File size: 14,951 Bytes
1f71502
 
 
 
 
 
 
 
8176e08
be1bc6c
1f71502
 
8176e08
 
 
1f71502
8176e08
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1f71502
307b502
1f71502
 
 
307b502
 
 
 
1f71502
 
 
 
 
8176e08
307b502
 
 
 
 
1f71502
 
 
 
6441bee
 
8176e08
 
 
1f71502
 
 
 
be1bc6c
 
 
 
 
 
1f71502
 
 
 
 
 
 
 
 
 
 
8176e08
1f71502
8176e08
1f71502
 
 
 
 
8176e08
 
1f71502
8176e08
1f71502
 
 
 
 
8176e08
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
be1bc6c
 
 
 
 
 
 
 
 
 
 
 
 
 
8176e08
1f71502
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8176e08
 
 
 
1f71502
8176e08
 
 
1f71502
8176e08
 
 
 
1f71502
 
8176e08
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1f71502
8176e08
 
1f71502
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
307b502
 
1f71502
 
8176e08
 
 
1f71502
 
307b502
8176e08
 
 
307b502
 
 
 
 
 
1f71502
8176e08
1f71502
307b502
1f71502
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8176e08
 
 
 
1f71502
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
Prediction generation module for Talmud language classifier
Generates predictions for all dafim using a trained model
"""

import torch
import requests
import os
import re
import warnings
from train import TalmudClassifierLSTM, TalmudDataset, MAX_LEN

# Preprocessing regex to match Vercel's preprocessing exactly
# Vercel uses: /[\u0591-\u05C7]|[,\-?!:\.״]+|<[^>]+>/g
PREPROCESSING_REGEX = re.compile(r'[\u0591-\u05C7]|[,\-?!:\.״]+|<[^>]+>')

def preprocess_text(text: str) -> tuple[str, dict, dict]:
    """
    Preprocess text by removing nikud, punctuation, and HTML tags.
    Matches Vercel's preprocessing exactly.
    Returns (preprocessed_text, prep_to_orig, orig_to_prep) where:
    - prep_to_orig maps preprocessed position -> original position
    - orig_to_prep maps original position -> preprocessed position (or -1 if removed)
    """
    preprocessed = ''
    prep_to_orig = {}  # Maps preprocessed_pos -> original_pos
    orig_to_prep = {}  # Maps original_pos -> preprocessed_pos (or -1 if removed)
    preprocessed_pos = 0
    i = 0
    
    # Process text character by character, handling HTML tags as units
    while i < len(text):
        # Check for HTML tags (they are removed as units)
        if text[i] == '<':
            # Find the end of the HTML tag
            tag_end = text.find('>', i)
            if tag_end != -1:
                # Mark all characters in the tag as removed
                for orig_pos in range(i, tag_end + 1):
                    orig_to_prep[orig_pos] = -1
                i = tag_end + 1
                continue
        
        char = text[i]
        char_code = ord(char)
        
        # Check if character should be removed:
        # 1. Nikud range: \u0591-\u05C7 (0x0591 to 0x05C7)
        # 2. Punctuation: , - ? ! : . ״
        should_remove = (
            (0x0591 <= char_code <= 0x05C7) or
            char in [',', '-', '?', '!', ':', '.', '״']
        )
        
        if should_remove:
            orig_to_prep[i] = -1  # Mark as removed
        else:
            prep_to_orig[preprocessed_pos] = i
            orig_to_prep[i] = preprocessed_pos
            preprocessed += char
            preprocessed_pos += 1
        i += 1
    
    return preprocessed, prep_to_orig, orig_to_prep

def fetch_daf_texts(vercel_base_url: str, auth_token: str) -> list:
    """
    Fetch all daf texts from Vercel API endpoint.
    Returns list of daf objects with id and text_content.
    
    Args:
        vercel_base_url: Base URL of the Vercel app
        auth_token: Authentication token for Vercel API (TRAINING_CALLBACK_TOKEN)
    """
    url = f"{vercel_base_url}/api/dafim-texts"
    print(f"Fetching daf texts from {url}...")
    
    try:
        # Include authentication token in header
        headers = {
            'x-auth-token': auth_token,
            'Content-Type': 'application/json'
        }
        response = requests.get(url, headers=headers, timeout=60)
        response.raise_for_status()
        data = response.json()
        print(f"Fetched {data.get('count', 0)} dafim")
        return data.get('dafim', [])
    except Exception as e:
        print(f"Error fetching daf texts: {e}")
        if hasattr(e, 'response') and e.response is not None:
            print(f"Response status: {e.response.status_code}")
            print(f"Response text: {e.response.text}")
        raise

def text_to_sequence(text: str, word_to_idx: dict) -> list:
    """Convert text to sequence of word indices"""
    # Validate that required keys exist
    if '<UNK>' not in word_to_idx:
        raise ValueError("Vocabulary must contain '<UNK>' key")
    if '<PAD>' not in word_to_idx:
        raise ValueError("Vocabulary must contain '<PAD>' key")
    
    words = text.split()
    return [word_to_idx.get(word, word_to_idx['<UNK>']) for word in words]

def generate_predictions_for_daf(
    model: torch.nn.Module,
    daf_text: str,
    word_to_idx: dict,
    label_encoder,
    max_len: int = MAX_LEN
) -> list:
    """
    Generate predictions for a single daf text (original text, not preprocessed).
    Returns list of ranges: [{'start': int, 'end': int, 'type': int}, ...]
    Positions are relative to the original text.
    
    Strategy: Sliding window approach - predict on overlapping windows of text
    """
    model.eval()
    
    # Preprocess the text and get character mappings
    preprocessed_text, prep_to_orig, orig_to_prep = preprocess_text(daf_text)
    
    # Split into words and track character positions accurately
    words = preprocessed_text.split()
    
    if len(words) == 0:
        return []
    
    # Build word boundaries in preprocessed text by tracking positions as we iterate
    # This is more reliable than using find() which could match wrong occurrences
    word_boundaries = []
    char_pos = 0
    word_idx = 0
    
    # Iterate through preprocessed text to find word boundaries
    while char_pos < len(preprocessed_text) and word_idx < len(words):
        # Skip leading spaces
        while char_pos < len(preprocessed_text) and preprocessed_text[char_pos] == ' ':
            char_pos += 1
        
        if char_pos >= len(preprocessed_text):
            break
        
        # Find the current word
        word = words[word_idx]
        word_start = char_pos
        
        # Check if the word starts at this position
        if preprocessed_text[char_pos:char_pos + len(word)] == word:
            word_end = char_pos + len(word)
            word_boundaries.append((word_start, word_end))
            char_pos = word_end
            word_idx += 1
        else:
            # Word doesn't match - this shouldn't happen, but handle gracefully
            # Try to find the word starting from current position
            found_pos = preprocessed_text.find(word, char_pos)
            if found_pos != -1:
                word_boundaries.append((found_pos, found_pos + len(word)))
                char_pos = found_pos + len(word)
                word_idx += 1
            else:
                # Couldn't find word - this indicates a mismatch between words and preprocessed_text
                # This can happen if preprocessing changed the text in an unexpected way
                # Log a warning and use a fallback: estimate position based on character count
                warnings.warn(f"Word '{word}' at index {word_idx} not found in preprocessed text. Using estimated position.")
                # Estimate position: assume words are separated by single spaces
                estimated_start = char_pos
                estimated_end = estimated_start + len(word)
                word_boundaries.append((estimated_start, min(estimated_end, len(preprocessed_text))))
                char_pos = estimated_end
                word_idx += 1
    
    # Validate that we found boundaries for all words
    if len(word_boundaries) < len(words):
        warnings.warn(f"Only found boundaries for {len(word_boundaries)} out of {len(words)} words. Some predictions may be inaccurate.")
    
    # Use sliding window approach
    window_size = max_len
    stride = window_size // 2  # 50% overlap
    
    predictions = []
    ranges = []
    
    with torch.no_grad():
        for i in range(0, len(words), stride):
            # Get window of words
            window_words = words[i:i + window_size]
            
            if len(window_words) == 0:
                break
            
            # Convert to sequence
            seq = text_to_sequence(' '.join(window_words), word_to_idx)
            
            # Pad or truncate to max_len
            if len(seq) > max_len:
                seq = seq[:max_len]
            else:
                seq = seq + [word_to_idx['<PAD>']] * (max_len - len(seq))
            
            # Convert to tensor and add batch dimension
            seq_tensor = torch.tensor([seq], dtype=torch.long)
            
            # Get prediction
            output = model(seq_tensor)
            _, predicted = torch.max(output.data, 1)
            predicted_label_idx = predicted.item()
            
            # Calculate character positions in preprocessed text using word boundaries
            # Ensure we don't go out of bounds
            if i >= len(word_boundaries):
                continue
            
            last_word_idx = min(i + len(window_words) - 1, len(word_boundaries) - 1)
            if last_word_idx < i:
                continue
            
            # Start position is the start of the first word in the window
            window_start_prep = word_boundaries[i][0]
            # End position is the end of the last word in the window
            window_end_prep = word_boundaries[last_word_idx][1]
            
            # Only add if we have a valid range
            if window_end_prep > window_start_prep:
                # Map preprocessed text positions to original text positions
                # Find the original start position
                original_start = prep_to_orig.get(window_start_prep)
                if original_start is None:
                    # Find the closest mapped position before or at window_start_prep
                    for prep_pos in sorted(prep_to_orig.keys(), reverse=True):
                        if prep_pos <= window_start_prep:
                            original_start = prep_to_orig[prep_pos]
                            break
                    if original_start is None:
                        continue  # Skip if we can't map start position
                
                # Find the original end position
                # window_end_prep points to the character after the last character in the window
                # We need to map this to the original text
                window_end_prep_clamped = min(window_end_prep, len(preprocessed_text))
                
                # Find the original position corresponding to the end of the window
                # If window_end_prep_clamped is at the end of preprocessed text, use end of original text
                if window_end_prep_clamped >= len(preprocessed_text):
                    original_end = len(daf_text)
                else:
                    # Find the original position for the character at window_end_prep_clamped
                    # (the character right after the window ends)
                    end_char_orig = prep_to_orig.get(window_end_prep_clamped)
                    if end_char_orig is not None:
                        original_end = end_char_orig
                    else:
                        # Character at window_end_prep_clamped was removed, find the next non-removed character
                        # Look for the next preprocessed position >= window_end_prep_clamped
                        next_prep_pos = None
                        for prep_pos in sorted(prep_to_orig.keys()):
                            if prep_pos >= window_end_prep_clamped:
                                next_prep_pos = prep_pos
                                break
                        
                        if next_prep_pos is not None:
                            original_end = prep_to_orig[next_prep_pos]
                        else:
                            # No more characters in preprocessed text, use end of original text
                            original_end = len(daf_text)
                
                # Ensure end is after start and within bounds
                if original_end <= original_start:
                    # Fallback: ensure at least one character
                    original_end = min(original_start + 1, len(daf_text))
                original_end = min(original_end, len(daf_text))
                
                ranges.append({
                    'start': original_start,
                    'end': original_end,
                    'type': int(predicted_label_idx)
                })
    
    # Merge overlapping ranges of the same type
    if len(ranges) == 0:
        return []
    
    # Sort by start position
    ranges.sort(key=lambda x: x['start'])
    
    # Merge consecutive ranges of same type
    merged_ranges = []
    current_range = ranges[0].copy()
    
    for next_range in ranges[1:]:
        # If same type and overlapping or adjacent, merge
        if (next_range['type'] == current_range['type'] and 
            next_range['start'] <= current_range['end']):
            current_range['end'] = max(current_range['end'], next_range['end'])
        else:
            merged_ranges.append(current_range)
            current_range = next_range.copy()
    
    merged_ranges.append(current_range)
    
    return merged_ranges

def generate_all_predictions(
    model: torch.nn.Module,
    word_to_idx: dict,
    label_encoder,
    vercel_base_url: str,
    auth_token: str
) -> list:
    """
    DEPRECATED: This function is no longer used in the training flow.
    It's kept for reference but should not be called.
    
    Generate predictions for all dafim.
    Returns list of prediction objects: [{'daf_id': str, 'ranges': [...]}, ...]
    
    NOTE: This function expects preprocessed text from the API, but generate_predictions_for_daf
    now expects original text. This function needs to be updated if it's ever used again.
    
    Args:
        model: Trained model
        word_to_idx: Word to index mapping
        label_encoder: Label encoder
        vercel_base_url: Base URL of the Vercel app
        auth_token: Authentication token for Vercel API (TRAINING_CALLBACK_TOKEN)
    """
    print("WARNING: generate_all_predictions is deprecated and may not work correctly.")
    print("Fetching daf texts from Vercel...")
    dafim = fetch_daf_texts(vercel_base_url, auth_token)
    
    if len(dafim) == 0:
        print("No dafim found")
        return []
    
    predictions = []
    
    print(f"Generating predictions for {len(dafim)} dafim...")
    
    for idx, daf in enumerate(dafim):
        if (idx + 1) % 100 == 0:
            print(f"Processed {idx + 1}/{len(dafim)} dafim...")
        
        try:
            daf_id = daf['id']
            # NOTE: The API returns preprocessed text, but generate_predictions_for_daf
            # now expects original text. This will cause incorrect character position mapping.
            # This function should fetch original text or be updated to handle preprocessed text.
            text_content = daf['text_content']
            
            ranges = generate_predictions_for_daf(
                model, text_content, word_to_idx, label_encoder
            )
            
            predictions.append({
                'daf_id': daf_id,
                'ranges': ranges
            })
        except Exception as e:
            print(f"Error generating predictions for daf {daf.get('id', 'unknown')}: {e}")
            # Continue with next daf
            continue
    
    print(f"Generated predictions for {len(predictions)} dafim")
    return predictions