------------------------------------------------

#[test]
fn random_block_mask_indices_in_bounds() {
    let n_rois = 20;
    let n_time = 10;
    let n = n_rois * n_time;
    let mask = brainharmony::masks::random_block_mask::<B>(
        n_rois, n_time, 0.5, 0.5, 4, &device(),
    );
    assert_eq!(mask.dims()[0], 1);
    let data = mask.squeeze::<1>().into_data();
    let indices: Vec<i64> = data.to_vec::<i64>().unwrap();
    for &idx in &indices {
        assert!(idx >= 0 && idx < n as i64, "index {idx} out of range [0, {n})");
    }
}

#[test]
fn random_block_mask_respects_min_keep() {
    let n_rois = 5;
    let n_time = 5;
    let min_keep = 10;
    let mask = brainharmony::masks::random_block_mask::<B>(
        n_rois, n_time, 0.1, 0.1, min_keep, &device(),
    );
    let k = mask.dims()[1];
    assert!(k >= min_keep, "expected at least {min_keep} patches, got {k}");
}

#[test]
fn random_block_mask_sorted() {
    let mask = brainharmony::masks::random_block_mask::<B>(
        15, 8, 0.4, 0.3, 4, &device(),
    );
    let data = mask.squeeze::<1>().into_data();
    let indices: Vec<i64> = data.to_vec::<i64>().unwrap();
    for w in indices.windows(2) {
        assert!(w[0] <= w[1], "indices not sorted: {} > {}", w[0], w[1]);
    }
}

// -- jepa_masks -------------------------------------------------------------------

#[test]
fn jepa_masks_returns_enc_plus_three_pred() {
    let cfg = brainharmony::MaskConfig {
        n_rois: 20,
        n_time_patches: 8,
        seed: Some(42),
        ..Default::default()
    };
    let (enc_mask, pred_masks) = brainharmony::jepa_masks::<B>(&cfg, &device());
    assert_eq!(enc_mask.dims()[0], 1);
    assert_eq!(pred_masks.len(), 3);
}

#[test]
fn jepa_masks_indices_in_range() {
    let n_rois = 15;
    let n_time = 6;
    let n = n_rois * n_time;
    let cfg = brainharmony::MaskConfig {
        n_rois,
        n_time_patches: n_time,
        seed: Some(123),
        ..Default::default()
    };
    let (enc_mask, pred_masks) = brainharmony::jepa_masks::<B>(&cfg, &device());

    let check = |mask: Tensor<B, 2, Int>, label: &str| {
        let data = mask.squeeze::<1>().into_data();
        let indices: Vec<i64> = data.to_vec::<i64>().unwrap();
        for &idx in &indices {
            assert!(
                idx >= 0 && idx < n as i64,
                "{label}: index {idx} out of range [0, {n})"
            );
        }
    };

    check(enc_mask, "enc_mask");
    for (i, pm) in pred_masks.into_iter().enumerate() {
        check(pm, &format!("pred_mask[{i}]"));
    }
}

#[test]
fn jepa_masks_seed_reproducibility() {
    let cfg = brainharmony::MaskConfig {
        n_rois: 20,
        n_time_patches: 8,
        seed: Some(999),
        ..Default::default()
    };

    let (enc1, pred1) = brainharmony::jepa_masks::<B>(&cfg, &device());
    let (enc2, pred2) = brainharmony::jepa_masks::<B>(&cfg, &device());

    let to_vec = |t: Tensor<B, 2, Int>| -> Vec<i64> {
        t.squeeze::<1>().into_data().to_vec::<i64>().unwrap()
    };

    assert_eq!(to_vec(enc1), to_vec(enc2), "encoder masks differ with same seed");
    for (a, b) in pred1.into_iter().zip(pred2.into_iter()) {
        assert_eq!(to_vec(a), to_vec(b), "predictor masks differ with same seed");
    }
}

#[test]
fn jepa_masks_pred_not_in_encoder() {
    let n_rois = 20;
    let n_time = 8;
    let cfg = brainharmony::MaskConfig {
        n_rois,
        n_time_patches: n_time,
        seed: Some(77),
        ..Default::default()
    };
    let (enc_mask, pred_masks) = brainharmony::jepa_masks::<B>(&cfg, &device());

    let enc_data = enc_mask.squeeze::<1>().into_data();
    let enc_indices: Vec<i64> = enc_data.to_vec::<i64>().unwrap();
    let enc_set: std::collections::HashSet<i64> = enc_indices.into_iter().collect();

    for (i, pm) in pred_masks.into_iter().enumerate() {
        let data = pm.squeeze::<1>().into_data();
        let pred_indices: Vec<i64> = data.to_vec::<i64>().unwrap();
        for &idx in &pred_indices {
            assert!(
                !enc_set.contains(&idx),
                "pred_mask[{i}] index {idx} overlaps with encoder mask"
            );
        }
    }
}