File size: 7,722 Bytes
456ae8a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import json
from pathlib import Path

import torch
import torch.nn as nn
import torch.nn.functional as F
from safetensors.torch import load_file
from torchcrf import CRF

from model import ModernBertConfig, ModernBertModel


CWS_ID2TAG = {0: "B", 1: "I", 2: "E", 3: "S"}
LTP_POS_TAGS = [
    "a", "b", "c", "d", "e", "h", "i", "j", "k", "m",
    "n", "nd", "nh", "ni", "nl", "ns", "nt", "nz", "o", "p",
    "q", "r", "u", "v", "wp", "x", "z",
]
LTP_ID2POS = {i: t for i, t in enumerate(LTP_POS_TAGS)}
LTP_NER_TAGS = [
    "O",
    "B-Nh", "I-Nh", "E-Nh", "S-Nh",
    "B-Ns", "I-Ns", "E-Ns", "S-Ns",
    "B-Ni", "I-Ni", "E-Ni", "S-Ni",
]
LTP_ID2NER = {i: t for i, t in enumerate(LTP_NER_TAGS)}


def bies_to_words(chars, tag_ids):
    words = []
    buf = []
    for ch, tid in zip(chars, tag_ids):
        tag = CWS_ID2TAG.get(int(tid), "S")
        if tag == "S":
            if buf:
                words.append("".join(buf))
                buf = []
            words.append(ch)
        elif tag == "B":
            if buf:
                words.append("".join(buf))
            buf = [ch]
        elif tag == "I":
            buf.append(ch)
        elif tag == "E":
            buf.append(ch)
            words.append("".join(buf))
            buf = []
        else:
            words.append(ch)
    if buf:
        words.append("".join(buf))
    return words


def bies_tags_to_spans(tag_ids):
    spans = []
    i = 0
    n = len(tag_ids)
    while i < n:
        tag = LTP_ID2NER.get(int(tag_ids[i]), "O")
        if tag.startswith("S-"):
            spans.append({"type": tag[2:], "start": i, "end": i + 1})
            i += 1
            continue
        if tag.startswith("B-"):
            ent = tag[2:]
            j = i + 1
            while j < n:
                nxt = LTP_ID2NER.get(int(tag_ids[j]), "O")
                if nxt == f"I-{ent}":
                    j += 1
                    continue
                if nxt == f"E-{ent}":
                    spans.append({"type": ent, "start": i, "end": j + 1})
                j += 1
                break
            i = j
            continue
        i += 1
    return spans


class PieceCharTokenizer:
    def __init__(self, model_dir):
        import piece_tokenizer as pt

        model_dir = Path(model_dir)
        self._tok = pt.Tokenizer()
        self._tok.load(str(model_dir / "piece.model"), cn_dict="no")
        self.pad_token_id = self._tok.piece_to_id("<pad>")
        self.unk_token_id = 0
        mask_path = model_dir / "mask_token_id.txt"
        self.mask_token_id = int(mask_path.read_text().strip()) if mask_path.exists() else self._tok.vocab_size()
        self.vocab_size = self._tok.vocab_size() + 1
        self._cache = {}

    def char_to_id(self, char):
        if char in self._cache:
            return self._cache[char]
        ids = self._tok.encode_as_ids(char)
        tid = ids[0] if ids else self.unk_token_id
        self._cache[char] = tid
        return tid


class BERTcForMT(nn.Module):
    def __init__(self, config, num_pos=27, num_cws=4, num_ner=13, dropout=0.1):
        super().__init__()
        self.config = config
        self.bert = ModernBertModel(config)
        hidden = config.hidden_size
        self.dropout = nn.Dropout(dropout)
        self.cws_classifier = nn.Linear(hidden, num_cws)
        self.cws_crf = CRF(num_cws, batch_first=True)
        self.pos_classifier = nn.Linear(hidden, num_pos)
        self.ner_classifier = nn.Linear(hidden, num_ner)
        self.ner_crf = CRF(num_ner, batch_first=True)

    @classmethod
    def from_pretrained(cls, model_dir, map_location="cpu"):
        model_dir = Path(model_dir)
        cfg = ModernBertConfig(**json.loads((model_dir / "config.json").read_text()))
        model = cls(cfg)
        state = load_file(str(model_dir / "model.safetensors"), device=str(map_location))
        missing, unexpected = model.load_state_dict(state, strict=True)
        if missing or unexpected:
            raise RuntimeError(f"Bad state dict: missing={missing}, unexpected={unexpected}")
        model.eval()
        return model

    def forward(self, input_ids, attention_mask, cws_labels=None, pos_labels=None, ner_labels=None):
        hs = self.bert(input_ids, attention_mask=attention_mask)
        hs = self.dropout(hs)
        cws_emi = self.cws_classifier(hs)
        pos_logits = self.pos_classifier(hs)
        ner_emi = self.ner_classifier(hs)
        mask = attention_mask.bool()
        losses = {}
        if cws_labels is not None:
            losses["cws"] = -self.cws_crf(cws_emi, cws_labels, mask=mask, reduction="mean")
        if pos_labels is not None:
            losses["pos"] = F.cross_entropy(
                pos_logits.view(-1, pos_logits.size(-1)),
                pos_labels.view(-1),
                ignore_index=-100,
            )
        if ner_labels is not None:
            losses["ner"] = -self.ner_crf(ner_emi, ner_labels, mask=mask, reduction="mean")
        return losses, (cws_emi, pos_logits, ner_emi)

    @torch.no_grad()
    def decode_cws(self, input_ids, attention_mask):
        hs = self.bert(input_ids, attention_mask=attention_mask)
        emi = self.cws_classifier(self.dropout(hs)).float()
        return self.cws_crf.decode(emi, mask=attention_mask.bool())

    @torch.no_grad()
    def decode_ner(self, input_ids, attention_mask):
        hs = self.bert(input_ids, attention_mask=attention_mask)
        emi = self.ner_classifier(self.dropout(hs)).float()
        return self.ner_crf.decode(emi, mask=attention_mask.bool())

    @torch.no_grad()
    def predict_pos(self, input_ids, attention_mask):
        hs = self.bert(input_ids, attention_mask=attention_mask)
        logits = self.pos_classifier(self.dropout(hs))
        return logits.argmax(-1)

    @torch.no_grad()
    def predict(self, text, tokenizer=None, max_len=254, device=None):
        if tokenizer is None:
            tokenizer = PieceCharTokenizer(Path(__file__).resolve().parent)
        if isinstance(text, str):
            single = True
            texts = [text]
        else:
            single = False
            texts = list(text)
        device = device or next(self.parameters()).device
        self.eval()
        lengths = [min(len(t), max_len) for t in texts]
        max_l = max(lengths) if lengths else 0
        input_ids = torch.full((len(texts), max_l), tokenizer.pad_token_id, dtype=torch.long, device=device)
        attn = torch.zeros((len(texts), max_l), dtype=torch.long, device=device)
        for i, s in enumerate(texts):
            ids = [tokenizer.char_to_id(c) for c in s[:lengths[i]]]
            if ids:
                input_ids[i, :len(ids)] = torch.tensor(ids, dtype=torch.long, device=device)
                attn[i, :len(ids)] = 1
        cws_preds = self.decode_cws(input_ids, attn)
        pos_pred = self.predict_pos(input_ids, attn).cpu().tolist()
        ner_preds = self.decode_ner(input_ids, attn)

        out = []
        for i, s in enumerate(texts):
            chars = list(s[:lengths[i]])
            cws_ids = cws_preds[i][:len(chars)]
            pos_ids = pos_pred[i][:len(chars)]
            ner_ids = ner_preds[i][:len(chars)]
            words = bies_to_words(chars, cws_ids)
            pos = []
            offset = 0
            for w in words:
                if offset < len(pos_ids):
                    pos.append(LTP_ID2POS.get(int(pos_ids[offset]), "x"))
                else:
                    pos.append("x")
                offset += len(w)
            out.append({
                "text": s,
                "words": words,
                "pos": pos,
                "ner": bies_tags_to_spans(ner_ids),
            })
        return out[0] if single else out