File size: 8,522 Bytes
5a95e08
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
indus_ngram.py — Standalone module for InduNgramModel
======================================================
This file MUST exist so that pickle can import InduNgramModel
when loading ngram_model.pkl in 07_ensemble.py and 08_electra_train.py.

The pickle fix:
  When you save a class with pickle, Python records the module path.
  If the class was defined in __main__ (i.e. inside 06_ngram_model.py
  when run directly), pickle saves it as __main__.InduNgramModel.
  When another script tries to load it, __main__ refers to THAT script,
  which doesn't have InduNgramModel — hence the AttributeError.

  Solution: define the class in THIS standalone module (indus_ngram.py).
  Both 06_ngram_model.py and 07_ensemble.py import from here.
  Pickle records the path as indus_ngram.InduNgramModel — always findable.

Do not rename or move this file.
"""

import math
import pickle
from pathlib import Path
from collections import Counter, defaultdict


class InduNgramModel:
    """
    Kneser-Ney smoothed N-gram LM for Indus Script.

    RTL mode (default, recommended):
      Sequences reversed before training/scoring.
      RTL bigram entropy (3.18) < LTR (3.72) → supports RTL hypothesis.

    Sign roles in RTL reading direction:
      INITIAL  = reading-start sign (data position [-1])
      TERMINAL = reading-end sign   (data position [0])
      MEDIAL   = appears in middle positions
    """

    def __init__(self, rtl=True):
        self.rtl             = rtl
        self.unigram         = Counter()
        self.bigram          = defaultdict(Counter)
        self.trigram         = defaultdict(Counter)
        self.start_cnt       = Counter()
        self.end_cnt         = Counter()
        self.total_seqs      = 0
        self.total_tokens    = 0
        self.vocab_size      = 0
        self.D               = 0.75
        self.score_mean      = 0.0
        self.score_std       = 1.0
        self.score_min       = -20.0
        self.score_max       = 0.0
        self._pairwise_acc   = 0.0
        self._cont_right     = Counter()
        self._total_bi_types = 0

    def _orient(self, seq):
        cleaned = [t for t in seq if t is not None]
        return list(reversed(cleaned)) if self.rtl else list(cleaned)

    def train(self, sequences):
        mode = "RTL" if self.rtl else "LTR"
        print(f"  Training [{mode}] on {len(sequences):,} sequences...")

        for seq in sequences:
            s = self._orient(seq)
            if not s:
                continue
            self.total_seqs += 1
            self.unigram.update(s)
            self.start_cnt[s[0]]  += 1
            self.end_cnt[s[-1]]   += 1
            for i in range(len(s) - 1):
                self.bigram[s[i]][s[i+1]] += 1
            for i in range(len(s) - 2):
                self.trigram[(s[i], s[i+1])][s[i+2]] += 1

        self.total_tokens    = sum(self.unigram.values())
        self.vocab_size      = len(self.unigram)

        for a, followers in self.bigram.items():
            for b in followers:
                self._cont_right[b] += 1
        self._total_bi_types = sum(self._cont_right.values())

        self._calibrate(sequences)
        print(f"  Vocab      : {self.vocab_size}")
        print(f"  Pairwise   : {self._pairwise_acc*100:.1f}%")
        print(f"  Score range: [{self.score_min:.3f}, {self.score_max:.3f}]")

    def _calibrate(self, sequences):
        import random, statistics
        random.seed(42)
        all_toks = list(self.unigram.keys())

        def corrupt(seq):
            r = random.randint(0, 3)
            c = list(seq)
            if r == 0:
                random.shuffle(c)
            elif r == 1:
                c[0] = random.choice(list(self.end_cnt.keys()))
            elif r == 2:
                c[-1] = random.choice(list(self.start_cnt.keys()))
            else:
                for p in random.sample(range(len(c)), max(1, len(c)//2)):
                    c[p] = random.choice(all_toks)
            return c

        sample = sequences[:500]
        good   = [self._raw_score(s) for s in sample]
        bad    = [self._raw_score(corrupt(s)) for s in sample]
        all_s  = good + bad

        self.score_mean    = statistics.mean(all_s)
        self.score_std     = statistics.stdev(all_s)
        self.score_min     = min(all_s)
        self.score_max     = max(all_s)
        self._pairwise_acc = sum(g > b for g, b in zip(good, bad)) / len(good)

    def _p_uni_kn(self, w):
        return (self._cont_right[w] + 1) / (self._total_bi_types + self.vocab_size)

    def _p_bi_kn(self, w, given):
        gt = sum(self.bigram[given].values())
        if gt == 0:
            return self._p_uni_kn(w)
        cnt   = self.bigram[given].get(w, 0)
        first = max(cnt - self.D, 0) / gt
        lam   = (self.D / gt) * len(self.bigram[given])
        return first + lam * self._p_uni_kn(w)

    def _p_tri_kn(self, w, a, b):
        gt = sum(self.trigram[(a, b)].values())
        if gt == 0:
            return self._p_bi_kn(w, b)
        cnt   = self.trigram[(a, b)].get(w, 0)
        first = max(cnt - self.D, 0) / gt
        lam   = (self.D / gt) * len(self.trigram[(a, b)])
        return first + lam * self._p_bi_kn(w, b)

    def _p_initial(self, w):
        return (self.start_cnt[w] + 0.1) / (self.total_seqs + 0.1 * self.vocab_size)

    def _p_terminal(self, w):
        return (self.end_cnt[w] + 0.1) / (self.total_seqs + 0.1 * self.vocab_size)

    def _raw_score(self, seq):
        if not seq:
            return self.score_min
        s   = self._orient(seq)
        eps = 1e-10
        lp  = math.log(self._p_initial(s[0])  + eps)
        lp += math.log(self._p_terminal(s[-1]) + eps)
        for i, w in enumerate(s):
            if i == 0:
                p = self._p_uni_kn(w)
            elif i == 1:
                p = self._p_bi_kn(w, s[i-1])
            else:
                p = self._p_tri_kn(w, s[i-2], s[i-1])
            lp += math.log(p + eps)
        return lp / (len(s) + 2)

    def validity_score(self, seq):
        raw  = self._raw_score(seq)
        norm = (raw - self.score_min) / (self.score_max - self.score_min + 1e-10)
        return float(max(0.02, min(0.98, norm)))

    def predict_masked(self, seq_with_none, top_k=10):
        masked = [i for i, t in enumerate(seq_with_none) if t is None]
        results = {}
        n = len(seq_with_none)

        for orig_pos in masked:
            ort_pos  = (n - 1 - orig_pos) if self.rtl else orig_pos
            oriented = self._orient(seq_with_none)
            if ort_pos >= len(oriented):
                continue

            prev  = oriented[ort_pos-1] if ort_pos > 0 and oriented[ort_pos-1] is not None else None
            prev2 = oriented[ort_pos-2] if ort_pos > 1 and oriented[ort_pos-2] is not None else None

            cands = []
            for cand in self.unigram:
                if prev2 is not None and prev is not None:
                    p = self._p_tri_kn(cand, prev2, prev)
                elif prev is not None:
                    p = self._p_bi_kn(cand, prev)
                else:
                    p = self._p_uni_kn(cand)

                if ort_pos == 0:
                    p *= max(self._p_initial(cand) * self.vocab_size, 0.01)
                elif ort_pos == n - 1:
                    p *= max(self._p_terminal(cand) * self.vocab_size, 0.01)

                cands.append((cand, p))

            cands.sort(key=lambda x: -x[1])
            total = sum(p for _, p in cands[:top_k * 3]) or 1
            results[orig_pos] = [
                {"id": c, "prob": p / total, "rank": i + 1}
                for i, (c, p) in enumerate(cands[:top_k])
            ]

        return results

    def sign_role(self, sign_id):
        """Positional role in reading direction."""
        init_p = self.start_cnt[sign_id] / (self.total_seqs + 1)
        term_p = self.end_cnt[sign_id]   / (self.total_seqs + 1)
        if init_p > 0.05 and init_p > term_p * 2:
            return "INITIAL"
        elif term_p > 0.05 and term_p > init_p * 2:
            return "TERMINAL"
        elif self.unigram[sign_id] > 5:
            return "MEDIAL"
        return "RARE"

    def save(self, path):
        path = Path(path)
        path.parent.mkdir(parents=True, exist_ok=True)
        with open(path, "wb") as f:
            pickle.dump(self, f, protocol=pickle.HIGHEST_PROTOCOL)
        print(f"  Saved → {path}")

    @staticmethod
    def load(path):
        with open(path, "rb") as f:
            return pickle.load(f)