File size: 17,880 Bytes
ed10267
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
#!/usr/bin/env python3
from __future__ import annotations

import json
import tempfile
from pathlib import Path
from typing import Any

import numpy as np
from huggingface_hub import HfApi, hf_hub_download
from transformers import AutoConfig, AutoTokenizer

TOKENIZER_FILES = [
    "tokenizer_config.json",
    "tokenizer.json",
    "special_tokens_map.json",
    "vocab.txt",
    "vocab.json",
    "merges.txt",
    "added_tokens.json",
    "sentencepiece.bpe.model",
    "spiece.model",
]
DEFAULT_LABEL_MAX_SPAN_TOKENS = {
    # Token-piece limits, not word limits. These need to reflect how the
    # underlying tokenizer actually fragments compact identifiers.
    "PPSN": 9,
    "POSTCODE": 7,
    "PHONE_NUMBER": 10,
    "PASSPORT_NUMBER": 8,
    "BANK_ROUTING_NUMBER": 5,
    "ACCOUNT_NUMBER": 19,
    "CREDIT_DEBIT_CARD": 12,
    "SWIFT_BIC": 8,
    "EMAIL": 15,
    "FIRST_NAME": 5,
    "LAST_NAME": 8,
}
DEFAULT_LABEL_MIN_NONSPACE_CHARS = {
    "PPSN": 8,
    "POSTCODE": 6,
    "PHONE_NUMBER": 7,
    "PASSPORT_NUMBER": 7,
    "BANK_ROUTING_NUMBER": 6,
    "ACCOUNT_NUMBER": 6,
    "CREDIT_DEBIT_CARD": 12,
    "SWIFT_BIC": 8,
    "EMAIL": 6,
    "FIRST_NAME": 2,
    "LAST_NAME": 2,
}
WHITESPACE_BRIDGE_LABELS = {
    "PPSN",
    "POSTCODE",
    "PHONE_NUMBER",
    "PASSPORT_NUMBER",
    "BANK_ROUTING_NUMBER",
    "ACCOUNT_NUMBER",
    "CREDIT_DEBIT_CARD",
    "SWIFT_BIC",
    "EMAIL",
}
CONSERVATIVE_BOUNDARY_REFINEMENT_LABELS = {
    "PPSN",
    "POSTCODE",
    "PHONE_NUMBER",
    "PASSPORT_NUMBER",
    "BANK_ROUTING_NUMBER",
    "ACCOUNT_NUMBER",
    "CREDIT_DEBIT_CARD",
    "SWIFT_BIC",
    "EMAIL",
}
OUTPUT_PRIORITY = {
    "PPSN": 0,
    "PASSPORT_NUMBER": 1,
    "ACCOUNT_NUMBER": 2,
    "BANK_ROUTING_NUMBER": 3,
    "CREDIT_DEBIT_CARD": 4,
    "PHONE_NUMBER": 5,
    "SWIFT_BIC": 6,
    "POSTCODE": 7,
    "EMAIL": 8,
    "FIRST_NAME": 9,
    "LAST_NAME": 10,
}


def normalize_entity_name(label: str) -> str:
    label = (label or "").strip()
    if label.startswith("B-") or label.startswith("I-"):
        label = label[2:]
    return label.upper()


def _sanitize_tokenizer_dir(tokenizer_path: Path) -> str:
    tokenizer_cfg_path = tokenizer_path / "tokenizer_config.json"
    if not tokenizer_cfg_path.exists():
        return str(tokenizer_path)
    data = json.loads(tokenizer_cfg_path.read_text(encoding="utf-8"))
    if "fix_mistral_regex" not in data:
        return str(tokenizer_path)
    tmpdir = Path(tempfile.mkdtemp(prefix="openmed_span_tokenizer_"))
    keep = set(TOKENIZER_FILES)
    for child in tokenizer_path.iterdir():
        if child.is_file() and child.name in keep:
            (tmpdir / child.name).write_bytes(child.read_bytes())
    data.pop("fix_mistral_regex", None)
    (tmpdir / "tokenizer_config.json").write_text(json.dumps(data, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
    return str(tmpdir)


def safe_auto_tokenizer(tokenizer_ref: str):
    tokenizer_path = Path(tokenizer_ref)
    if tokenizer_path.exists():
        tokenizer_ref = _sanitize_tokenizer_dir(tokenizer_path)
    else:
        api = HfApi()
        files = set(api.list_repo_files(repo_id=tokenizer_ref, repo_type="model"))
        tmpdir = Path(tempfile.mkdtemp(prefix="openmed_remote_span_tokenizer_"))
        copied = False
        for name in TOKENIZER_FILES:
            if name not in files:
                continue
            src = hf_hub_download(repo_id=tokenizer_ref, filename=name, repo_type="model")
            (tmpdir / Path(name).name).write_bytes(Path(src).read_bytes())
            copied = True
        if copied:
            tokenizer_ref = _sanitize_tokenizer_dir(tmpdir)

    try:
        return AutoTokenizer.from_pretrained(tokenizer_ref, use_fast=True, fix_mistral_regex=True)
    except Exception:
        pass
    try:
        return AutoTokenizer.from_pretrained(tokenizer_ref, use_fast=True, fix_mistral_regex=False)
    except TypeError:
        pass
    try:
        return AutoTokenizer.from_pretrained(tokenizer_ref, use_fast=True)
    except Exception:
        return AutoTokenizer.from_pretrained(tokenizer_ref, use_fast=False)


def label_names_from_config(config) -> list[str]:
    names = list(getattr(config, "span_label_names", []))
    if not names:
        raise ValueError("Missing span_label_names in config")
    return [normalize_entity_name(name) for name in names]


def label_thresholds_from_config(config, default_threshold: float) -> dict[str, float]:
    raw = getattr(config, "span_label_thresholds", None) or {}
    out = {normalize_entity_name(key): float(value) for key, value in raw.items()}
    for label in label_names_from_config(config):
        out.setdefault(label, float(default_threshold))
    return out


def token_label_thresholds_from_config(config, default_threshold: float) -> dict[str, float]:
    raw = getattr(config, "token_label_thresholds", None) or {}
    out = {normalize_entity_name(key): float(value) for key, value in raw.items()}
    for label in label_names_from_config(config):
        out.setdefault(label, float(default_threshold))
    return out


def token_extend_thresholds_from_config(config, default_fraction: float = 0.6) -> dict[str, float]:
    raw = getattr(config, "token_extend_thresholds", None) or {}
    out = {normalize_entity_name(key): float(value) for key, value in raw.items()}
    for label in label_names_from_config(config):
        out.setdefault(label, max(0.0, min(1.0, float(token_label_thresholds_from_config(config, 0.5).get(label, 0.5)) * default_fraction)))
    return out


def boundary_label_thresholds_from_config(config, default_threshold: float = 0.0) -> dict[str, float]:
    raw = getattr(config, "boundary_label_thresholds", None) or {}
    out = {normalize_entity_name(key): float(value) for key, value in raw.items()}
    for label in label_names_from_config(config):
        out.setdefault(label, float(default_threshold))
    return out


def label_max_span_tokens_from_config(config) -> dict[str, int]:
    raw = getattr(config, "span_label_max_span_tokens", None) or {}
    out = {normalize_entity_name(key): int(value) for key, value in raw.items()}
    for label, value in DEFAULT_LABEL_MAX_SPAN_TOKENS.items():
        out.setdefault(label, value)
    for label in label_names_from_config(config):
        out.setdefault(label, 8)
    return out


def label_min_nonspace_chars_from_config(config) -> dict[str, int]:
    raw = getattr(config, "span_label_min_nonspace_chars", None) or {}
    out = {normalize_entity_name(key): int(value) for key, value in raw.items()}
    for label, value in DEFAULT_LABEL_MIN_NONSPACE_CHARS.items():
        out.setdefault(label, value)
    for label in label_names_from_config(config):
        out.setdefault(label, 1)
    return out


def overlaps(a: dict, b: dict) -> bool:
    return not (a["end"] <= b["start"] or b["end"] <= a["start"])


def dedupe_spans(spans: list[dict]) -> list[dict]:
    ordered = sorted(
        spans,
        key=lambda item: (-float(item.get("score", 0.0)), item["start"], item["end"], OUTPUT_PRIORITY.get(item["label"], 99)),
    )
    kept = []
    for span in ordered:
        if any(overlaps(span, other) for other in kept):
            continue
        kept.append(span)
    kept.sort(key=lambda item: (item["start"], item["end"], OUTPUT_PRIORITY.get(item["label"], 99)))
    return kept


def _valid_offset(offset: tuple[int, int]) -> bool:
    return bool(offset) and offset[1] > offset[0]


def _has_skippable_bridge(text: str, left: tuple[int, int], right: tuple[int, int], label: str) -> bool:
    bridge = text[int(left[1]) : int(right[0])]
    if bridge == "":
        return True
    return label in WHITESPACE_BRIDGE_LABELS and bridge.isspace()


def _has_left_extension_bridge(text: str, left: tuple[int, int], right: tuple[int, int]) -> bool:
    bridge = text[int(left[1]) : int(right[0])]
    return bridge == ""


def _nonspace_length(text: str, start: int, end: int) -> int:
    return sum(0 if ch.isspace() else 1 for ch in text[int(start) : int(end)])


def decode_span_logits(
    text: str,
    offsets: list[tuple[int, int]],
    start_scores: np.ndarray,
    end_scores: np.ndarray,
    label_names: list[str],
    default_threshold: float,
    label_thresholds: dict[str, float] | None = None,
    label_max_span_tokens: dict[str, int] | None = None,
) -> list[dict]:
    thresholds = {label: float(default_threshold) for label in label_names}
    if label_thresholds:
        thresholds.update({normalize_entity_name(key): float(value) for key, value in label_thresholds.items()})
    max_tokens = dict(DEFAULT_LABEL_MAX_SPAN_TOKENS)
    if label_max_span_tokens:
        max_tokens.update({normalize_entity_name(key): int(value) for key, value in label_max_span_tokens.items()})

    spans: list[dict] = []
    for label_index, label in enumerate(label_names):
        threshold = thresholds.get(label, float(default_threshold))
        max_span = max_tokens.get(label, 8)
        start_candidates = [idx for idx in range(len(offsets)) if _valid_offset(offsets[idx]) and float(start_scores[idx, label_index]) >= threshold]
        for start_idx in start_candidates:
            best = None
            for end_idx in range(start_idx, min(len(offsets), start_idx + max_span)):
                if not _valid_offset(offsets[end_idx]):
                    continue
                end_score = float(end_scores[end_idx, label_index])
                if end_score < threshold:
                    continue
                score = min(float(start_scores[start_idx, label_index]), end_score)
                if best is None or score > best["score"]:
                    best = {
                        "label": label,
                        "start": int(offsets[start_idx][0]),
                        "end": int(offsets[end_idx][1]),
                        "score": score,
                    }
            if best is not None and best["start"] < best["end"]:
                best["text"] = text[best["start"]:best["end"]]
                spans.append(best)
    return dedupe_spans(spans)


def decode_token_presence_segments(
    text: str,
    offsets: list[tuple[int, int]],
    token_scores: np.ndarray,
    label_names: list[str],
    default_threshold: float,
    label_thresholds: dict[str, float] | None = None,
    label_extend_thresholds: dict[str, float] | None = None,
    label_max_span_tokens: dict[str, int] | None = None,
    label_min_nonspace_chars: dict[str, int] | None = None,
    boundary_label_thresholds: dict[str, float] | None = None,
    start_scores: np.ndarray | None = None,
    end_scores: np.ndarray | None = None,
) -> list[dict]:
    thresholds = {label: float(default_threshold) for label in label_names}
    if label_thresholds:
        thresholds.update({normalize_entity_name(key): float(value) for key, value in label_thresholds.items()})
    extend_thresholds = {label: max(0.0, min(1.0, thresholds[label] * 0.6)) for label in label_names}
    if label_extend_thresholds:
        extend_thresholds.update({normalize_entity_name(key): float(value) for key, value in label_extend_thresholds.items()})
    max_tokens = dict(DEFAULT_LABEL_MAX_SPAN_TOKENS)
    if label_max_span_tokens:
        max_tokens.update({normalize_entity_name(key): int(value) for key, value in label_max_span_tokens.items()})
    min_nonspace_chars = dict(DEFAULT_LABEL_MIN_NONSPACE_CHARS)
    if label_min_nonspace_chars:
        min_nonspace_chars.update({normalize_entity_name(key): int(value) for key, value in label_min_nonspace_chars.items()})
    boundary_thresholds = {label: 0.0 for label in label_names}
    if boundary_label_thresholds:
        boundary_thresholds.update({normalize_entity_name(key): float(value) for key, value in boundary_label_thresholds.items()})

    spans: list[dict] = []
    valid = [_valid_offset(offset) for offset in offsets]
    num_tokens = len(offsets)
    for label_index, label in enumerate(label_names):
        threshold = thresholds.get(label, float(default_threshold))
        extend_threshold = min(threshold, extend_thresholds.get(label, threshold))
        max_span = max_tokens.get(label, 8)
        idx = 0
        while idx < num_tokens:
            if not valid[idx] or float(token_scores[idx, label_index]) < threshold:
                idx += 1
                continue
            start_idx = idx
            end_idx = idx
            while end_idx + 1 < num_tokens and valid[end_idx + 1] and float(token_scores[end_idx + 1, label_index]) >= threshold and (end_idx + 1 - start_idx + 1) <= max_span:
                end_idx += 1
            while (
                start_idx - 1 >= 0
                and valid[start_idx - 1]
                and _has_left_extension_bridge(text, offsets[start_idx - 1], offsets[start_idx])
                and float(token_scores[start_idx - 1, label_index]) >= extend_threshold
                and (end_idx - (start_idx - 1) + 1) <= max_span
            ):
                start_idx -= 1
            while (
                end_idx + 1 < num_tokens
                and valid[end_idx + 1]
                and _has_skippable_bridge(text, offsets[end_idx], offsets[end_idx + 1], label)
                and float(token_scores[end_idx + 1, label_index]) >= extend_threshold
                and ((end_idx + 1) - start_idx + 1) <= max_span
            ):
                end_idx += 1
            presence_slice = token_scores[start_idx : end_idx + 1, label_index]
            score = float(presence_slice.mean())
            out_start_idx = start_idx
            out_end_idx = end_idx
            if start_scores is not None and end_scores is not None:
                refine_window = min(3, end_idx - start_idx + 1)
                start_window = start_scores[start_idx : start_idx + refine_window, label_index]
                best_start_rel = int(np.argmax(start_window))
                best_start_idx = start_idx + best_start_rel
                end_window_start = max(best_start_idx, end_idx - refine_window + 1)
                end_window = end_scores[end_window_start : end_idx + 1, label_index]
                best_end_rel = int(np.argmax(end_window))
                best_end_idx = end_window_start + best_end_rel
                if (
                    float(start_scores[best_start_idx, label_index]) < boundary_thresholds.get(label, 0.0)
                    or float(end_scores[best_end_idx, label_index]) < boundary_thresholds.get(label, 0.0)
                ):
                    idx = end_idx + 1
                    continue
                out_start_idx = best_start_idx
                out_end_idx = best_end_idx
                if label in CONSERVATIVE_BOUNDARY_REFINEMENT_LABELS and (
                    best_start_idx != start_idx or best_end_idx != end_idx
                ):
                    outer_boundary = min(float(start_scores[start_idx, label_index]), float(end_scores[end_idx, label_index]))
                    refined_boundary = min(
                        float(start_scores[best_start_idx, label_index]),
                        float(end_scores[best_end_idx, label_index]),
                    )
                    if refined_boundary < outer_boundary + 0.08:
                        out_start_idx = start_idx
                        out_end_idx = end_idx
                score = (
                    0.65 * score
                    + 0.175 * float(start_scores[out_start_idx, label_index])
                    + 0.175 * float(end_scores[out_end_idx, label_index])
                )
            min_chars = int(min_nonspace_chars.get(label, 1))
            if _nonspace_length(text, offsets[out_start_idx][0], offsets[out_end_idx][1]) < min_chars:
                idx = end_idx + 1
                continue
            spans.append(
                {
                    "label": label,
                    "start": int(offsets[out_start_idx][0]),
                    "end": int(offsets[out_end_idx][1]),
                    "score": score,
                    "text": text[int(offsets[out_start_idx][0]) : int(offsets[out_end_idx][1])],
                }
            )
            idx = end_idx + 1
    return dedupe_spans(spans)


def load_onnx_session(model_ref: str, onnx_file: str = "model_quantized.onnx", onnx_subfolder: str = "onnx"):
    import onnxruntime as ort

    model_path = Path(model_ref)
    if model_path.exists():
        candidates = []
        if onnx_subfolder:
            candidates.append(model_path / onnx_subfolder / onnx_file)
        candidates.append(model_path / onnx_file)
        onnx_path = next((path for path in candidates if path.exists()), candidates[0])
        config = AutoConfig.from_pretrained(model_ref)
        tokenizer = safe_auto_tokenizer(model_ref)
    else:
        remote_name = f"{onnx_subfolder}/{onnx_file}" if onnx_subfolder else onnx_file
        onnx_path = Path(hf_hub_download(repo_id=model_ref, filename=remote_name, repo_type="model"))
        config = AutoConfig.from_pretrained(model_ref)
        tokenizer = safe_auto_tokenizer(model_ref)
    session = ort.InferenceSession(str(onnx_path), providers=["CPUExecutionProvider"])
    return session, tokenizer, config


def run_onnx(session, encoded: dict[str, Any]) -> tuple[np.ndarray, np.ndarray]:
    feed = {}
    input_names = {item.name for item in session.get_inputs()}
    for key, value in encoded.items():
        if key == "offset_mapping":
            continue
        if key in input_names:
            feed[key] = value
    outputs = session.run(None, feed)
    return outputs[0], outputs[1]


def run_onnx_all(session, encoded: dict[str, Any]) -> list[np.ndarray]:
    feed = {}
    input_names = {item.name for item in session.get_inputs()}
    for key, value in encoded.items():
        if key == "offset_mapping":
            continue
        if key in input_names:
            feed[key] = value
    return session.run(None, feed)