File size: 9,950 Bytes
9302284 | 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 | from __future__ import annotations
import json
import re
from dataclasses import dataclass
from pathlib import Path
from typing import Any, Dict, List, Sequence
import numpy as np
from ..config import load_config
from ..data.dataset import load_and_prepare_dataset
from ..logging_utils import get_logger
from ..models.baseline_detoxify import DetoxifyPredictor
from ..models.hf_model import HFPredictor
from ..training.metrics import evaluate_multilabel
from ..utils import resolve_paths, sha256_text
logger = get_logger(__name__)
_RE_URL = re.compile(r"https?://\S+|www\.\S+", flags=re.IGNORECASE)
_RE_EMAIL = re.compile(r"[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Za-z]{2,}", flags=re.IGNORECASE)
_RE_REPEAT = re.compile(r"(.)\1{2,}")
_PUNCT = set("!?.:,;\"'()[]{}<>/\\|@#$%^&*_+-=~`")
@dataclass
class ErrorAnalysisReport:
label_fields: List[str]
n_samples: int
threshold: float
model: Dict[str, Any]
overall_metrics: Dict[str, Any]
confusion_per_label: Dict[str, Any]
feature_summary: Dict[str, Any]
top_error_cases: List[Dict[str, Any]]
def to_dict(self) -> Dict[str, Any]:
return {
"label_fields": self.label_fields,
"n_samples": self.n_samples,
"threshold": self.threshold,
"model": self.model,
"overall_metrics": self.overall_metrics,
"confusion_per_label": self.confusion_per_label,
"feature_summary": self.feature_summary,
"top_error_cases": self.top_error_cases,
}
def _cuda_available() -> bool:
try:
import torch
return bool(torch.cuda.is_available())
except Exception:
return False
def _extract_features(texts: Sequence[str]) -> Dict[str, np.ndarray]:
n = len(texts)
length_chars = np.zeros((n,), dtype=np.int32)
length_words = np.zeros((n,), dtype=np.int32)
uppercase_ratio = np.zeros((n,), dtype=np.float32)
punct_ratio = np.zeros((n,), dtype=np.float32)
exclam = np.zeros((n,), dtype=np.int32)
question = np.zeros((n,), dtype=np.int32)
has_url = np.zeros((n,), dtype=np.int8)
has_email = np.zeros((n,), dtype=np.int8)
has_repeat = np.zeros((n,), dtype=np.int8)
non_ascii_ratio = np.zeros((n,), dtype=np.float32)
has_quote = np.zeros((n,), dtype=np.int8)
for i, t in enumerate(texts):
s = t or ""
length_chars[i] = len(s)
words = s.split()
length_words[i] = len(words)
if len(s) > 0:
upper = sum(1 for c in s if c.isalpha() and c.isupper())
alpha = sum(1 for c in s if c.isalpha())
uppercase_ratio[i] = (upper / alpha) if alpha > 0 else 0.0
punct = sum(1 for c in s if c in _PUNCT)
punct_ratio[i] = punct / len(s)
non_ascii = sum(1 for c in s if ord(c) > 127)
non_ascii_ratio[i] = non_ascii / len(s)
exclam[i] = s.count("!")
question[i] = s.count("?")
has_url[i] = 1 if _RE_URL.search(s) else 0
has_email[i] = 1 if _RE_EMAIL.search(s) else 0
has_repeat[i] = 1 if _RE_REPEAT.search(s) else 0
has_quote[i] = 1 if ('"' in s or "'" in s) else 0
return {
"length_chars": length_chars,
"length_words": length_words,
"uppercase_ratio": uppercase_ratio,
"punct_ratio": punct_ratio,
"exclam": exclam,
"question": question,
"has_url": has_url,
"has_email": has_email,
"has_repeat": has_repeat,
"non_ascii_ratio": non_ascii_ratio,
"has_quote": has_quote,
}
def _summarize_features(features: Dict[str, np.ndarray], mask: np.ndarray) -> Dict[str, Any]:
"""Return a compact numeric summary of features for subset mask."""
idx = np.where(mask)[0]
if idx.size == 0:
return {"n": 0}
out: Dict[str, Any] = {"n": int(idx.size)}
for k, arr in features.items():
sub = arr[idx]
# Use robust stats
out[k] = {
"mean": float(np.mean(sub)),
"median": float(np.median(sub)),
"p90": float(np.percentile(sub, 90)),
}
return out
def _confusion_counts(y_true: np.ndarray, y_pred: np.ndarray) -> Dict[str, int]:
# y_true/y_pred are 0/1 arrays for one label
tp = int(((y_true == 1) & (y_pred == 1)).sum())
fp = int(((y_true == 0) & (y_pred == 1)).sum())
fn = int(((y_true == 1) & (y_pred == 0)).sum())
tn = int(((y_true == 0) & (y_pred == 0)).sum())
return {"tp": tp, "fp": fp, "fn": fn, "tn": tn}
def run_error_analysis(
*,
train_config_path: str,
split: str = "test",
threshold: float = 0.5,
max_samples: int | None = None,
model_kind: str = "finetuned",
top_k_cases: int = 50,
) -> Dict[str, Any]:
"""Run privacy-preserving error analysis.
- No raw text is printed or written.
- We only store hashed ids + numeric features for a few top error cases.
model_kind:
- "finetuned" (default) -> models/finetuned/latest
- "detoxify-unbiased" -> detoxify baseline
"""
cfg = load_config(train_config_path)
paths_cfg = cfg.get("paths", {})
paths = resolve_paths(
data_dir_cfg=str(paths_cfg.get("data_dir", "")),
artifacts_dir_cfg=str(paths_cfg.get("artifacts_dir", "")),
)
loaded = load_and_prepare_dataset(cfg)
label_fields = loaded.label_fields
ds = loaded.dataset.get(split)
if ds is None:
raise ValueError(f"Split '{split}' not found. Available: {list(loaded.dataset.keys())}")
if max_samples is not None:
ds = ds.select(range(min(int(max_samples), len(ds))))
texts = ds["text"]
y_true = np.array(ds["labels"], dtype=np.float32)
device = "cuda" if _cuda_available() else "cpu"
probs: np.ndarray
model_info: Dict[str, Any]
if model_kind == "finetuned":
model_dir = paths.models_dir / "finetuned" / "latest"
if not model_dir.exists():
raise FileNotFoundError(f"Fine-tuned model not found: {model_dir}. Run training first.")
predictor = HFPredictor(model_dir=model_dir, device=device, max_length=int(cfg["model"]["max_length"]))
probs = predictor.predict_proba_matrix(texts, label_order=label_fields, batch_size=64)
model_info = {"kind": model_kind, "model_dir": str(model_dir)}
elif model_kind == "detoxify-unbiased":
detox = DetoxifyPredictor(model_type="unbiased", device=device)
label_map = {"identity_hate": "identity_attack"}
probs = detox.predict_proba_matrix(texts, label_order=label_fields, label_map=label_map, batch_size=64)
model_info = {"kind": model_kind, "detoxify_model_type": "unbiased"}
else:
raise ValueError(f"Unknown model_kind: {model_kind}")
y_pred = (probs >= float(threshold)).astype(np.int32)
overall = evaluate_multilabel(y_true, probs, label_fields, threshold=float(threshold)).to_dict()
# Per-label confusion
confusion = {}
for j, lf in enumerate(label_fields):
confusion[lf] = _confusion_counts(y_true[:, j].astype(np.int32), y_pred[:, j].astype(np.int32))
# Feature extraction
feats = _extract_features(texts)
# Overall error mask: any label differs
any_error = np.any(y_pred != y_true.astype(np.int32), axis=1)
# For each label, fp and fn masks
per_label_summaries: Dict[str, Any] = {}
for j, lf in enumerate(label_fields):
yt = y_true[:, j].astype(np.int32)
yp = y_pred[:, j].astype(np.int32)
fp = (yt == 0) & (yp == 1)
fn = (yt == 1) & (yp == 0)
per_label_summaries[lf] = {
"fp_features": _summarize_features(feats, fp),
"fn_features": _summarize_features(feats, fn),
}
feature_summary = {
"all_samples": _summarize_features(feats, np.ones((len(texts),), dtype=bool)),
"any_error": _summarize_features(feats, any_error),
"per_label": per_label_summaries,
"notes": [
"Feature summaries are numeric aggregates (mean/median/p90).",
"No raw text is stored. 'top_error_cases' uses sha256(text) only.",
],
}
# Top error cases: choose by max absolute error (|prob - true|)
# This provides a stable 'most wrong' set without exposing text.
abs_err = np.abs(probs - y_true)
# score per row: max over labels
row_score = abs_err.max(axis=1)
top_idx = np.argsort(-row_score)[: int(top_k_cases)]
top_cases: List[Dict[str, Any]] = []
for i in top_idx:
t = texts[i]
# hashed id only
case = {
"sha256": sha256_text(t),
"length_chars": int(feats["length_chars"][i]),
"length_words": int(feats["length_words"][i]),
"has_url": int(feats["has_url"][i]),
"has_email": int(feats["has_email"][i]),
"has_quote": int(feats["has_quote"][i]),
"uppercase_ratio": float(feats["uppercase_ratio"][i]),
"max_abs_error": float(row_score[i]),
"true_labels": {lf: int(y_true[i, j] > 0.5) for j, lf in enumerate(label_fields)},
"pred_labels": {lf: int(y_pred[i, j]) for j, lf in enumerate(label_fields)},
"pred_probs": {lf: float(probs[i, j]) for j, lf in enumerate(label_fields)},
}
top_cases.append(case)
report = ErrorAnalysisReport(
label_fields=label_fields,
n_samples=int(len(texts)),
threshold=float(threshold),
model=model_info,
overall_metrics=overall,
confusion_per_label=confusion,
feature_summary=feature_summary,
top_error_cases=top_cases,
)
return report.to_dict()
def save_error_analysis(out_path: Path, report: Dict[str, Any]) -> Path:
out_path.parent.mkdir(parents=True, exist_ok=True)
out_path.write_text(json.dumps(report, indent=2, ensure_ascii=False), encoding="utf-8")
logger.info(f"Saved error analysis report to {out_path}")
return out_path
|