anno 0.10.0

NER, coreference resolution, relation extraction, PII detection, and zero-shot entity types
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

use super::*;

/// Span representation layer that computes span embeddings from token-level hidden states.
pub struct SpanRepLayer {
    /// MLP for projecting start positions (Linear -> ReLU -> Dropout -> Linear)
    project_start_0: Linear,
    project_start_3: Linear,
    /// MLP for projecting end positions
    project_end_0: Linear,
    project_end_3: Linear,
    /// Final projection layer
    out_project_0: Linear,
    out_project_3: Linear,
    hidden_size: usize,
}

#[cfg(feature = "candle")]
impl SpanRepLayer {
    /// Create a new span representation layer from GLiNER weights.
    ///
    /// GLiNER uses the SpanMarker architecture with:
    /// - project_start: Linear(D, 4D) -> ReLU -> Dropout -> Linear(4D, D)
    /// - project_end: Linear(D, 4D) -> ReLU -> Dropout -> Linear(4D, D)
    /// - out_project: Linear(2D, 4D) -> ReLU -> Dropout -> Linear(4D, D)
    pub fn new(hidden_size: usize, _max_width: usize, vb: VarBuilder) -> Result<Self> {
        // Load project_start MLP (layers 0 and 3, indices match PyTorch Sequential)
        // Hidden multiplier is 4x for these models
        let project_start_0 = linear(hidden_size, hidden_size * 4, vb.pp("project_start").pp("0"))
            .map_err(|e| Error::Retrieval(format!("SpanRepLayer project_start.0: {}", e)))?;
        let project_start_3 = linear(hidden_size * 4, hidden_size, vb.pp("project_start").pp("3"))
            .map_err(|e| Error::Retrieval(format!("SpanRepLayer project_start.3: {}", e)))?;

        // Load project_end MLP
        let project_end_0 = linear(hidden_size, hidden_size * 4, vb.pp("project_end").pp("0"))
            .map_err(|e| Error::Retrieval(format!("SpanRepLayer project_end.0: {}", e)))?;
        let project_end_3 = linear(hidden_size * 4, hidden_size, vb.pp("project_end").pp("3"))
            .map_err(|e| Error::Retrieval(format!("SpanRepLayer project_end.3: {}", e)))?;

        // Load out_project MLP (input is 2*hidden_size = concatenated start+end)
        let out_project_0 = linear(
            hidden_size * 2,
            hidden_size * 4,
            vb.pp("out_project").pp("0"),
        )
        .map_err(|e| Error::Retrieval(format!("SpanRepLayer out_project.0: {}", e)))?;
        let out_project_3 = linear(hidden_size * 4, hidden_size, vb.pp("out_project").pp("3"))
            .map_err(|e| Error::Retrieval(format!("SpanRepLayer out_project.3: {}", e)))?;

        Ok(Self {
            project_start_0,
            project_start_3,
            project_end_0,
            project_end_3,
            out_project_0,
            out_project_3,
            hidden_size,
        })
    }

    /// Compute span embeddings from token embeddings using SpanMarker approach.
    ///
    /// # Arguments
    /// * `token_embeddings` - [batch, seq_len, hidden]
    /// * `span_indices` - [batch, num_spans, 2] (start, end)
    ///
    /// # Returns
    /// [batch, num_spans, hidden]
    pub fn forward(&self, token_embeddings: &Tensor, span_indices: &Tensor) -> Result<Tensor> {
        let (batch_size, seq_len, _hidden) = token_embeddings
            .dims3()
            .map_err(|e| Error::Parse(format!("token_embeddings dims: {}", e)))?;
        let (_, _num_spans, _) = span_indices
            .dims3()
            .map_err(|e| Error::Parse(format!("span_indices dims: {}", e)))?;

        // Project start and end representations for all tokens first
        // project_start: Linear -> ReLU (at layer 2, which is dropout in PyTorch) -> Linear
        let start_rep = self.project_start_0.forward(token_embeddings)?;
        let start_rep = start_rep.relu()?;
        let start_rep = self.project_start_3.forward(&start_rep)?;

        let end_rep = self.project_end_0.forward(token_embeddings)?;
        let end_rep = end_rep.relu()?;
        let end_rep = self.project_end_3.forward(&end_rep)?;

        // Extract start and end indices
        let start_idx = span_indices.i((.., .., 0))?.to_dtype(DType::U32)?;
        let end_idx = span_indices.i((.., .., 1))?.to_dtype(DType::U32)?;

        let mut span_embs = Vec::new();

        for b in 0..batch_size {
            let batch_start_rep = start_rep.i(b)?;
            let batch_end_rep = end_rep.i(b)?;
            let batch_starts = start_idx.i(b)?;
            let batch_ends = end_idx.i(b)?;

            // Clamp indices to valid range
            let max_idx = (seq_len - 1) as u32;
            let batch_starts = batch_starts.clamp(0f64, max_idx as f64)?;
            let batch_ends = batch_ends.clamp(0f64, max_idx as f64)?;

            // Extract start and end representations for each span
            let start_span_rep = batch_start_rep
                .index_select(&batch_starts.to_dtype(DType::U32)?, 0)
                .map_err(|e| Error::Parse(format!("start index_select: {}", e)))?;
            let end_span_rep = batch_end_rep
                .index_select(&batch_ends.to_dtype(DType::U32)?, 0)
                .map_err(|e| Error::Parse(format!("end index_select: {}", e)))?;

            // Concatenate and apply ReLU
            let cat = Tensor::cat(&[&start_span_rep, &end_span_rep], D::Minus1)?;
            let cat = cat.relu()?;

            // Apply output projection: Linear -> ReLU -> Linear
            let out = self.out_project_0.forward(&cat)?;
            let out = out.relu()?;
            let out = self.out_project_3.forward(&out)?;

            span_embs.push(out);
        }

        Tensor::stack(&span_embs, 0).map_err(|e| Error::Parse(format!("stack span_embs: {}", e)))
    }
}

// =============================================================================
// Label Encoder (prompt_rep_layer in GLiNER)
// =============================================================================

/// Projects label embeddings to matching space.
/// Maps to GLiNER's prompt_rep_layer MLP.
#[cfg(feature = "candle")]
pub struct LabelEncoder {
    linear_0: Linear,
    linear_3: Linear,
}

#[cfg(feature = "candle")]
impl LabelEncoder {
    /// Create a new label encoder from GLiNER prompt_rep_layer weights.
    ///
    /// GLiNER structure: Linear(D, 4D) -> ReLU -> Dropout -> Linear(4D, D)
    pub fn new(hidden_size: usize, vb: VarBuilder) -> Result<Self> {
        let linear_0 = linear(hidden_size, hidden_size * 4, vb.pp("0"))
            .map_err(|e| Error::Retrieval(format!("LabelEncoder.0: {}", e)))?;
        let linear_3 = linear(hidden_size * 4, hidden_size, vb.pp("3"))
            .map_err(|e| Error::Retrieval(format!("LabelEncoder.3: {}", e)))?;

        Ok(Self { linear_0, linear_3 })
    }

    /// Project label embeddings to matching space.
    pub fn forward(&self, label_embeddings: &Tensor) -> Result<Tensor> {
        let out = self
            .linear_0
            .forward(label_embeddings)
            .map_err(|e| Error::Parse(format!("label projection 0: {}", e)))?;
        let out = out
            .relu()
            .map_err(|e| Error::Parse(format!("label relu: {}", e)))?;
        self.linear_3
            .forward(&out)
            .map_err(|e| Error::Parse(format!("label projection 3: {}", e)))
    }
}

// =============================================================================
// Span-Label Matcher
// =============================================================================

/// Computes similarity between spans and labels.
#[cfg(feature = "candle")]
pub struct SpanLabelMatcher {
    temperature: f64,
}

#[cfg(feature = "candle")]
impl SpanLabelMatcher {
    /// Create a new span-label matcher with temperature scaling.
    pub fn new(temperature: f64) -> Self {
        Self { temperature }
    }

    /// Match spans to labels via cosine similarity.
    ///
    /// # Arguments
    /// * `span_embeddings` - [batch, num_spans, hidden]
    /// * `label_embeddings` - [num_labels, hidden]
    ///
    /// # Returns
    /// [batch, num_spans, num_labels] scores in [0, 1]
    pub fn forward(&self, span_embeddings: &Tensor, label_embeddings: &Tensor) -> Result<Tensor> {
        let span_norm = l2_normalize(span_embeddings, D::Minus1)?;
        let label_norm = l2_normalize(label_embeddings, D::Minus1)?;

        let batch_size = span_norm.dims()[0];
        let label_t = label_norm.t()?;
        let label_t = label_t.unsqueeze(0)?.broadcast_as((
            batch_size,
            label_t.dims()[0],
            label_t.dims()[1],
        ))?;

        let scores = span_norm.matmul(&label_t)?;
        let scaled = (scores * self.temperature)?;

        candle_nn::ops::sigmoid(&scaled).map_err(|e| Error::Parse(format!("sigmoid: {}", e)))
    }
}

#[cfg(feature = "candle")]
pub(crate) fn l2_normalize(tensor: &Tensor, dim: D) -> Result<Tensor> {
    let norm = tensor.sqr()?.sum(dim)?.sqrt()?;
    let norm = norm.unsqueeze(D::Minus1)?;
    // Clamp norm to prevent division by zero
    let norm_clamped = norm
        .clamp(1e-12, f32::MAX)
        .map_err(|e| Error::Parse(format!("clamp: {}", e)))?;
    tensor
        .broadcast_div(&norm_clamped)
        .map_err(|e| Error::Parse(format!("l2_normalize: {}", e)))
}