luna-rs 0.1.0

LUNA EEG Foundation Model — inference in Rust (Burn and/or RLX)
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

/// Cross-Attention Block for LUNA's channel unification.
///
/// Python: `CrossAttentionBlock` in LUNA.py.
///
/// Python uses:
///   self.cross_attention = nn.MultiheadAttention(embed_dim, num_heads, dropout, batch_first=True)
///   self.ffn = Mlp(input_embed_dim, ff_dim, output_embed_dim, GELU, drop, norm_layer=LayerNorm)
///   self.keys_norm = LayerNorm(D)
///   self.values_norm = LayerNorm(D)
///   self.queries_norm = LayerNorm(D)
///   self.query_self_attn = nn.TransformerEncoder(
///       nn.TransformerEncoderLayer(D, nhead, activation='gelu', dim_feedforward=ff_dim,
///                                  batch_first=True, norm_first=True),
///       num_layers=3)
///
/// nn.MultiheadAttention stores a fused in_proj_weight \[3*D, D\] and in_proj_bias \[3*D\],
/// plus out_proj.weight \[D, D\] and out_proj.bias \[D\].
///
/// nn.TransformerEncoderLayer (norm_first=True) structure:
///   norm1 → self_attn (nn.MultiheadAttention) → dropout → residual
///   norm2 → linear1 → activation → dropout → linear2 → dropout → residual

use burn::prelude::*;
use burn::module::{Param, ParamId};
use burn::nn::{Linear, LinearConfig};
use burn::tensor::activation::{gelu, softmax};
use crate::model::norm::LunaLayerNorm;

// ── MultiheadAttention (fused in_proj) ──────────────────────────────────────

/// Matches nn.MultiheadAttention with fused in_proj_weight/bias.
/// Weight key: in_proj_weight \[3*D, D\], in_proj_bias \[3*D\], out_proj.weight \[D,D\], out_proj.bias \[D\]
#[derive(Module, Debug)]
pub struct FusedMultiheadAttention<B: Backend> {
    pub in_proj:  Linear<B>,   // [D, 3*D] (burn stores transposed)
    pub out_proj: Linear<B>,   // [D, D]
    pub n_heads:  usize,
    pub head_dim: usize,
}

impl<B: Backend> FusedMultiheadAttention<B> {
    pub fn new(dim: usize, n_heads: usize, device: &B::Device) -> Self {
        let head_dim = dim / n_heads;
        Self {
            in_proj:  LinearConfig::new(dim, dim * 3).with_bias(true).init(device),
            out_proj: LinearConfig::new(dim, dim).with_bias(true).init(device),
            n_heads,
            head_dim,
        }
    }

    /// q_input, k_input, v_input: [B, S_q, D] / [B, S_kv, D]
    /// Returns: (output [B, S_q, D], attn_weights [B, S_q, S_kv])
    pub fn forward(
        &self,
        q_input: Tensor<B, 3>,
        k_input: Tensor<B, 3>,
        v_input: Tensor<B, 3>,
    ) -> (Tensor<B, 3>, Tensor<B, 3>) {
        let [b, s_q, _] = q_input.dims();
        let s_kv = k_input.dims()[1];
        let (h, dh) = (self.n_heads, self.head_dim);
        let dim = h * dh;

        // Fused projection for Q
        let qkv_q = self.in_proj.forward(q_input);  // [B, S_q, 3*D]
        let q = qkv_q.narrow(2, 0, dim).reshape([b, s_q, h, dh]).swap_dims(1, 2);

        // Fused projection for K
        let qkv_k = self.in_proj.forward(k_input);
        let k = qkv_k.narrow(2, dim, dim).reshape([b, s_kv, h, dh]).swap_dims(1, 2);

        // Fused projection for V
        let qkv_v = self.in_proj.forward(v_input);
        let v = qkv_v.narrow(2, dim * 2, dim).reshape([b, s_kv, h, dh]).swap_dims(1, 2);

        let scale = (dh as f64).powf(-0.5) as f32;
        let attn_weights = softmax(q.matmul(k.transpose()).mul_scalar(scale), 3);  // [B, H, S_q, S_kv]

        let out = attn_weights.clone().matmul(v);  // [B, H, S_q, D]
        let out = self.out_proj.forward(
            out.swap_dims(1, 2).reshape([b, s_q, dim])
        );

        // Average attention weights across heads: [B, H, S_q, S_kv] → [B, S_q, S_kv]
        let avg_attn = attn_weights.mean_dim(1).squeeze::<3>();

        (out, avg_attn)
    }
}

// ── TransformerEncoderLayer (norm_first=True) ───────────────────────────────

/// Matches nn.TransformerEncoderLayer(norm_first=True, activation='gelu').
/// Structure:
///   x = x + self_attn(norm1(x))
///   x = x + ffn(norm2(x))
/// where ffn = linear1 → gelu → linear2
#[derive(Module, Debug)]
pub struct TransformerEncoderLayer<B: Backend> {
    pub norm1:     LunaLayerNorm<B>,
    pub self_attn: FusedMultiheadAttention<B>,
    pub norm2:     LunaLayerNorm<B>,
    pub linear1:   Linear<B>,
    pub linear2:   Linear<B>,
}

impl<B: Backend> TransformerEncoderLayer<B> {
    pub fn new(dim: usize, n_heads: usize, ff_dim: usize, norm_eps: f64, device: &B::Device) -> Self {
        Self {
            norm1:     LunaLayerNorm::new(dim, norm_eps, device),
            self_attn: FusedMultiheadAttention::new(dim, n_heads, device),
            norm2:     LunaLayerNorm::new(dim, norm_eps, device),
            linear1:   LinearConfig::new(dim, ff_dim).with_bias(true).init(device),
            linear2:   LinearConfig::new(ff_dim, dim).with_bias(true).init(device),
        }
    }

    /// x: [B, S, D] → [B, S, D]
    pub fn forward(&self, x: Tensor<B, 3>) -> Tensor<B, 3> {
        // norm_first: norm before attention
        let normed = self.norm1.forward(x.clone());
        let (attn_out, _) = self.self_attn.forward(normed.clone(), normed.clone(), normed);
        let x = x + attn_out;

        // FFN with norm_first
        let normed = self.norm2.forward(x.clone());
        let ff_out = self.linear2.forward(gelu(self.linear1.forward(normed)));
        x + ff_out
    }
}

// ── CrossAttentionBlock ─────────────────────────────────────────────────────

#[derive(Module, Debug)]
pub struct CrossAttentionBlock<B: Backend> {
    /// Learnable query prototypes: [1, Q, D].
    pub query_embed: Param<Tensor<B, 3>>,
    /// Temperature parameter (stored but unused in forward).
    pub temperature: Param<Tensor<B, 1>>,
    /// Cross-attention: nn.MultiheadAttention (fused in_proj)
    pub cross_attention: FusedMultiheadAttention<B>,
    /// FFN: timm.Mlp(D, ff_dim, D, GELU, drop, norm_layer=LayerNorm)
    /// At inference: fc1 → GELU → LayerNorm → fc2
    pub ffn_fc1:  Linear<B>,
    pub ffn_norm: LunaLayerNorm<B>,
    pub ffn_fc2:  Linear<B>,
    /// Pre-attention norms
    pub queries_norm: LunaLayerNorm<B>,
    pub keys_norm:    LunaLayerNorm<B>,
    pub values_norm:  LunaLayerNorm<B>,
    /// 3-layer TransformerEncoder for query self-attention
    pub self_attn_layers: Vec<TransformerEncoderLayer<B>>,
    pub num_queries: usize,
}

impl<B: Backend> CrossAttentionBlock<B> {
    pub fn new(
        num_queries: usize,
        input_embed_dim: usize,
        output_embed_dim: usize,
        num_heads: usize,
        ff_dim: usize,
        norm_eps: f64,
        device: &B::Device,
    ) -> Self {
        let self_attn_layers = (0..3)
            .map(|_| TransformerEncoderLayer::new(
                input_embed_dim, num_heads, ff_dim, norm_eps, device,
            ))
            .collect();

        Self {
            query_embed: Param::initialized(
                ParamId::new(),
                Tensor::zeros([1, num_queries, input_embed_dim], device),
            ),
            temperature: Param::initialized(
                ParamId::new(),
                Tensor::ones([1], device),
            ),
            cross_attention: FusedMultiheadAttention::new(input_embed_dim, num_heads, device),
            ffn_fc1:  LinearConfig::new(input_embed_dim, ff_dim).with_bias(true).init(device),
            ffn_norm: LunaLayerNorm::new(ff_dim, norm_eps, device),
            ffn_fc2:  LinearConfig::new(ff_dim, output_embed_dim).with_bias(true).init(device),
            queries_norm: LunaLayerNorm::new(input_embed_dim, norm_eps, device),
            keys_norm:    LunaLayerNorm::new(input_embed_dim, norm_eps, device),
            values_norm:  LunaLayerNorm::new(input_embed_dim, norm_eps, device),
            self_attn_layers,
            num_queries,
        }
    }

    /// x: [B_eff, C, D] — channel tokens for one time patch
    /// Returns: ([B_eff, Q, D], [B_eff, Q, C]) — unified queries + attention scores
    pub fn forward(&self, x: Tensor<B, 3>) -> (Tensor<B, 3>, Tensor<B, 3>) {
        let [batch_size, _num_channels, dim] = x.dims();

        // Expand learnable queries: [1, Q, D] → [B_eff, Q, D]
        let queries = self.query_embed.val()
            .expand([batch_size, self.num_queries, dim]);

        let queries = self.queries_norm.forward(queries);
        let keys = self.keys_norm.forward(x.clone());
        let values = self.values_norm.forward(x);

        // Cross-attention via nn.MultiheadAttention
        let (attention_out, attention_scores) =
            self.cross_attention.forward(queries, keys, values);
        // attention_out: [B_eff, Q, D], attention_scores: [B_eff, Q, C]

        // FFN with residual: timm.Mlp(D, ff_dim, D, GELU, norm_layer=LayerNorm)
        let ffn_out = self.ffn_fc2.forward(
            self.ffn_norm.forward(gelu(self.ffn_fc1.forward(attention_out.clone())))
        );
        let attention_out = ffn_out + attention_out;

        // 3-layer TransformerEncoder for query self-attention
        let mut out = attention_out;
        for layer in &self.self_attn_layers {
            out = layer.forward(out);
        }

        (out, attention_scores)
    }
}