File size: 20,847 Bytes
51b3b77 | 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 | """
INFERENCE & EVALUATION
Greedy Search, Beam Search, BLEU Score evaluation
"""
import torch
import torch.nn.functional as F
from collections import Counter
import math
from tqdm import tqdm
# ============================================================================
# 1. GREEDY SEARCH DECODING
# ============================================================================
def greedy_decode(model, src, src_vocab, tgt_vocab, device, max_len=100, repetition_penalty=1.5, no_repeat_ngram_size=3):
"""
Greedy Decoding - Chọn token có xác suất cao nhất mỗi bước với Repetition Penalty
Args:
model: Transformer model
src: Source sequence [1, src_len] hoặc list of tokens
src_vocab: Source vocabulary
tgt_vocab: Target vocabulary
device: Device
max_len: Maximum length để generate
repetition_penalty: Penalty factor cho tokens lặp lại (>1.0 để giảm repetition)
no_repeat_ngram_size: Kích thước n-gram để tránh lặp lại (0 = tắt)
Returns:
decoded_tokens: List of decoded token indices
decoded_sentence: Decoded sentence (string)
"""
model.eval()
# Nếu src là list, convert sang tensor
if isinstance(src, list):
src = torch.LongTensor([src]).to(device)
elif src.dim() == 1:
src = src.unsqueeze(0)
with torch.no_grad():
# Encode source
encoder_output, src_mask = model.encode(src)
# Khởi tạo target với <sos> token
tgt_tokens = [tgt_vocab.SOS_IDX]
for step in range(max_len):
# Tạo target tensor
tgt = torch.LongTensor([tgt_tokens]).to(device)
# Decode
output = model.decode(tgt, encoder_output, src_mask)
# Lấy prediction cho token cuối cùng
next_token_logits = output[0, -1, :]
# Áp dụng repetition penalty cho từng token
token_counts = {}
for token_id in tgt_tokens:
if token_id not in [tgt_vocab.SOS_IDX, tgt_vocab.EOS_IDX, tgt_vocab.PAD_IDX]:
token_counts[token_id] = token_counts.get(token_id, 0) + 1
# Áp dụng penalty: giảm logit của các tokens đã xuất hiện
for token_id, count in token_counts.items():
if count > 0 and token_id < len(next_token_logits):
# Penalty tăng theo số lần lặp lại
penalty = repetition_penalty ** count
next_token_logits[token_id] = next_token_logits[token_id] / penalty
# N-gram repetition penalty: tránh lặp cụm từ
if no_repeat_ngram_size > 0 and len(tgt_tokens) >= no_repeat_ngram_size:
# Lấy n-gram cuối cùng
last_ngram = tuple(tgt_tokens[-(no_repeat_ngram_size-1):])
# Kiểm tra xem n-gram này đã xuất hiện trước đó chưa
for i in range(len(tgt_tokens) - no_repeat_ngram_size + 1):
ngram = tuple(tgt_tokens[i:i+no_repeat_ngram_size-1])
if ngram == last_ngram:
# Nếu n-gram đã xuất hiện, giảm logit của token tiếp theo
if i + no_repeat_ngram_size - 1 < len(tgt_tokens):
repeated_token = tgt_tokens[i + no_repeat_ngram_size - 1]
if repeated_token < len(next_token_logits):
next_token_logits[repeated_token] = next_token_logits[repeated_token] / (repetition_penalty ** 2)
# Lấy token có xác suất cao nhất
next_token = next_token_logits.argmax().item()
# Thêm vào sequence
tgt_tokens.append(next_token)
# Dừng nếu gặp <eos>
if next_token == tgt_vocab.EOS_IDX:
break
# Decode thành sentence
decoded_sentence = tgt_vocab.decode(tgt_tokens)
return tgt_tokens, decoded_sentence
# ============================================================================
# 2. BEAM SEARCH DECODING
# ============================================================================
class BeamSearchNode:
"""
Node trong Beam Search
"""
def __init__(self, tokens, log_prob, length):
self.tokens = tokens
self.log_prob = log_prob
self.length = length
def eval(self, alpha=0.6):
"""
Tính score với length normalization
Score = log_prob / (length^alpha)
Args:
alpha: Length penalty factor
"""
return self.log_prob / (self.length ** alpha)
def beam_search_decode(model, src, src_vocab, tgt_vocab, device, beam_size=5, max_len=100, alpha=0.6, repetition_penalty=1.5, no_repeat_ngram_size=3):
"""
Beam Search Decoding - Giữ top-k candidates tốt nhất với Repetition Penalty
Args:
model: Transformer model
src: Source sequence
src_vocab: Source vocabulary
tgt_vocab: Target vocabulary
device: Device
beam_size: Beam size (số lượng candidates)
max_len: Maximum length
alpha: Length penalty factor
repetition_penalty: Penalty factor cho tokens lặp lại (>1.0 để giảm repetition)
no_repeat_ngram_size: Kích thước n-gram để tránh lặp lại (0 = tắt)
Returns:
best_tokens: List of best token indices
best_sentence: Best decoded sentence
"""
model.eval()
# Nếu src là list, convert sang tensor
if isinstance(src, list):
src = torch.LongTensor([src]).to(device)
elif src.dim() == 1:
src = src.unsqueeze(0)
with torch.no_grad():
# Encode source
encoder_output, src_mask = model.encode(src)
# Khởi tạo beam với <sos> token
beams = [BeamSearchNode(
tokens=[tgt_vocab.SOS_IDX],
log_prob=0.0,
length=1
)]
completed_beams = []
for step in range(max_len):
candidates = []
for beam in beams:
# Nếu đã kết thúc, thêm vào completed
if beam.tokens[-1] == tgt_vocab.EOS_IDX:
completed_beams.append(beam)
continue
# Tạo target tensor
tgt = torch.LongTensor([beam.tokens]).to(device)
# Decode
output = model.decode(tgt, encoder_output, src_mask)
# Lấy log probabilities cho token cuối
next_token_logits = output[0, -1, :]
# Áp dụng repetition penalty
# Đếm số lần xuất hiện của mỗi token trong sequence hiện tại
token_counts = {}
for token_id in beam.tokens:
if token_id not in [tgt_vocab.SOS_IDX, tgt_vocab.EOS_IDX, tgt_vocab.PAD_IDX]:
token_counts[token_id] = token_counts.get(token_id, 0) + 1
# Áp dụng penalty: giảm logit của các tokens đã xuất hiện (mạnh hơn)
for token_id, count in token_counts.items():
if count > 0 and token_id < len(next_token_logits):
# Penalty tăng mạnh theo số lần lặp lại
# count=1: penalty nhẹ, count=2+: penalty rất mạnh
penalty = repetition_penalty ** (count * 1.5) # Tăng penalty mạnh hơn
next_token_logits[token_id] = next_token_logits[token_id] / penalty
# N-gram repetition penalty: tránh lặp cụm từ
if no_repeat_ngram_size > 0 and len(beam.tokens) >= no_repeat_ngram_size:
# Lấy n-gram cuối cùng
last_ngram = tuple(beam.tokens[-(no_repeat_ngram_size-1):])
# Kiểm tra xem n-gram này đã xuất hiện trước đó chưa
for i in range(len(beam.tokens) - no_repeat_ngram_size + 1):
ngram = tuple(beam.tokens[i:i+no_repeat_ngram_size-1])
if ngram == last_ngram:
# Nếu n-gram đã xuất hiện, giảm logit của token tiếp theo rất mạnh
if i + no_repeat_ngram_size - 1 < len(beam.tokens):
repeated_token = beam.tokens[i + no_repeat_ngram_size - 1]
if repeated_token < len(next_token_logits):
# Penalty rất mạnh cho n-gram repetition
next_token_logits[repeated_token] = next_token_logits[repeated_token] / (repetition_penalty ** 3)
log_probs = F.log_softmax(next_token_logits, dim=-1)
# Lấy top-k tokens
top_log_probs, top_tokens = torch.topk(log_probs, beam_size)
# Tạo candidates mới
for log_prob, token in zip(top_log_probs, top_tokens):
new_beam = BeamSearchNode(
tokens=beam.tokens + [token.item()],
log_prob=beam.log_prob + log_prob.item(),
length=beam.length + 1
)
candidates.append(new_beam)
# Nếu không còn candidates, dừng
if not candidates:
break
# Chọn top beam_size candidates tốt nhất
beams = sorted(candidates, key=lambda x: x.eval(alpha), reverse=True)[:beam_size]
# Nếu tất cả beams đã complete, dừng
if len(completed_beams) >= beam_size:
break
# Thêm các beams chưa complete vào completed
completed_beams.extend(beams)
# Chọn beam tốt nhất
best_beam = max(completed_beams, key=lambda x: x.eval(alpha))
# Decode thành sentence
best_sentence = tgt_vocab.decode(best_beam.tokens)
return best_beam.tokens, best_sentence
# ============================================================================
# 3. BLEU SCORE CALCULATION
# ============================================================================
def calculate_ngrams(tokens, n):
"""
Tính n-grams từ list of tokens
Args:
tokens: List of tokens
n: n-gram size
Returns:
ngrams: Counter of n-grams
"""
ngrams = []
for i in range(len(tokens) - n + 1):
ngram = tuple(tokens[i:i+n])
ngrams.append(ngram)
return Counter(ngrams)
def calculate_bleu_score(references, hypotheses, max_n=4, weights=None):
"""
Tính BLEU score
BLEU = BP * exp(sum(w_n * log(p_n)))
Args:
references: List of reference sentences (list of token lists)
hypotheses: List of hypothesis sentences (list of token lists)
max_n: Maximum n-gram size (mặc định 4)
weights: Weights cho mỗi n-gram (mặc định uniform)
Returns:
bleu_score: BLEU score (0-100)
"""
if weights is None:
weights = [1.0/max_n] * max_n
# Tính precision cho mỗi n-gram
precisions = []
for n in range(1, max_n + 1):
matched = 0
total = 0
for ref, hyp in zip(references, hypotheses):
# Tính n-grams
ref_ngrams = calculate_ngrams(ref, n)
hyp_ngrams = calculate_ngrams(hyp, n)
# Đếm matched n-grams
for ngram, count in hyp_ngrams.items():
matched += min(count, ref_ngrams.get(ngram, 0))
total += max(len(hyp) - n + 1, 0)
# Tính precision
if total > 0:
precision = matched / total
else:
precision = 0
precisions.append(precision)
# Tính Brevity Penalty (BP)
ref_length = sum(len(ref) for ref in references)
hyp_length = sum(len(hyp) for hyp in hypotheses)
if hyp_length > ref_length:
bp = 1.0
elif hyp_length == 0:
bp = 0.0
else:
bp = math.exp(1 - ref_length / hyp_length)
# Tính BLEU score
if min(precisions) > 0:
log_precisions = [w * math.log(p) for w, p in zip(weights, precisions)]
bleu = bp * math.exp(sum(log_precisions))
else:
bleu = 0.0
return bleu * 100 # Convert sang 0-100 scale
# ============================================================================
# 4. EVALUATE ON TEST SET
# ============================================================================
def evaluate_model(model, test_loader, src_vocab, tgt_vocab, device,
use_beam_search=True, beam_size=5, max_len=100):
"""
Đánh giá model trên test set
Args:
model: Transformer model
test_loader: Test DataLoader
src_vocab: Source vocabulary
tgt_vocab: Target vocabulary
device: Device
use_beam_search: Sử dụng beam search hay greedy search
beam_size: Beam size (nếu dùng beam search)
max_len: Maximum decode length
Returns:
bleu_score: BLEU score
translations: List of (source, reference, hypothesis) tuples
"""
model.eval()
references = []
hypotheses = []
translations = []
print(f"\n{'='*70}")
print(f"ĐÁNH GIÁ TRÊN TEST SET")
print(f"{'='*70}")
print(f"Decoding method: {'Beam Search' if use_beam_search else 'Greedy Search'}")
if use_beam_search:
print(f"Beam size: {beam_size}")
print(f"{'='*70}\n")
with torch.no_grad():
for src, tgt, _, _ in tqdm(test_loader, desc='Evaluating'):
src = src.to(device)
batch_size = src.size(0)
for i in range(batch_size):
src_seq = src[i]
tgt_seq = tgt[i]
# Decode
if use_beam_search:
_, hypothesis = beam_search_decode(
model, src_seq, src_vocab, tgt_vocab,
device, beam_size, max_len, alpha=0.6, repetition_penalty=1.5, no_repeat_ngram_size=3
)
else:
_, hypothesis = greedy_decode(
model, src_seq, src_vocab, tgt_vocab,
device, max_len, repetition_penalty=1.5, no_repeat_ngram_size=3
)
# Reference
reference = tgt_vocab.decode(tgt_seq.tolist())
# Source
source = src_vocab.decode(src_seq.tolist())
# Tokenize để tính BLEU
ref_tokens = reference.split()
hyp_tokens = hypothesis.split()
references.append(ref_tokens)
hypotheses.append(hyp_tokens)
translations.append((source, reference, hypothesis))
# Tính BLEU score
bleu = calculate_bleu_score(references, hypotheses)
print(f"\n{'='*70}")
print(f"KẾT QUẢ ĐÁNH GIÁ")
print(f"{'='*70}")
print(f"BLEU Score: {bleu:.2f}")
print(f"{'='*70}\n")
return bleu, translations
# ============================================================================
# 5. PRINT SAMPLE TRANSLATIONS
# ============================================================================
def print_sample_translations(translations, num_samples=10):
"""
In một số ví dụ dịch
Args:
translations: List of (source, reference, hypothesis) tuples
num_samples: Số lượng samples để in
"""
print(f"\n{'='*70}")
print(f"MỘT SỐ VÍ DỤ DỊCH")
print(f"{'='*70}\n")
for i, (src, ref, hyp) in enumerate(translations[:num_samples], 1):
print(f"Ví dụ {i}:")
print(f" Source: {src}")
print(f" Reference: {ref}")
print(f" Hypothesis: {hyp}")
print()
# ============================================================================
# 6. TRANSLATE SINGLE SENTENCE
# ============================================================================
def translate_sentence(model, sentence, src_vocab, tgt_vocab, device,
use_beam_search=True, beam_size=5, max_len=100, src_lang='vi',
repetition_penalty=1.5, no_repeat_ngram_size=3):
"""
Dịch một câu đơn
Args:
model: Transformer model
sentence: Source sentence (string)
src_vocab: Source vocabulary
tgt_vocab: Target vocabulary
device: Device
use_beam_search: Sử dụng beam search hay greedy
beam_size: Beam size
max_len: Maximum length
src_lang: Ngôn ngữ source ('vi' hoặc 'en')
Returns:
translation: Translated sentence
"""
model.eval()
# Tiền xử lý câu
from data_preprocessing import clean_text
sentence = clean_text(sentence, src_lang)
# Encode
tokens = src_vocab.encode(sentence)
src = torch.LongTensor([tokens]).to(device)
# Decode
if use_beam_search:
_, translation = beam_search_decode(
model, src, src_vocab, tgt_vocab,
device, beam_size, max_len, alpha=0.6,
repetition_penalty=repetition_penalty,
no_repeat_ngram_size=no_repeat_ngram_size
)
else:
_, translation = greedy_decode(
model, src, src_vocab, tgt_vocab,
device, max_len,
repetition_penalty=repetition_penalty,
no_repeat_ngram_size=no_repeat_ngram_size
)
return translation
# ============================================================================
# 7. INTERACTIVE TRANSLATION
# ============================================================================
def interactive_translation(model, src_vocab, tgt_vocab, device,
use_beam_search=True, beam_size=5, direction='vi2en'):
"""
Chế độ dịch tương tác
Args:
model: Transformer model
src_vocab: Source vocabulary
tgt_vocab: Target vocabulary
device: Device
use_beam_search: Sử dụng beam search
beam_size: Beam size
direction: 'vi2en' hoặc 'en2vi'
"""
print("\n" + "="*70)
print("CHẾ ĐỘ DỊCH TƯƠNG TÁC")
print("="*70)
if direction == 'vi2en':
print("Nhập câu tiếng Việt để dịch sang tiếng Anh")
else:
print("Nhập câu tiếng Anh để dịch sang tiếng Việt")
print("Gõ 'quit' hoặc 'exit' để thoát")
print("="*70 + "\n")
src_lang = 'en' if direction == 'en2vi' else 'vi'
src_label = "Tiếng Anh" if direction == 'en2vi' else "Tiếng Việt"
tgt_label = "Tiếng Việt" if direction == 'en2vi' else "Tiếng Anh"
while True:
sentence = input(f"{src_label}: ").strip()
if sentence.lower() in ['quit', 'exit', '']:
print("Tạm biệt!")
break
translation = translate_sentence(
model, sentence, src_vocab, tgt_vocab, device,
use_beam_search, beam_size, src_lang=src_lang, max_len=100
)
print(f"{tgt_label}: {translation}\n")
# ============================================================================
# 8. SAVE TRANSLATIONS
# ============================================================================
def save_translations(translations, output_file='translations.txt'):
"""
Lưu translations ra file
Args:
translations: List of (source, reference, hypothesis) tuples
output_file: Output file path
"""
with open(output_file, 'w', encoding='utf-8') as f:
for i, (src, ref, hyp) in enumerate(translations, 1):
f.write(f"Example {i}:\n")
f.write(f"Source: {src}\n")
f.write(f"Reference: {ref}\n")
f.write(f"Hypothesis: {hyp}\n")
f.write("\n")
print(f"✓ Saved translations to {output_file}") |