File size: 6,584 Bytes
ee7f83c | 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 | import sys
import json
import yaml
import torch
import logging
from pathlib import Path
from collections import Counter
sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
from src.models.recurrent.seq2seq import build_model
logging.basicConfig(
level=logging.INFO,
format="%(asctime)s | %(levelname)s | %(message)s",
)
logger = logging.getLogger(__name__)
LANGUAGE_PAIRS = {
"fil_war": "fil_war",
"war_fil": "war_fil",
}
def load_config(path: str = "config/model_config.yml") -> dict:
with open(path, "r") as f:
return yaml.safe_load(f)
# BLEU implementation (corpus-level, no external deps)
def _ngram_counts(tokens: list[str], n: int) -> Counter:
return Counter(tuple(tokens[i : i + n]) for i in range(len(tokens) - n + 1))
def _clipped_precision(
hypotheses: list[list[str]], references: list[list[str]], n: int
) -> tuple[int, int]:
total_clipped, total_count = 0, 0
for hyp, ref in zip(hypotheses, references):
hyp_counts = _ngram_counts(hyp, n)
ref_counts = _ngram_counts(ref, n)
clipped = {ng: min(c, ref_counts[ng]) for ng, c in hyp_counts.items()}
total_clipped += sum(clipped.values())
total_count += sum(hyp_counts.values())
return total_clipped, total_count
def corpus_bleu_n(
hypotheses: list[list[str]], references: list[list[str]], n: int
) -> float:
"""Corpus-level BLEU-N."""
import math
hyp_len = sum(len(h) for h in hypotheses)
ref_len = sum(len(r) for r in references)
if hyp_len == 0:
return 0.0
bp = 1.0 if hyp_len >= ref_len else math.exp(1 - ref_len / hyp_len)
log_avg = 0.0
for i in range(1, n + 1):
clipped, total = _clipped_precision(hypotheses, references, i)
if total == 0 or clipped == 0:
return 0.0
log_avg += math.log(clipped / total)
return bp * math.exp(log_avg / n) * 100
def chrf(hypotheses: list[list[str]], references: list[list[str]], n: int = 6) -> float:
"""
Corpus-level chrF — character n-gram F-score.
Joins tokens back to string then compares character n-grams.
"""
def char_ngrams(text: str, n: int) -> Counter:
return Counter(text[i : i + n] for i in range(len(text) - n + 1))
total_prec, total_rec, count = 0.0, 0.0, 0
for hyp_tokens, ref_tokens in zip(hypotheses, references):
hyp_str = " ".join(hyp_tokens)
ref_str = " ".join(ref_tokens)
hyp_counts = char_ngrams(hyp_str, n)
ref_counts = char_ngrams(ref_str, n)
if not hyp_counts or not ref_counts:
continue
matched = sum(min(hyp_counts[k], ref_counts[k]) for k in hyp_counts)
prec = matched / sum(hyp_counts.values())
rec = matched / sum(ref_counts.values())
total_prec += prec
total_rec += rec
count += 1
if count == 0:
return 0.0
avg_prec = total_prec / count
avg_rec = total_rec / count
if avg_prec + avg_rec == 0:
return 0.0
return 2 * avg_prec * avg_rec / (avg_prec + avg_rec) * 100
# Evaluation
def evaluate_language(lang_code: str, config: dict, device: torch.device) -> dict:
from src.models.recurrent.tokenizer import SPMTokenizer, BOS_IDX, EOS_IDX
artifact_dir = Path("models/recurrent") / LANGUAGE_PAIRS[lang_code]
checkpoint_path = artifact_dir / "best_model.pt"
if not checkpoint_path.exists():
raise FileNotFoundError(
f"No checkpoint at {checkpoint_path}. Run train_seq2seq.py first."
)
spm_path = str(
Path(config["paths"]["processed_data_dir"])
/ LANGUAGE_PAIRS[lang_code]
/ "spm.model"
)
source_tokenizer = SPMTokenizer(spm_path)
target_tokenizer = SPMTokenizer(spm_path)
checkpoint = torch.load(checkpoint_path, map_location=device)
train_config = checkpoint.get("config", config)
model = build_model(
train_config, len(source_tokenizer), len(target_tokenizer), device
)
model.load_state_dict(checkpoint["model_state_dict"])
model.eval()
test_jsonl = (
Path(config["paths"]["processed_data_dir"])
/ LANGUAGE_PAIRS[lang_code]
/ "test.jsonl"
)
if not test_jsonl.exists():
raise FileNotFoundError(f"Test file not found: {test_jsonl}")
hypotheses, references = [], []
with open(test_jsonl, "r", encoding="utf-8") as f:
for line in f:
line = line.strip()
if not line:
continue
record = json.loads(line)
source_ids = source_tokenizer.encode(record["source_text"])
source_tensor = (
torch.tensor(source_ids, dtype=torch.long).unsqueeze(0).to(device)
)
source_lengths = torch.tensor([len(source_ids)], dtype=torch.long)
predicted_ids, _ = model.translate(
source_tensor, source_lengths, BOS_IDX, EOS_IDX
)
hyp_tokens = target_tokenizer.decode(
predicted_ids, skip_special_tokens=True
).split()
ref_tokens = record["target_text"].split()
hypotheses.append(hyp_tokens)
references.append(ref_tokens)
results = {
"language_pair": lang_code,
"num_sentences": len(hypotheses),
"bleu_1": round(corpus_bleu_n(hypotheses, references, 1), 4),
"bleu_2": round(corpus_bleu_n(hypotheses, references, 2), 4),
"bleu_3": round(corpus_bleu_n(hypotheses, references, 3), 4),
"bleu_4": round(corpus_bleu_n(hypotheses, references, 4), 4),
"chrf": round(chrf(hypotheses, references), 4),
}
logger.info(
f"{lang_code} | Sentences: {len(hypotheses)} | "
f"BLEU-1: {results['bleu_1']} | BLEU-2: {results['bleu_2']} | "
f"BLEU-3: {results['bleu_3']} | BLEU-4: {results['bleu_4']} | "
f"chrF: {results['chrf']}"
)
return results
if __name__ == "__main__":
config = load_config()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
all_results = []
for lang_code in LANGUAGE_PAIRS:
try:
result = evaluate_language(lang_code, config, device)
all_results.append(result)
except FileNotFoundError as e:
logger.warning(f"Skipping {lang_code}: {e}")
results_path = Path("results/rnn_evaluation.json")
results_path.parent.mkdir(parents=True, exist_ok=True)
with open(results_path, "w") as f:
json.dump(all_results, f, indent=2)
logger.info(f"Results saved to {results_path}")
|