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
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
use super::feature_gen::{FeatureGenerator, FeatureType};
use crate::dataset::Instance;
/// CRF scoring context for computing state/transition scores and Viterbi decoding.
pub struct ScoreContext {
/// Number of labels
num_labels: usize,
/// State scores [time][label]
state_scores: Vec<Vec<f64>>,
/// Transition scores [prev_label][label]
trans_scores: Vec<Vec<f64>>,
}
impl ScoreContext {
/// Create a new scoring context
pub fn new(num_labels: usize, max_items: usize) -> Self {
Self {
num_labels,
state_scores: vec![vec![0.0; num_labels]; max_items],
trans_scores: vec![vec![0.0; num_labels]; num_labels],
}
}
/// Compute state and transition scores for an instance
pub fn compute_scores(&mut self, inst: &Instance, fgen: &FeatureGenerator) {
let seq_len = inst.num_items as usize;
// Reset scores
for t in 0..seq_len {
for l in 0..self.num_labels {
self.state_scores[t][l] = 0.0;
}
}
for i in 0..self.num_labels {
for j in 0..self.num_labels {
self.trans_scores[i][j] = 0.0;
}
}
// Compute state scores
for t in 0..seq_len {
for attr in &inst.items[t] {
let aid = attr.id as usize;
if aid < fgen.attr_refs.len() {
for &fid in &fgen.attr_refs[aid].fids {
let feature = &fgen.features[fid as usize];
if feature.ftype == FeatureType::State {
let lid = feature.dst as usize;
self.state_scores[t][lid] += feature.weight * attr.value;
}
}
}
}
}
// Compute transition scores
for l in 0..self.num_labels {
if l < fgen.label_refs.len() {
for &fid in &fgen.label_refs[l].fids {
let feature = &fgen.features[fid as usize];
if feature.ftype == FeatureType::Transition {
let prev_lid = feature.src as usize;
let lid = feature.dst as usize;
self.trans_scores[prev_lid][lid] += feature.weight;
}
}
}
}
}
/// Viterbi decoding to find the best label sequence.
///
/// This method assumes that [`compute_scores`](Self::compute_scores) has already been
/// called. It uses dynamic programming to find the label sequence with maximum score.
///
/// # Arguments
///
/// * `seq_len` - The length of the sequence
///
/// # Returns
///
/// A vector of label IDs representing the best label sequence
pub fn viterbi_decode(&self, seq_len: usize) -> Vec<u32> {
let mut delta = vec![vec![f64::NEG_INFINITY; self.num_labels]; seq_len];
let mut psi = vec![vec![0usize; self.num_labels]; seq_len];
// Initialization: delta[0][l] = state_score[0][l]
delta[0][..self.num_labels].copy_from_slice(&self.state_scores[0][..self.num_labels]);
// Forward pass: find best previous state for each current state
for t in 1..seq_len {
for l in 0..self.num_labels {
let (best_prev, best_score) = (0..self.num_labels)
.map(|prev_l| {
let score = delta[t - 1][prev_l]
+ self.trans_scores[prev_l][l]
+ self.state_scores[t][l];
(prev_l, score)
})
.max_by(|a, b| a.1.partial_cmp(&b.1).unwrap())
.unwrap_or((0, f64::NEG_INFINITY));
delta[t][l] = best_score;
psi[t][l] = best_prev;
}
}
// Backtrack: find the best path
let mut labels = vec![0u32; seq_len];
// Find the best final state
let best_final_label = delta[seq_len - 1][..self.num_labels]
.iter()
.enumerate()
.max_by(|a, b| a.1.partial_cmp(b.1).unwrap())
.map(|(l, _)| l)
.unwrap_or(0);
labels[seq_len - 1] = best_final_label as u32;
// Backtrack through the sequence
for t in (0..seq_len - 1).rev() {
labels[t] = psi[t + 1][labels[t + 1] as usize] as u32;
}
labels
}
/// Compute the score for a given label sequence using pre-computed scores.
///
/// This method assumes that [`compute_scores`](Self::compute_scores) has already been
/// called. It sums state and transition scores along the provided path.
pub fn sequence_score(&self, labels: &[u32]) -> f64 {
if labels.is_empty() {
return 0.0;
}
let mut score = 0.0;
for t in 0..labels.len() {
let label = labels[t] as usize;
score += self.state_scores[t][label];
if t > 0 {
let prev_label = labels[t - 1] as usize;
score += self.trans_scores[prev_label][label];
}
}
score
}
}
fn logsumexp(values: &[f64]) -> f64 {
if values.is_empty() {
return f64::NEG_INFINITY;
}
let max_val = values.iter().copied().fold(f64::NEG_INFINITY, f64::max);
if max_val.is_infinite() {
return max_val;
}
let sum: f64 = values.iter().map(|&v| (v - max_val).exp()).sum();
max_val + sum.ln()
}
/// CRF context for forward-backward inference.
pub struct ForwardBackwardContext {
scores: ScoreContext,
/// Forward variables (in log space) [time][label]
alpha: Vec<Vec<f64>>,
/// Backward variables (in log space) [time][label]
beta: Vec<Vec<f64>>,
/// Marginal probabilities [time][label]
marginals: Vec<Vec<f64>>,
/// Transition marginals [time][prev_label][label]
trans_marginals: Vec<Vec<Vec<f64>>>,
/// Reusable buffer for log-sum-exp computations
log_buffer: Vec<f64>,
}
impl ForwardBackwardContext {
/// Create a new CRF context
pub fn new(num_labels: usize, max_items: usize) -> Self {
Self {
scores: ScoreContext::new(num_labels, max_items),
alpha: vec![vec![f64::NEG_INFINITY; num_labels]; max_items],
beta: vec![vec![f64::NEG_INFINITY; num_labels]; max_items],
marginals: vec![vec![0.0; num_labels]; max_items],
trans_marginals: vec![vec![vec![0.0; num_labels]; num_labels]; max_items],
log_buffer: vec![0.0; num_labels],
}
}
/// Compute state and transition scores for an instance
pub fn compute_scores(&mut self, inst: &Instance, fgen: &FeatureGenerator) {
self.scores.compute_scores(inst, fgen);
}
/// Forward algorithm in log space
pub fn forward(&mut self, seq_len: usize) -> f64 {
let num_labels = self.scores.num_labels;
// Initialize at t=0
for l in 0..num_labels {
self.alpha[0][l] = self.scores.state_scores[0][l];
}
// Forward recursion - reuse log_buffer to avoid allocations
for t in 1..seq_len {
for l in 0..num_labels {
for prev_l in 0..num_labels {
self.log_buffer[prev_l] = self.alpha[t - 1][prev_l]
+ self.scores.trans_scores[prev_l][l]
+ self.scores.state_scores[t][l];
}
self.alpha[t][l] = logsumexp(&self.log_buffer[..num_labels]);
}
}
// Compute log partition function
for l in 0..num_labels {
self.log_buffer[l] = self.alpha[seq_len - 1][l];
}
logsumexp(&self.log_buffer[..num_labels])
}
/// Backward algorithm in log space
pub fn backward(&mut self, seq_len: usize) {
let num_labels = self.scores.num_labels;
// Initialize at t=T-1
for l in 0..num_labels {
self.beta[seq_len - 1][l] = 0.0; // log(1) = 0
}
// Backward recursion - reuse log_buffer to avoid allocations
for t in (0..seq_len - 1).rev() {
for l in 0..num_labels {
for next_l in 0..num_labels {
self.log_buffer[next_l] = self.beta[t + 1][next_l]
+ self.scores.trans_scores[l][next_l]
+ self.scores.state_scores[t + 1][next_l];
}
self.beta[t][l] = logsumexp(&self.log_buffer[..num_labels]);
}
}
}
/// Compute marginal probabilities
pub fn compute_marginals(&mut self, seq_len: usize, log_z: f64) {
let num_labels = self.scores.num_labels;
// State marginals
for t in 0..seq_len {
for l in 0..num_labels {
let log_marginal = self.alpha[t][l] + self.beta[t][l] - log_z;
self.marginals[t][l] = log_marginal.exp();
}
}
// Transition marginals
for t in 1..seq_len {
for prev_l in 0..num_labels {
for l in 0..num_labels {
let log_marginal = self.alpha[t - 1][prev_l]
+ self.scores.trans_scores[prev_l][l]
+ self.scores.state_scores[t][l]
+ self.beta[t][l]
- log_z;
self.trans_marginals[t][prev_l][l] = log_marginal.exp();
}
}
}
}
/// Compute expected feature counts into a pre-allocated vector
pub fn expected_counts_into(
&self,
inst: &Instance,
fgen: &FeatureGenerator,
counts: &mut [f64],
) {
let seq_len = inst.num_items as usize;
// State feature expectations
for t in 0..seq_len {
for attr in &inst.items[t] {
let aid = attr.id as usize;
if aid < fgen.attr_refs.len() {
for &fid in &fgen.attr_refs[aid].fids {
let feature = &fgen.features[fid as usize];
if feature.ftype == FeatureType::State {
let lid = feature.dst as usize;
counts[fid as usize] += self.marginals[t][lid] * attr.value;
}
}
}
}
}
// Transition feature expectations
for t in 1..seq_len {
for prev_l in 0..self.scores.num_labels {
if prev_l < fgen.label_refs.len() {
for &fid in &fgen.label_refs[prev_l].fids {
let feature = &fgen.features[fid as usize];
if feature.ftype == FeatureType::Transition {
let lid = feature.dst as usize;
counts[fid as usize] += self.trans_marginals[t][prev_l][lid];
}
}
}
}
}
}
/// Compute observed feature counts into a pre-allocated vector
pub fn observed_counts_into(
&self,
inst: &Instance,
fgen: &FeatureGenerator,
counts: &mut [f64],
) {
let seq_len = inst.num_items as usize;
// State feature observations
for t in 0..seq_len {
let label_id = inst.labels[t]; // u32 label ID for this timestep
for attr in &inst.items[t] {
let aid = attr.id as usize;
if aid < fgen.attr_refs.len() {
for &fid in &fgen.attr_refs[aid].fids {
let feature = &fgen.features[fid as usize];
// feature.dst is the target label ID for this state feature
if feature.ftype == FeatureType::State && feature.dst == label_id {
counts[fid as usize] += attr.value;
}
}
}
}
}
// Transition feature observations
for t in 1..seq_len {
let prev_label = inst.labels[t - 1];
let label = inst.labels[t];
let prev_l = prev_label as usize;
if prev_l < fgen.label_refs.len() {
for &fid in &fgen.label_refs[prev_l].fids {
let feature = &fgen.features[fid as usize];
if feature.ftype == FeatureType::Transition
&& feature.src == prev_label
&& feature.dst == label
{
counts[fid as usize] += 1.0;
}
}
}
}
}
/// Compute log-likelihood for an instance using pre-computed scores.
///
/// This method assumes that [`compute_scores`](Self::compute_scores) has already been
/// called and that `forward()` has been run to populate `self.alpha`.
/// It computes the score of the correct label sequence and subtracts
/// the partition function (log_z) to get the log-likelihood.
///
/// # Arguments
///
/// * `inst` - The training instance
/// * `log_z` - The log partition function from the forward algorithm
pub fn log_likelihood(&self, inst: &Instance, log_z: f64) -> f64 {
let seq_len = inst.num_items as usize;
// Compute score of the correct label sequence
let mut score = 0.0;
for t in 0..seq_len {
let label = inst.labels[t] as usize;
score += self.scores.state_scores[t][label];
if t > 0 {
let prev_label = inst.labels[t - 1] as usize;
score += self.scores.trans_scores[prev_label][label];
}
}
score - log_z
}
}