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

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

type B = NdArray;

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

// -- FlexiPatchEmbed --------------------------------------------------------------

#[test]
fn patch_embed_output_shape() {
    let patch_size = 48;
    let n_rois = 10;
    let signal_length = 48 * 4; // 4 patches
    let embed_dim = 768;

    let pe = brainharmony::model::patch_embed::FlexiPatchEmbed::<B>::new(
        (n_rois, signal_length),
        patch_size,
        1,
        embed_dim,
        &device(),
    );

    let input = Tensor::<B, 4>::zeros([1, 1, n_rois, signal_length], &device());
    let output = pe.forward(input, None);

    let expected_patches = n_rois * (signal_length / patch_size);
    assert_eq!(output.dims(), [1, expected_patches, embed_dim]);
}

#[test]
fn patch_embed_num_patches() {
    let pe = brainharmony::model::patch_embed::FlexiPatchEmbed::<B>::new(
        (400, 864),
        48,
        1,
        768,
        &device(),
    );
    assert_eq!(pe.num_patches, 400 * 18);
    assert_eq!(pe.num_patches_2d, (400, 18));
}

#[test]
fn patch_embed_batch_forward() {
    let pe = brainharmony::model::patch_embed::FlexiPatchEmbed::<B>::new(
        (8, 96),
        48,
        1,
        128,
        &device(),
    );
    let input = Tensor::<B, 4>::zeros([4, 1, 8, 96], &device());
    let output = pe.forward(input, None);
    assert_eq!(output.dims(), [4, 8 * 2, 128]);
}

// -- Block ------------------------------------------------------------------------

#[test]
fn block_forward_preserves_shape() {
    let dim = 64;
    let num_heads = 4;
    let block = brainharmony::model::block::Block::<B>::new(
        dim, num_heads, 4.0, true, 1e-6, &device(),
    );

    let input = Tensor::<B, 3>::zeros([1, 100, dim], &device());
    let output = block.forward(input, None);
    assert_eq!(output.dims(), [1, 100, dim]);
}

#[test]
fn block_forward_batch() {
    let dim = 64;
    let block = brainharmony::model::block::Block::<B>::new(
        dim, 4, 4.0, true, 1e-6, &device(),
    );

    let input = Tensor::<B, 3>::zeros([2, 50, dim], &device());
    let output = block.forward(input, None);
    assert_eq!(output.dims(), [2, 50, dim]);
}

// -- Attention --------------------------------------------------------------------

#[test]
fn attention_forward_preserves_shape() {
    let dim = 64;
    let num_heads = 4;
    let attn = brainharmony::model::attention::Attention::<B>::new(
        dim, num_heads, true, &device(),
    );

    let input = Tensor::<B, 3>::zeros([1, 100, dim], &device());
    let output = attn.forward(input, None);
    assert_eq!(output.dims(), [1, 100, dim]);
}

#[test]
fn attention_head_dim_and_scale() {
    let dim = 128;
    let num_heads = 8;
    let attn = brainharmony::model::attention::Attention::<B>::new(
        dim, num_heads, true, &device(),
    );
    assert_eq!(attn.num_heads, 8);
    assert_eq!(attn.head_dim, 16);
    let expected_scale = (16.0f64).powf(-0.5) as f32;
    assert!((attn.scale - expected_scale).abs() < 1e-6);
}

#[test]
fn attention_no_qkv_bias() {
    let attn = brainharmony::model::attention::Attention::<B>::new(
        64, 4, false, &device(),
    );
    assert!(attn.qkv.bias.is_none());
}

// -- MLP --------------------------------------------------------------------------

#[test]
fn mlp_forward_preserves_shape() {
    let dim = 64;
    let hidden = 256;
    let mlp = brainharmony::model::feedforward::MLP::<B>::new(dim, hidden, &device());

    let input = Tensor::<B, 3>::zeros([1, 100, dim], &device());
    let output = mlp.forward(input);
    assert_eq!(output.dims(), [1, 100, dim]);
}

#[test]
fn mlp_zeros_produce_zeros() {
    let dim = 16;
    let mlp = brainharmony::model::feedforward::MLP::<B>::new(dim, 64, &device());

    let input = Tensor::<B, 3>::zeros([1, 5, dim], &device());
    let output = mlp.forward(input);

    use burn::prelude::ElementConversion;
    let max_abs: f32 = output.abs().max().into_scalar().elem();
    assert!(max_abs < 1e-6, "expected all zeros, max abs = {max_abs}");
}

// -- MLPHead ----------------------------------------------------------------------

#[test]
fn mlp_head_forward_shape() {
    let head = brainharmony::MLPHead::<B>::new(768, 384, 2, &device());
    let input = Tensor::<B, 2>::zeros([2, 768], &device());
    let output = head.forward(input);
    assert_eq!(output.dims(), [2, 2]);
}