File size: 37,946 Bytes
fec5dda |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 |
import pandas as pd
from nltk.translate import AlignedSent
from nltk.translate.ibm1 import IBMModel1
from nltk.lm import MLE
from nltk.lm.preprocessing import padded_everygram_pipeline
from collections import defaultdict, Counter
import math
import os
from tqdm import tqdm
import pickle
import random
import gc
import matplotlib.pyplot as plt
import numpy as np
import contractions
BILINGUAL_DATA_PATH = "bilingual_cleaned_dataset.csv" # Default bilingual dataset path
VIE_DATA_PATH = "vie_cleaned_dataset.csv" # Default Vietnamese dataset path
VISUALIZATION_PATH = "visualizations" # Default visualization output path
BEAM_SIZE = 3
MAX_PHRASE_LENGTH = 7
LM_ORDER = 3
ALPHA = 0.7
BETA = 0.3
BATCH_SIZE = 1000 # For processing data in batches
MIN_PHRASE_COUNT = 3 # Increased threshold to reduce phrase table size
LIMIT_VOCAB = 100000 # Limit vocabulary size to 10 words
MODE_VISUALIZATION = False # Enable visualization
from pyvi import ViTokenizer
from nltk.tokenize import word_tokenize
################################################## 1. Language Model ##################################################
class LanguageModel:
"""Memory-optimized Language Model"""
def __init__(self, order=LM_ORDER, MODE_VISUALIZATION=MODE_VISUALIZATION):
self.order = order
self.lm = None
self.vocab_size = 0
self.MODE_VISUALIZATION = MODE_VISUALIZATION
def preprocess(self, text):
"""Tokenize Vietnamese words"""
# return text.lower().split()
return ViTokenizer.tokenize(text.lower()).split()
def visualize_iterations(self, word_freq, iteration, batch_tokens, output_dir="/kaggle/working/visualizations"):
if "KAGGLE_KERNEL_RUN_TYPE" in os.environ:
# Đang chạy trên Kaggle
output_dir = "/kaggle/working/visualizations"
else:
output_dir = VISUALIZATION_PATH
os.makedirs(output_dir, exist_ok=True)
"""Visualize word frequency for a given iteration"""
if not self.MODE_VISUALIZATION:
return
print(f"\nIteration {iteration} - Word Frequency (Top 5):")
top_words = word_freq.most_common(5)
for word, count in top_words:
print(f" {word}: {count}")
if not os.path.exists(output_dir):
os.makedirs(output_dir)
words, counts = zip(*word_freq.most_common(10)) if word_freq else ([], [])
if words:
plt.figure(figsize=(8, 6))
plt.bar(words, counts, color='purple', alpha=0.7)
plt.title(f'Word Frequency - Iteration {iteration}')
plt.xlabel('Words')
plt.ylabel('Frequency')
plt.xticks(rotation=45)
plt.grid(True, axis='y')
plt.savefig(os.path.join(output_dir, f'word_freq_iter_{iteration}.png'))
plt.close()
def get_probability(self, tokens):
"""Calculate probability P(V) for a vietnamese tokens sequence"""
if not tokens or not self.lm:
return 0.0
start_tokens = ['<s>'] * (self.order - 1)
tokens = start_tokens + tokens
log_prob = 0.0
for i in range(self.order - 1, len(tokens)):
context = tokens[max(0, i - self.order + 1):i]
word = tokens[i]
prob = self.lm.score(word, context) or 1e-10
log_prob += math.log(prob)
return log_prob
def visualize_log_probabilities(self, sentences, max_sentences=100, output_dir="/kaggle/working/visualizations"):
if "KAGGLE_KERNEL_RUN_TYPE" in os.environ:
# Đang chạy trên Kaggle
output_dir = "/kaggle/working/visualizations"
else:
# Chạy local
output_dir = VISUALIZATION_PATH
os.makedirs(output_dir, exist_ok=True)
"""Visualize the log probabilities of a sample of sentences"""
if not self.MODE_VISUALIZATION:
return
if not self.lm:
print("Cannot visualize log probabilities: Language model not trained.")
return
# Sample sentences to reduce computation
sample_size = min(len(sentences), max_sentences)
sample_sentences = random.sample(sentences, sample_size) if len(sentences) > max_sentences else sentences
# Compute log probabilities
log_probs = []
for sent in sample_sentences:
tokens = self.preprocess(sent)
log_prob = self.get_probability(tokens)
log_probs.append(log_prob)
# Print summary statistics
print(f"\nLog Probabilities for {len(log_probs)} sentences:")
print(f" Mean Log Probability: {np.mean(log_probs):.2f}")
print(f" Min Log Probability: {min(log_probs):.2f}")
print(f" Max Log Probability: {max(log_probs):.2f}")
# Plot histogram of log probabilities
if not os.path.exists(output_dir):
os.makedirs(output_dir)
plt.figure(figsize=(8, 6))
plt.hist(log_probs, bins=30, color='blue', alpha=0.7)
plt.title('Distribution of Log Probabilities for Sentences')
plt.xlabel('Log Probability')
plt.ylabel('Frequency')
plt.grid(True)
plt.savefig(os.path.join(output_dir, 'log_probabilities.png'))
plt.close()
print(f"Log probabilities visualization saved to {output_dir}/log_probabilities.png")
def train(self, vietnamese_sentences, max_sentences=200000):
"""Training Language Model with memory optimization"""
print(f"Training Language Model on {min(len(vietnamese_sentences), max_sentences)} sentences...")
# Limit training data for LM to reduce memory
if len(vietnamese_sentences) > max_sentences:
print(f"Sampling {max_sentences} sentences from {len(vietnamese_sentences)} for LM training")
vietnamese_sentences = random.sample(vietnamese_sentences, max_sentences)
# Process in batches to reduce memory usage
all_tokens = []
batch_size = 10000
word_freq = Counter()
iteration = 0
for i in range(0, len(vietnamese_sentences), batch_size):
batch = vietnamese_sentences[i:i+batch_size]
batch_tokens = [self.preprocess(sent) for sent in batch]
all_tokens.extend(batch_tokens)
# Update word frequency for visualization
if self.MODE_VISUALIZATION and iteration < 2: # Limit to 2 iterations
for tokens in batch_tokens:
word_freq.update(tokens)
self.visualize_iterations(word_freq, iteration + 1, batch_tokens)
iteration += 1
# Force garbage collection
if i % (batch_size * 5) == 0:
gc.collect()
vocab = set()
for tokens in all_tokens:
vocab.update(tokens)
# Limit vocabulary size to most frequent words
if len(vocab) > LIMIT_VOCAB:
word_freq = Counter()
for tokens in all_tokens:
word_freq.update(tokens)
# Keep only top words
most_common = word_freq.most_common(LIMIT_VOCAB)
vocab = set(word for word, _ in most_common)
print(f"Limited vocabulary to {len(vocab)} most frequent words")
self.vocab_size = len(vocab)
# Filter sentences to contain only vocabulary words
filtered_sentences = []
for tokens in all_tokens:
filtered_tokens = [token for token in tokens if token in vocab]
if filtered_tokens: # Only add non-empty sentences
filtered_sentences.append(filtered_tokens)
# Clear original data
del all_tokens
gc.collect()
# Train N-gram model
train_data, padded_sents = padded_everygram_pipeline(self.order, filtered_sentences)
self.lm = MLE(self.order)
self.lm.fit(train_data, padded_sents)
# Visualize log probabilities after training
if self.MODE_VISUALIZATION:
self.visualize_log_probabilities(vietnamese_sentences)
# Clear training data
del filtered_sentences, train_data, padded_sents
gc.collect()
return {"vocab_size": self.vocab_size, "ngram_order": self.order}
############################################# 2. Translation Model #############################################
class TranslationModel:
"""Memory-optimized Translation Model"""
def __init__(self, max_phrase_length=MAX_PHRASE_LENGTH, MODE_VISUALIZATION=MODE_VISUALIZATION):
self.max_phrase_length = max_phrase_length
self.phrase_table = {}
self.word_alignments = []
self.MODE_VISUALIZATION = MODE_VISUALIZATION
def preprocess(self, text, lang):
"""Preprocess text for both languages"""
text = text.lower()
if lang == 'eng':
text = contractions.fix(text)
return word_tokenize(text)
elif lang == 'vie':
return ViTokenizer.tokenize(text).split()
else:
return text.split()
def load_bilingual_data_batch(self, file_path, batch_size=BATCH_SIZE):
"""Load bilingual data in batches to reduce memory usage"""
print(f"Loading bilingual data from {file_path} in batches")
# default = '/kaggle/input/general-data/bilingual_cleaned_dataset.csv'
try:
df = pd.read_csv(file_path)
except FileNotFoundError:
file_path = os.path.join('datatest', BILINGUAL_DATA_PATH)
df = pd.read_csv(file_path)
total_rows = len(df)
print(f"Total rows: {total_rows}")
for start_idx in range(0, total_rows, batch_size):
end_idx = min(start_idx + batch_size, total_rows)
batch_df = df.iloc[start_idx:end_idx]
aligned_sentences = []
for _, row in batch_df.iterrows():
eng_tokens = self.preprocess(row['en'], 'eng')
vie_tokens = self.preprocess(row['vi'], 'vie')
# Filter out very long sentences to save memory
if len(eng_tokens) <= 50 and len(vie_tokens) <= 50:
aligned_sentences.append(AlignedSent(eng_tokens, vie_tokens))
yield aligned_sentences
# Clean up batch
del batch_df, aligned_sentences
gc.collect()
def visualize_alignments(self, aligned_sentences, max_sentences=2, output_dir="/kaggle/working/visualizations"):
if "KAGGLE_KERNEL_RUN_TYPE" in os.environ:
# Đang chạy trên Kaggle
output_dir = "/kaggle/working/visualizations"
else:
# Chạy local
output_dir = VISUALIZATION_PATH
os.makedirs(output_dir, exist_ok=True)
"""Visualize word alignments for a sample of sentence pairs"""
if not self.MODE_VISUALIZATION:
return
if not self.ibm_model:
print("Cannot visualize alignments: IBM Model 1 not trained.")
return
# Sample sentences to reduce computation
sample_size = min(len(aligned_sentences), max_sentences)
sample_sentences = random.sample(aligned_sentences, sample_size) if len(aligned_sentences) > max_sentences else aligned_sentences
if not os.path.exists(output_dir):
os.makedirs(output_dir)
for idx, sent in enumerate(sample_sentences):
src_words = sent.words # English
tgt_words = sent.mots # Vietnamese
alignment = sent.alignment
# Create alignment matrix
matrix = np.zeros((len(tgt_words), len(src_words)))
for src_idx, tgt_idx in alignment:
if tgt_idx is not None and src_idx < len(src_words) and tgt_idx < len(tgt_words):
matrix[tgt_idx, src_idx] = 1
# Plot alignment matrix
plt.figure(figsize=(8, 6))
plt.imshow(matrix, cmap='Blues', interpolation='nearest')
plt.title(f'Alignment Matrix - Sentence Pair {idx + 1}')
plt.xlabel('English Words')
plt.ylabel('Vietnamese Words')
plt.xticks(range(len(src_words)), src_words, rotation=45, ha='right')
plt.yticks(range(len(tgt_words)), tgt_words)
plt.tight_layout()
plt.savefig(os.path.join(output_dir, f'alignment_matrix_{idx + 1}.png'))
plt.close()
# Print alignment details
print(f"\nSentence Pair {idx + 1}:")
print(f" English: {' '.join(src_words)}")
print(f" Vietnamese: {' '.join(tgt_words)}")
print(f" Alignments: {[(src_words[src], tgt_words[tgt]) for src, tgt in alignment if tgt is not None]}")
print(f"Alignment visualizations saved to {output_dir}/")
def _extract_alignments_memory_efficient(self, aligned_sentences, ibm_model):
"""Memory-efficient alignment extraction"""
alignments = []
# Process in smaller batches
batch_size = 5000
for i in range(0, len(aligned_sentences), batch_size):
batch_alignments = []
batch_sentences = aligned_sentences[i:i+batch_size]
for sent_pair in batch_sentences:
eng_tokens = sent_pair.words
vie_tokens = sent_pair.mots
# Only keep high-probability alignments
alignment = []
for eng_i, eng_word in enumerate(eng_tokens):
best_prob = 0
best_vie_i = -1
for vie_i, vie_word in enumerate(vie_tokens):
prob = ibm_model.translation_table.get(eng_word, {}).get(vie_word, 0)
if prob > best_prob:
best_prob = prob
best_vie_i = vie_i
# Only keep alignments above threshold
if best_prob > 0.01: # Increased threshold
alignment.append((eng_i, best_vie_i))
batch_alignments.append(alignment)
alignments.extend(batch_alignments)
# Periodic cleanup
if i % (batch_size * 10) == 0:
gc.collect()
return alignments
def extract_phrases_memory_efficient(self, aligned_sentences):
"""Memory-efficient phrase extraction"""
print("Extracting phrase pairs with memory optimization...")
# Use smaller data structures
phrase_counts = defaultdict(lambda: defaultdict(int))
# Process in batches
batch_size = 5000
for i in range(0, len(aligned_sentences), batch_size):
batch_sentences = aligned_sentences[i:i+batch_size]
batch_alignments = self.word_alignments[i:i+batch_size]
for sent_pair, alignments in zip(batch_sentences, batch_alignments):
if not alignments: # Skip sentences with no alignments
continue
eng_tokens = sent_pair.words
vie_tokens = sent_pair.mots
alignment_set = set(alignments)
# Extract word-level translations first
for eng_i, vie_i in alignments:
if eng_i < len(eng_tokens) and vie_i < len(vie_tokens):
eng_word = eng_tokens[eng_i]
vie_word = vie_tokens[vie_i]
phrase_counts[eng_word][vie_word] += 1
# Extract short phrases only (max length 3 to save memory)
max_len = min(3, self.max_phrase_length)
consistent_phrases = self._extract_consistent_phrases(
eng_tokens, vie_tokens, alignment_set, max_len
)
for eng_phrase, vie_phrase in consistent_phrases:
phrase_counts[eng_phrase][vie_phrase] += 1
# Periodic cleanup
if i % (batch_size * 5) == 0:
gc.collect()
print(f"Processed {i+batch_size} sentences...")
# Calculate probabilities with higher threshold
self.phrase_table = {}
for eng_phrase, vie_phrases in phrase_counts.items():
total_count = sum(vie_phrases.values())
if total_count >= MIN_PHRASE_COUNT: # Higher threshold
# Keep only top 3 translations per phrase to save memory
sorted_phrases = sorted(vie_phrases.items(), key=lambda x: x[1], reverse=True)[:3]
filtered_phrases = {}
for vie_phrase, count in sorted_phrases:
if count >= MIN_PHRASE_COUNT:
filtered_phrases[vie_phrase] = count / total_count
if filtered_phrases:
self.phrase_table[eng_phrase] = filtered_phrases
print(f"Extracted {len(self.phrase_table)} phrase pairs (filtered)")
# Visualize phrase table if enabled
if self.MODE_VISUALIZATION:
self.visualize_phrase_table()
return self.phrase_table
def _extract_consistent_phrases(self, eng_tokens, vie_tokens, alignments, max_length):
"""Extract consistent phrase pairs with length limit"""
consistent_phrases = []
eng_len = len(eng_tokens)
# Limit phrase extraction to reduce memory
for e_start in range(eng_len):
for e_end in range(e_start, min(eng_len, e_start + max_length)):
vie_positions = set()
for e_pos in range(e_start, e_end + 1):
for (eng_idx, vie_idx) in alignments:
if eng_idx == e_pos:
vie_positions.add(vie_idx)
if not vie_positions:
continue
v_start, v_end = min(vie_positions), max(vie_positions)
if v_end - v_start + 1 <= max_length:
if self._is_consistent_phrase_pair(e_start, e_end, v_start, v_end, alignments):
eng_phrase = ' '.join(eng_tokens[e_start:e_end+1])
vie_phrase = ' '.join(vie_tokens[v_start:v_end+1])
consistent_phrases.append((eng_phrase, vie_phrase))
return consistent_phrases
def _is_consistent_phrase_pair(self, e_start, e_end, v_start, v_end, alignments):
"""Check if a phrase pair is consistent"""
for (eng_idx, vie_idx) in alignments:
if (e_start <= eng_idx <= e_end) and not (v_start <= vie_idx <= v_end):
return False
if (v_start <= vie_idx <= v_end) and not (e_start <= eng_idx <= e_end):
return False
return True
def train_ibm_model_incremental(self, file_path="/kaggle/input/general-data/bilingual_cleaned_dataset.csv", iterations=5):
"""Train IBM Model 1 incrementally to reduce memory usage"""
if not os.path.exists(file_path):
file_path = os.path.join('datatest', BILINGUAL_DATA_PATH)
print(f"Training IBM Model 1 incrementally with {iterations} iterations...")
# First pass: collect vocabulary and create aligned sentences
all_aligned_sentences = []
eng_vocab = set()
vie_vocab = set()
for batch in self.load_bilingual_data_batch(file_path):
for sent_pair in batch:
eng_vocab.update(sent_pair.words)
vie_vocab.update(sent_pair.mots)
all_aligned_sentences.append(sent_pair)
# Limit total sentences to prevent memory issues
if len(all_aligned_sentences) >= 300000: # Reduced from 500k
print(f"Limited training to {len(all_aligned_sentences)} sentences")
break
print(f"Training on {len(all_aligned_sentences)} aligned sentences")
print(f"English vocab: {len(eng_vocab)}, Vietnamese vocab: {len(vie_vocab)}")
ibm_model = IBMModel1(all_aligned_sentences, iterations)
# Extract alignments with memory optimization
self.word_alignments = self._extract_alignments_memory_efficient(all_aligned_sentences, ibm_model)
# Clean up
del ibm_model
gc.collect()
return all_aligned_sentences
def visualize_phrase_table(self, max_phrases=10, output_dir="/kaggle/working/visualizations"):
if "KAGGLE_KERNEL_RUN_TYPE" in os.environ:
# Đang chạy trên Kaggle
output_dir = "/kaggle/working/visualizations"
else:
# Chạy local
output_dir = VISUALIZATION_PATH
os.makedirs(output_dir, exist_ok=True)
"""Visualize the phrase table as a heatmap with English phrases as columns and Vietnamese phrases as rows"""
if not self.MODE_VISUALIZATION:
return
if not self.phrase_table:
print("Cannot visualize phrase table: Phrase table is empty.")
return
# Select top English phrases and their top Vietnamese translations
eng_phrases = sorted(self.phrase_table.keys(), key=lambda x: sum(self.phrase_table[x].values()), reverse=True)[:max_phrases]
vie_phrases = set()
for eng in eng_phrases:
vie_phrases.update(self.phrase_table[eng].keys())
vie_phrases = sorted(list(vie_phrases))[:max_phrases] # Limit Vietnamese phrases
# Create matrix for probabilities
matrix = np.zeros((len(vie_phrases), len(eng_phrases)))
for i, vie in enumerate(vie_phrases):
for j, eng in enumerate(eng_phrases):
matrix[i, j] = self.phrase_table.get(eng, {}).get(vie, 0)
# Create heatmap
if not os.path.exists(output_dir):
os.makedirs(output_dir)
plt.figure(figsize=(12, 8))
plt.imshow(matrix, cmap='Blues', interpolation='nearest')
plt.title('Phrase Table Translation Probabilities')
plt.xlabel('English Phrases')
plt.ylabel('Vietnamese Phrases')
plt.xticks(range(len(eng_phrases)), eng_phrases, rotation=45, ha='right')
plt.yticks(range(len(vie_phrases)), vie_phrases)
plt.colorbar(label='Translation Probability')
plt.tight_layout()
plt.savefig(os.path.join(output_dir, 'phrase_table.png'))
plt.close()
# Print sample phrase pairs
print("\nSample Phrase Table Entries (Top 5 English phrases):")
for eng in eng_phrases[:5]:
print(f" English: {eng}")
for vie, prob in sorted(self.phrase_table[eng].items(), key=lambda x: x[1], reverse=True)[:3]:
print(f" -> Vietnamese: {vie}, Probability: {prob:.4f}")
print(f"Phrase table visualization saved to {output_dir}/phrase_table.png")
############################################# 3. Decoder Algorithm #############################################
class Decoder:
"""Memory-optimized decoder"""
def __init__(self, phrase_table, language_model, beam_size=BEAM_SIZE):
self.phrase_table = phrase_table
self.lm = language_model
self.beam_size = beam_size
def translate(self, sentence):
"""Translate sentence with memory optimization"""
tokens = sentence.lower().split()
if not tokens:
return ""
return self._greedy_translate(tokens)
def _greedy_translate(self, tokens):
"""Greedy translation to save memory"""
translation = []
i = 0
while i < len(tokens):
best_phrase_len = 1
best_translation = tokens[i] # fallback
# Try phrases of different lengths
for phrase_len in range(min(3, len(tokens) - i), 0, -1): # Max length 3
eng_phrase = ' '.join(tokens[i:i+phrase_len])
if eng_phrase in self.phrase_table:
# Get best translation
vie_translations = self.phrase_table[eng_phrase]
if vie_translations:
best_vie_phrase = max(vie_translations.items(), key=lambda x: x[1])
best_translation = best_vie_phrase[0]
best_phrase_len = phrase_len
break
translation.append(best_translation)
i += best_phrase_len
return ' '.join(translation)
class Hypothesis:
"""Lightweight hypothesis class"""
def __init__(self, translation, coverage, score, last_phrase_end):
self.translation = translation
self.coverage = coverage
self.score = score
self.last_phrase_end = last_phrase_end
################################################# 4. Combine all SMT System #############################################
class SMT:
"""Memory-optimized SMT system"""
def __init__(self):
self.lm = LanguageModel(order=LM_ORDER)
self.tm = TranslationModel(max_phrase_length=MAX_PHRASE_LENGTH)
self.decoder = None
def post_process(self, text):
"""Replaces underscores with spaces in the translated text."""
return text.replace("_", " ")
def train(self):
bilingual_path = "/kaggle/input/general-data/bilingual_cleaned_dataset.csv"
vie_path = "/kaggle/input/general-data/vie_cleaned_dataset.csv"
if not os.path.exists(bilingual_path):
bilingual_path = os.path.join("datatest", BILINGUAL_DATA_PATH)
vie_path = os.path.join("datatest", VIE_DATA_PATH)
print("=== Training Translation Model ===")
aligned_sentences = self.tm.train_ibm_model_incremental(bilingual_path)
phrase_table = self.tm.extract_phrases_memory_efficient(aligned_sentences)
del aligned_sentences
gc.collect()
# Train language model
print("\n=== Training Language Model ===")
vie_df = pd.read_csv(vie_path)
vietnamese_sentences = vie_df['vi'].tolist()
del vie_df # Free memory
gc.collect()
lm_stats = self.lm.train(vietnamese_sentences, max_sentences=50000) # Limit LM training data
del vietnamese_sentences # Free memory
gc.collect()
# Initialize decoder
self.decoder = Decoder(phrase_table, self.lm)
# Save model immediately
self.save_model()
return {
"phrase_pairs": len(phrase_table),
"lm_stats": lm_stats
}
def translate_sentence(self, sentence):
"""Translate a single sentence"""
if self.decoder is None:
raise ValueError("Model not trained or loaded.")
translated_text_with_underscores = self.decoder.translate(sentence)
return self.post_process(translated_text_with_underscores)
def save_model(self):
"""Save the trained model"""
if "KAGGLE_KERNEL_RUN_TYPE" in os.environ:
# Đang chạy trên Kaggle
model_dir = "/kaggle/working/checkpoints"
else:
# Chạy local
model_dir = "checkpoints"
os.makedirs(model_dir, exist_ok=True)
# Save with compression
with open(os.path.join(model_dir, "phrase_table.pkl"), 'wb') as f:
pickle.dump(self.tm.phrase_table, f, protocol=pickle.HIGHEST_PROTOCOL)
with open(os.path.join(model_dir, "lm_object.pkl"), 'wb') as f:
pickle.dump(self.lm, f, protocol=pickle.HIGHEST_PROTOCOL)
print(f"Model saved to {model_dir}")
def load_model(self, model_dir='checkpoints'):
"""Load a pre-trained model"""
with open(os.path.join(model_dir, "phrase_table.pkl"), 'rb') as f:
phrase_table = pickle.load(f)
with open(os.path.join(model_dir, "lm_object.pkl"), 'rb') as f:
self.lm = pickle.load(f)
self.decoder = Decoder(phrase_table, self.lm, BEAM_SIZE)
self.tm.phrase_table = phrase_table
print(f"Model loaded from {model_dir}")
def evaluate(self, test_file='/kaggle/input/general-data/test_cleaned_dataset.csv', sample_size=5):
"""Evaluate model on test set"""
try :
df = pd.read_csv(test_file)
except FileNotFoundError:
test_file = 'datatest/test_cleaned_dataset.csv'
df = pd.read_csv(test_file)
sample_size = min(sample_size, len(df))
sample_indices = random.sample(range(len(df)), sample_size)
results = []
for idx in sample_indices:
try:
source = df.iloc[idx]['en']
reference = df.iloc[idx]['vi']
translation = self.translate_sentence(source)
results.append({
"source": source,
"reference": reference,
"translation": translation
})
except Exception as e:
print(f"Error translating sentence {idx}: {e}")
results.append({
"source": df.iloc[idx]['en'],
"reference": df.iloc[idx]['vi'],
"translation": "Translation failed"
})
return results
def save_predictions_batch(self, test_file="/kaggle/input/general-data/test_cleaned_dataset.csv", output_file="/kaggle/working/predicted.csv", batch_size=1000):
"""Save predictions in batches to avoid memory issues"""
# Check if test_file exists, if not update to default path
if not os.path.exists(test_file):
test_file = "datatest/test_cleaned_dataset.csv"
output_file = "datatest/predicted1.csv"
print(f"Output file will be saved to: {output_file}")
df_info = pd.read_csv(test_file, nrows=0) # Just get column info
total_rows = len(pd.read_csv(test_file))
print(f"Processing {total_rows} sentences in batches of {batch_size}")
# Process in batches and write incrementally
first_batch = True
for start_idx in tqdm(range(0, total_rows, batch_size), desc="Processing batches"):
end_idx = min(start_idx + batch_size, total_rows)
# Read batch
batch_df = pd.read_csv(test_file, skiprows=range(1, start_idx+1), nrows=batch_size)
# Process batch
batch_predictions = []
for _, row in batch_df.iterrows():
try:
source = row['en']
reference = row['vi']
translation = self.translate_sentence(source)
batch_predictions.append({
"en": source,
"vi": reference,
"pre": translation
})
except Exception as e:
batch_predictions.append({
"en": row['en'],
"vi": row['vi'],
"pre": "Translation failed"
})
# Save batch
batch_pred_df = pd.DataFrame(batch_predictions)
if first_batch:
batch_pred_df.to_csv(output_file, index=False)
first_batch = False
else:
batch_pred_df.to_csv(output_file, mode='a', header=False, index=False)
# Clean up
del batch_df, batch_predictions, batch_pred_df
gc.collect()
print(f"Predictions saved to {output_file}")
return output_file
def main():
print("Starting Memory-Optimized SMT System...")
smt = SMT()
model_dir = "checkpoints"
if os.path.exists(model_dir) and os.path.isfile(os.path.join(model_dir, "phrase_table.pkl")):
print("Loading existing model...")
smt.load_model()
else:
print("Training new model...")
stats = smt.train()
print(f"Training complete: {stats}")
# Evaluate model
print("\nEvaluating model...")
results = smt.evaluate(sample_size=1)
print("\nExample translations:")
for i, result in enumerate(results):
print(f"\nExample {i+1}:")
print(f"English: {result['source']}")
print(f"Reference: {result['reference']}")
print(f"Translation: {result['translation']}")
# Save predictions in batches
print("\nSaving predictions in batches...")
output_file = smt.save_predictions_batch(batch_size=500) # Smaller batch size
print(f"All predictions saved to: {output_file}")
# Final memory cleanup
gc.collect()
print("Processing complete!")
class SMTExtended(SMT):
def infer(self, sentence):
"""Translate a single arbitrary English sentence into Vietnamese using beam search"""
if self.decoder is None:
raise ValueError("Model not trained or loaded.")
# Preprocess input sentence
tokens = self.tm.preprocess(sentence, 'eng')
if not tokens:
return ""
# Initialize beam: (score, translation_tokens, last_pos, covered_positions)
beam = [(0.0, [], 0, set())] # Score, translation tokens, last position, covered positions
best_score = float('-inf')
best_translation = []
# Beam search
while beam:
new_beam = []
for score, trans_tokens, last_pos, covered in beam:
# Check if all positions are covered
if len(covered) == len(tokens):
if score > best_score:
best_score = score
best_translation = trans_tokens
continue
# Find next uncovered position
next_pos = last_pos
while next_pos in covered and next_pos < len(tokens):
next_pos += 1
if next_pos >= len(tokens):
if score > best_score:
best_score = score
best_translation = trans_tokens
continue
# Try phrases starting at next_pos
for phrase_len in range(1, min(self.tm.max_phrase_length + 1, len(tokens) - next_pos + 1)):
eng_phrase = ' '.join(tokens[next_pos:next_pos + phrase_len])
# Get possible translations from phrase table
vie_translations = self.tm.phrase_table.get(eng_phrase, {})
if not vie_translations and phrase_len == 1:
# Fallback for single unknown word
vie_translations = {tokens[next_pos]: 1.0}
for vie_phrase, trans_prob in vie_translations.items():
# Split Vietnamese phrase into tokens for LM scoring
vie_tokens = vie_phrase.split()
# Calculate new score: combine translation prob and LM prob
log_trans_prob = math.log(trans_prob) if trans_prob > 0 else math.log(1e-10)
lm_score = self.lm.get_probability(trans_tokens + vie_tokens)
new_score = ALPHA * log_trans_prob + BETA * lm_score
# Update covered positions
new_covered = covered | set(range(next_pos, next_pos + phrase_len))
# Add to new beam
new_beam.append((score + new_score, trans_tokens + vie_tokens, next_pos + phrase_len, new_covered))
# Keep top BEAM_SIZE hypotheses
new_beam.sort(key=lambda x: x[0], reverse=True)
beam = new_beam[:self.decoder.beam_size]
# Return best translation
return ' '.join(best_translation) if best_translation else "Translation failed"
if __name__ == "__main__":
main()
|