File size: 15,912 Bytes
47869c2
 
097d4eb
47869c2
 
 
 
 
 
 
097d4eb
20d6878
47869c2
 
 
 
 
 
 
 
 
20d6878
47869c2
 
 
20d6878
47869c2
20d6878
47869c2
 
6330fc7
20d6878
47869c2
 
 
 
 
 
 
20d6878
097d4eb
47869c2
20d6878
47869c2
 
097d4eb
47869c2
 
20d6878
47869c2
 
20d6878
47869c2
097d4eb
47869c2
20d6878
47869c2
 
 
 
 
20d6878
6330fc7
20d6878
47869c2
 
 
 
 
 
 
 
 
 
20d6878
47869c2
 
 
097d4eb
 
 
 
47869c2
 
 
20d6878
47869c2
20d6878
47869c2
 
20d6878
47869c2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
20d6878
47869c2
097d4eb
 
 
47869c2
 
 
 
 
097d4eb
47869c2
 
097d4eb
20d6878
47869c2
097d4eb
47869c2
 
 
 
 
 
 
 
 
 
097d4eb
47869c2
 
 
 
20d6878
47869c2
 
 
 
097d4eb
 
47869c2
097d4eb
 
47869c2
097d4eb
 
 
47869c2
097d4eb
 
20d6878
47869c2
 
 
20d6878
47869c2
 
 
 
20d6878
097d4eb
 
 
 
20d6878
47869c2
 
 
 
 
 
 
 
20d6878
47869c2
 
 
 
 
 
 
097d4eb
20d6878
47869c2
 
 
 
 
 
20d6878
47869c2
 
 
20d6878
47869c2
 
 
 
 
097d4eb
47869c2
097d4eb
47869c2
 
 
 
 
 
 
 
20d6878
097d4eb
 
 
 
 
 
 
20d6878
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5659365
20d6878
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5659365
20d6878
 
 
097d4eb
5659365
20d6878
47869c2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
097d4eb
 
47869c2
097d4eb
47869c2
 
20d6878
47869c2
 
 
 
 
20d6878
47869c2
 
 
20d6878
47869c2
 
 
 
e9e2747
47869c2
df44061
 
 
20d6878
47869c2
20d6878
097d4eb
 
 
47869c2
 
20d6878
47869c2
 
 
20d6878
47869c2
 
 
 
 
 
 
 
 
 
 
20d6878
47869c2
 
 
 
20d6878
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
#!/usr/bin/env python3
"""
app.py — BeRestoral
"""

import html
import json
import math
import re
from pathlib import Path
from typing import Any, Dict, Optional

import numpy as np
import torch
import torch.nn as nn
import uvicorn
from fastapi import FastAPI, Request
from fastapi.staticfiles import StaticFiles
from fastapi.templating import Jinja2Templates
from pydantic import BaseModel
from transformers import AutoModelForMaskedLM, AutoTokenizer

app = FastAPI(title="BeRestoral")
app.mount("/static", StaticFiles(directory="static"), name="static")
templates = Jinja2Templates(directory="templates")

device = torch.device("cpu")

MODEL_PATH_BPE  = "MaximEremeev/RoFormer-slav"
MODEL_PATH_CHAR = "MaximEremeev/DualEmb-slav"
PROBE_DIR       = Path("probes")

BIN_START  = 1050
BIN_SIZE   = 50
N_BINS     = 9
BINS       = [(BIN_START + i * BIN_SIZE, BIN_START + (i + 1) * BIN_SIZE - 1)
              for i in range(N_BINS)]
BIN_MIDPOINTS = np.array([(lo + hi) / 2 for lo, hi in BINS])
BIN_LABELS    = [f"{lo}{hi}" for lo, hi in BINS]

CATEGORY_LABELS    = ["letters", "records", "religious", "other"]
CATEGORY_LABELS_RU = ["письма", "деловые записи", "религиозные тексты", "другое"]

print("Loading BPE model (RoFormer)...")
tokenizer_bpe = AutoTokenizer.from_pretrained(MODEL_PATH_BPE, trust_remote_code=True)
tokenizer_bpe.add_special_tokens({"additional_special_tokens": ["[GAP]"]})
model_bpe = AutoModelForMaskedLM.from_pretrained(MODEL_PATH_BPE, trust_remote_code=True).to(device)
model_bpe.eval()

print("Loading char model (DualEmbLM)...")
from huggingface_hub import hf_hub_download

model_char = AutoModelForMaskedLM.from_pretrained(
    MODEL_PATH_CHAR, trust_remote_code=True).to(device)
model_char.eval()

_char_vocab_path = hf_hub_download(repo_id=MODEL_PATH_CHAR, filename="char_vocab.json")
_word_vocab_path = hf_hub_download(repo_id=MODEL_PATH_CHAR, filename="word_vocab.json")
char_vocab = json.loads(Path(_char_vocab_path).read_text(encoding="utf-8"))
word_vocab = json.loads(Path(_word_vocab_path).read_text(encoding="utf-8"))
id_to_char = {v: k for k, v in char_vocab.items()}

EMBED_DIM = 512

class LinearProbe(nn.Module):
    def __init__(self, in_dim: int, out_dim: int, dropout: float = 0.1):
        super().__init__()
        self.net = nn.Sequential(
            nn.LayerNorm(in_dim),
            nn.Dropout(dropout),
            nn.Linear(in_dim, out_dim),
        )
    def forward(self, x):
        return self.net(x)

print("Loading probe classifiers...")
probe_category = LinearProbe(EMBED_DIM, len(CATEGORY_LABELS))
probe_date     = LinearProbe(EMBED_DIM, N_BINS)
probe_category.load_state_dict(torch.load(
    PROBE_DIR / "RoFormer_category_masked_probe.pth", map_location=device, weights_only=True))
probe_date.load_state_dict(torch.load(
    PROBE_DIR / "RoFormer_date_masked_probe.pth", map_location=device, weights_only=True))
probe_category.eval()
probe_date.eval()
print("All models loaded.")

SPECIAL_RE = re.compile(r"(\[GAP\]|\[MASK\]|\[PAD\]|\[UNK\]|\[CLS\]|\[SEP\]|[+:·])")

def split_special(text: str) -> list[str]:
    return [p for p in SPECIAL_RE.split(text) if p]

def align_char_to_word(text: str, char_v: dict, word_v: dict, max_len: int = 256):
    c_unk = char_v["[UNK]"]; c_sep = char_v["[SEP]"]; c_cls = char_v["[CLS]"]
    w_unk = word_v.get("[UNK_WORD]", 0)
    input_ids, word_ids = [c_cls], [word_v.get("[CLS]", w_unk)]
    for part in split_special(text.strip()):
        if SPECIAL_RE.fullmatch(part):
            input_ids.append(char_v.get(part, c_unk))
            word_ids.append(word_v.get(part, w_unk))
            continue
        for chunk in re.split(r"(\s+)", part):
            if not chunk: continue
            if chunk.isspace():
                for ch in chunk:
                    input_ids.append(char_v.get(ch, c_unk)); word_ids.append(w_unk)
            else:
                wid = word_v.get(chunk, w_unk)
                for ch in chunk:
                    input_ids.append(char_v.get(ch, c_unk)); word_ids.append(wid)
    input_ids.append(c_sep); word_ids.append(word_v.get("[SEP]", w_unk))
    if len(input_ids) > max_len:
        input_ids, word_ids = input_ids[:max_len], word_ids[:max_len]
        input_ids[-1] = c_sep; word_ids[-1] = word_v.get("[SEP]", w_unk)
    max_char_id = model_char.config.vocab_char_size - 1
    max_word_id = model_char.config.vocab_word_size - 1
    return {
        "input_ids": [x if x <= max_char_id else c_unk for x in input_ids],
        "word_ids":  [x if x <= max_word_id else w_unk for x in word_ids],
    }

def get_roformer_embedding(text: str) -> torch.Tensor:
    """Mean pooling over non-padding tokens from RoFormer encoder.
    text should already contain BPE mask tokens where lacunae are."""
    clean = re.sub(r"\s+", " ", text).strip()
    enc = tokenizer_bpe(clean, return_tensors="pt", truncation=True,
                        max_length=512, return_attention_mask=True)
    enc = {k: v.to(device) for k, v in enc.items()}
    with torch.no_grad():
        out = model_bpe(**enc, output_hidden_states=True)
    hidden = out.hidden_states[-1]
    mask = enc["attention_mask"].unsqueeze(-1).float()
    emb = (hidden * mask).sum(dim=1) / mask.sum(dim=1)
    return emb.squeeze(0)

def classify(text: str) -> dict:
    emb = get_roformer_embedding(text).unsqueeze(0)
    with torch.no_grad():
        cat_logits  = probe_category(emb)[0]
        date_logits = probe_date(emb)[0]
    cat_probs  = torch.softmax(cat_logits,  dim=-1).cpu().numpy().tolist()
    date_probs = torch.softmax(date_logits, dim=-1).cpu().numpy().tolist()
    best_cat   = int(np.argmax(cat_probs))
    pred_year  = float(np.dot(date_probs, BIN_MIDPOINTS))
    return {
        "category":       CATEGORY_LABELS[best_cat],
        "category_ru":    CATEGORY_LABELS_RU[best_cat],
        "category_probs": {CATEGORY_LABELS[i]: round(p, 4) for i, p in enumerate(cat_probs)},
        "pred_year":      round(pred_year),
        "date_probs":     [round(p, 4) for p in date_probs],
        "bin_labels":     BIN_LABELS,
    }

def generate_sequential(text: str, is_char: bool,
                         top_k: int = 5, temperature: float = 1.0) -> dict:
    if is_char:
        encoded = align_char_to_word(text, char_vocab, word_vocab)
        input_ids     = torch.tensor(encoded["input_ids"]).to(device)
        word_ids      = torch.tensor(encoded["word_ids"]).to(device)
        mask_token_id = char_vocab["[MASK]"]
        mask_str      = "[MASK]"
        model         = model_char
    else:
        inputs        = tokenizer_bpe(text, return_tensors="pt").to(device)
        input_ids     = inputs["input_ids"][0]
        word_ids      = None
        mask_token_id = tokenizer_bpe.mask_token_id
        mask_str      = tokenizer_bpe.mask_token
        model         = model_bpe

    original_mask_indices = torch.where(input_ids == mask_token_id)[0].tolist()
    if not original_mask_indices:
        return {"variants": [], "steps": []}

    current_states = [{"input_ids": input_ids.clone(), "log_prob": 0.0,
                        "inserted_tokens": {}}]
    unfilled_masks = original_mask_indices.copy()
    steps = []

    # For char_pos: track how many masks have been filled so far
    # to compute offset correctly
    masks_filled_count = 0
    mask_str_len = len(mask_str)

    with torch.no_grad():
        while unfilled_masks:
            probe_ids = current_states[0]["input_ids"].unsqueeze(0).to(device)
            if is_char:
                outputs = model(input_ids=probe_ids, word_ids=word_ids.unsqueeze(0))
            else:
                outputs = model(input_ids=probe_ids)
            logits = outputs.logits[0]

            best_mask_idx, highest_prob = None, -1.0
            for m_idx in unfilled_masks:
                probs = torch.softmax(logits[m_idx] / max(0.01, temperature), dim=-1)
                p = torch.max(probs).item()
                if p > highest_prob:
                    highest_prob, best_mask_idx = p, m_idx
            unfilled_masks.remove(best_mask_idx)
            steps.append({"pos": best_mask_idx, "confidence": round(highest_prob * 100, 1)})

            batch_ids = torch.stack([s["input_ids"] for s in current_states]).to(device)
            if is_char:
                batch_wids = word_ids.unsqueeze(0).expand(len(current_states), -1).to(device)
                outputs = model(input_ids=batch_ids, word_ids=batch_wids)
            else:
                outputs = model(input_ids=batch_ids)

            mask_logits = outputs.logits[:, best_mask_idx, :]
            probs_k = torch.softmax(mask_logits / max(0.01, temperature), dim=-1)
            top_probs, top_ids = torch.topk(probs_k, top_k, dim=-1)

            new_candidates = []
            for si, state in enumerate(current_states):
                for i in range(top_k):
                    tid  = top_ids[si, i].item()
                    prob = top_probs[si, i].item()
                    new_ids = state["input_ids"].clone()
                    new_ids[best_mask_idx] = tid
                    new_tok = dict(state["inserted_tokens"])
                    new_tok[best_mask_idx] = tid
                    new_candidates.append({
                        "input_ids": new_ids,
                        "log_prob":  state["log_prob"] + math.log(max(prob, 1e-9)),
                        "inserted_tokens": new_tok,
                    })
            current_states = sorted(new_candidates,
                                    key=lambda x: x["log_prob"], reverse=True)[:top_k]

            _best_id = current_states[0]["inserted_tokens"].get(best_mask_idx)
            if is_char:
                _pred = id_to_char.get(_best_id, "")
            else:
                _pred = tokenizer_bpe.decode(
                    [_best_id], clean_up_tokenization_spaces=False
                ).replace("Ġ", "").replace("##", "").strip() if _best_id else ""

            # Build partial sentence using direct token index addressing
            best_state = current_states[0]

            if is_char:
                # input_ids: [CLS, tok1, tok2, ..., SEP] — skip CLS(0) and SEP(-1)
                current_tokens = [id_to_char.get(tid.item(), "")
                                  for tid in input_ids[1:-1]]
                # Fill in predictions
                for op in original_mask_indices:
                    fid = best_state["inserted_tokens"].get(op)
                    idx_in_tokens = op - 1  # offset for [CLS]
                    if fid is not None and 0 <= idx_in_tokens < len(current_tokens):
                        current_tokens[idx_in_tokens] = id_to_char.get(fid, "")
                # Build marked string
                target_idx = best_mask_idx - 1
                parts = []
                for i, tok in enumerate(current_tokens):
                    if tok in ("[MASK]", "[GAP]", "[PAD]", "[UNK]", "[CLS]", "[SEP]"):
                        parts.append("[MASK]" if tok == "[MASK]" else tok)
                    elif i == target_idx:
                        parts.append(f"[[R]]{tok}[[/R]]")
                    else:
                        parts.append(tok)
                _partial_marked = "".join(parts)
            else:
                # BPE: input_ids includes special tokens, decode each
                current_tokens = [tokenizer_bpe.decode([tid.item()],
                                  clean_up_tokenization_spaces=False)
                                  for tid in input_ids]
                for op in original_mask_indices:
                    fid = best_state["inserted_tokens"].get(op)
                    if fid is not None and 0 <= op < len(current_tokens):
                        current_tokens[op] = tokenizer_bpe.decode(
                            [fid], clean_up_tokenization_spaces=False)
                target_idx = best_mask_idx
                parts = []
                for i, tok in enumerate(current_tokens):
                    clean = tok.replace("Ġ", " ").replace("##", "")
                    if tok == tokenizer_bpe.mask_token:
                        parts.append("[MASK]")
                    elif i == target_idx:
                        parts.append(f"[[R]]{clean}[[/R]]")
                    elif tok in (tokenizer_bpe.cls_token, tokenizer_bpe.sep_token,
                                 tokenizer_bpe.pad_token):
                        pass  # skip special tokens
                    else:
                        parts.append(clean)
                _partial_marked = re.sub(r" +", " ", "".join(parts)).strip()

            steps[-1]["token"] = _pred
            steps[-1]["partial_sentence"] = _partial_marked

    variants = []
    escaped_mask = html.escape(mask_str)
    for state in current_states:
        ordered_ids = [state["inserted_tokens"][i] for i in original_mask_indices]
        full_sentence = html.escape(text)
        if is_char:
            inserted = "".join(id_to_char.get(t, "") for t in ordered_ids).strip()
            for tid in ordered_ids:
                ch = id_to_char.get(tid, "")
                tok = "&nbsp;" if ch == " " else html.escape(ch)
                full_sentence = full_sentence.replace(
                    escaped_mask, f'<span class="highlight-restored">{tok}</span>', 1)
        else:
            inserted = tokenizer_bpe.decode(ordered_ids,
                                            clean_up_tokenization_spaces=True).strip()
            for tid in ordered_ids:
                tok = html.escape(tokenizer_bpe.decode([tid])
                                  .replace("Ġ","").replace("##","").replace(" ",""))
                full_sentence = full_sentence.replace(
                    escaped_mask, f'<span class="highlight-restored">{tok}</span>', 1)
        full_sentence = re.sub(r"\s+", " ", full_sentence.strip())
        variants.append({
            "word":          inserted or "...",
            "score":         round(math.exp(state["log_prob"]) * 100, 2),
            "full_sentence": full_sentence,
            "raw_log_prob":  state["log_prob"],
        })
    return {"variants": variants, "steps": steps}

class RestoreRequest(BaseModel):
    text:        str
    mode:        str   = "char"
    top_k:       int   = 5
    temperature: float = 1.0

@app.get("/")
async def read_root(request: Request):
    return templates.TemplateResponse(request=request, name="index.html")

@app.post("/api/restore")
async def restore_text(req: RestoreRequest) -> Dict[str, Any]:
    try:
        is_char = req.mode == "char"
        mask    = "[MASK]" if is_char else tokenizer_bpe.mask_token
        text    = req.text.replace("#", "[GAP]")
        # Lowering everything except special tokens
        parts = SPECIAL_RE.split(text)
        text = "".join(p if SPECIAL_RE.fullmatch(p) else p.lower() for p in parts if p)

        n_gaps = text.count("-") + text.count("[GAP]")

        # Classification — always use BPE mask regardless of mode
        bpe_mask = tokenizer_bpe.mask_token
        masked_for_classify = re.sub(r"-", bpe_mask, text)
        masked_for_classify = re.sub(r" +", " ", masked_for_classify).strip()
        classification = classify(masked_for_classify)

        # Restoration
        query = re.sub(r" +", " ", text.replace("-", mask)).strip()
        _res  = generate_sequential(query, is_char, req.top_k, req.temperature)

        return {
            "status":         "success",
            "results":        [_res["variants"]],
            "steps":          [_res["steps"]],
            "n_gaps":         n_gaps,
            "classification": classification,
        }
    except Exception as e:
        import traceback
        return {"status": "error", "message": str(e),
                "traceback": traceback.format_exc()}

if __name__ == "__main__":
    import os
    port = int(os.environ.get("PORT", 7860))
    uvicorn.run(app, host="0.0.0.0", port=port)