pet-scanner / test_framework.py
einfachstarten's picture
Upload test_framework.py with huggingface_hub
5a5841b verified
"""
Pet-ID Test Framework
Automatisiertes Testing mit Metriken, Threshold-Sweep und Per-Model-Analyse.
"""
from __future__ import annotations
import json
import logging
import shutil
import tempfile
import zipfile
from dataclasses import dataclass, field, asdict
from datetime import datetime
from pathlib import Path
from typing import Callable, Optional
from logic import PetIdentifier, SIMILARITY_THRESHOLD, ENSEMBLE_WEIGHTS
logger = logging.getLogger("pet_id")
TEST_DATA_DIR = Path(__file__).parent / "test_data"
IMAGE_EXTENSIONS = {".jpg", ".jpeg", ".png", ".bmp", ".webp"}
# --- Dataclasses ---
@dataclass
class ImageResult:
path: str
pet_name: str # ground-truth Tiername (aus Ordnername)
matched_name: Optional[str] # erkannter Name oder None
matched_id: Optional[str]
ensemble_score: float
mega_score: float
dino_score: float
classification: str # "TP", "FN", "FP"
@dataclass
class PetTestResult:
name: str
total: int
tp: int
fn: int
fp: int
scores: list[float] = field(default_factory=list)
@dataclass
class ThresholdPoint:
threshold: float
tp: int
fn: int
fp: int
precision: float
recall: float
f1: float
@dataclass
class ModelAnalysis:
model_key: str # "mega", "dino", "ensemble"
avg_tp_score: float
avg_fn_score: float
min_tp_score: float
max_fn_score: float
@dataclass
class TestRunResult:
timestamp: str
threshold: float
num_registration: int
total_images: int
total_test_images: int
tp: int
fn: int
fp: int
precision: float
recall: float
f1: float
pet_results: list[dict] = field(default_factory=list)
image_results: list[dict] = field(default_factory=list)
threshold_sweep: list[dict] = field(default_factory=list)
model_analysis: list[dict] = field(default_factory=list)
optimal_threshold: float = 0.0
optimal_f1: float = 0.0
# --- Discovery ---
def discover_test_pets() -> dict[str, list[Path]]:
"""Scannt test_data/ nach Tier-Ordnern, extrahiert Zips automatisch.
Returns: {pet_name: [image_paths]}"""
if not TEST_DATA_DIR.exists():
return {}
pets = {}
for pet_dir in sorted(TEST_DATA_DIR.iterdir()):
if not pet_dir.is_dir():
continue
name = pet_dir.name
# Zips automatisch entpacken
for zf in pet_dir.glob("*.zip"):
logger.info("Entpacke %s ...", zf.name)
with zipfile.ZipFile(zf, "r") as z:
z.extractall(pet_dir)
logger.info("Zip entpackt: %d Dateien", len(list(pet_dir.iterdir())) - 1)
# Bilder sammeln (rekursiv, da Zips Unterordner haben koennen)
images = sorted([
p for p in pet_dir.rglob("*")
if p.is_file()
and p.suffix.lower() in IMAGE_EXTENSIONS
and not p.name.startswith(".")
and "__MACOSX" not in str(p)
])
if images:
pets[name] = images
logger.info("Test-Daten: %s -> %d Bilder", name, len(images))
return pets
# --- Test-Kernfunktionen ---
def _test_single_image(
identifier: PetIdentifier,
image_path: Path,
ground_truth_name: str,
registered_names: set[str],
) -> ImageResult:
"""Klassifiziert ein einzelnes Bild als TP/FN/FP."""
pet_id, name, score, details = identifier.identify(str(image_path))
# Score-Details extrahieren
mega_score = 0.0
dino_score = 0.0
if details.get("scores"):
# Besten Match pro Modell finden
for d in details["scores"]:
if d["name"].lower() == ground_truth_name.lower() or (pet_id and d["id"] == pet_id):
mega_score = d["mega_score"]
dino_score = d["dino_score"]
break
# Falls kein spezifischer Match: hoechsten Score nehmen
if mega_score == 0.0 and dino_score == 0.0 and details["scores"]:
best = max(details["scores"], key=lambda x: x["ensemble_score"])
mega_score = best["mega_score"]
dino_score = best["dino_score"]
# Klassifikation
if pet_id and name and name.lower() == ground_truth_name.lower():
classification = "TP"
elif pet_id and name and name.lower() != ground_truth_name.lower():
classification = "FP"
else:
classification = "FN"
return ImageResult(
path=str(image_path),
pet_name=ground_truth_name,
matched_name=name,
matched_id=pet_id,
ensemble_score=score,
mega_score=mega_score,
dino_score=dino_score,
classification=classification,
)
def _threshold_sweep(
image_results: list[ImageResult],
start: float = 0.20,
end: float = 0.80,
step: float = 0.01,
) -> tuple[list[ThresholdPoint], float, float]:
"""Sweep ueber Thresholds, berechnet Metriken bei jedem Punkt.
Returns: (points, optimal_threshold, optimal_f1)"""
points = []
best_f1 = 0.0
best_thresh = start
thresh = start
while thresh <= end + 1e-9:
tp = fn = fp = 0
for r in image_results:
if r.ensemble_score >= thresh:
if r.pet_name.lower() == (r.matched_name or "").lower():
tp += 1
else:
fp += 1
else:
fn += 1
precision = tp / (tp + fp) if (tp + fp) > 0 else 0.0
recall = tp / (tp + fn) if (tp + fn) > 0 else 0.0
f1 = 2 * precision * recall / (precision + recall) if (precision + recall) > 0 else 0.0
points.append(ThresholdPoint(
threshold=round(thresh, 4),
tp=tp, fn=fn, fp=fp,
precision=round(precision, 4),
recall=round(recall, 4),
f1=round(f1, 4),
))
if f1 > best_f1:
best_f1 = f1
best_thresh = round(thresh, 4)
thresh += step
return points, best_thresh, best_f1
def _per_model_analysis(image_results: list[ImageResult]) -> list[ModelAnalysis]:
"""Analysiert die Performance jedes Modells einzeln."""
analyses = []
for key, score_fn in [
("mega", lambda r: r.mega_score),
("dino", lambda r: r.dino_score),
("ensemble", lambda r: r.ensemble_score),
]:
tp_scores = [score_fn(r) for r in image_results if r.classification == "TP"]
fn_scores = [score_fn(r) for r in image_results if r.classification == "FN"]
analyses.append(ModelAnalysis(
model_key=key,
avg_tp_score=round(sum(tp_scores) / len(tp_scores), 4) if tp_scores else 0.0,
avg_fn_score=round(sum(fn_scores) / len(fn_scores), 4) if fn_scores else 0.0,
min_tp_score=round(min(tp_scores), 4) if tp_scores else 0.0,
max_fn_score=round(max(fn_scores), 4) if fn_scores else 0.0,
))
return analyses
def run_test(
source_identifier: PetIdentifier,
num_registration: int = 3,
progress_callback: Optional[Callable[[float, str], None]] = None,
) -> TestRunResult:
"""Hauptfunktion: Fuehrt kompletten Test durch.
- Erstellt temporaere DB (produktive DB wird NICHT angefasst)
- Registriert erste N Bilder pro Tier, testet den Rest
- Berechnet alle Metriken inkl. Threshold-Sweep
"""
# Test-Daten laden
test_pets = discover_test_pets()
if not test_pets:
raise ValueError("Keine Test-Daten gefunden in test_data/")
# Temporaeres Verzeichnis fuer Test-DB
tmp_dir = Path(tempfile.mkdtemp(prefix="petid_test_"))
tmp_db = tmp_dir / "test_db.json"
tmp_storage = tmp_dir / "storage"
def _progress(frac: float, msg: str):
if progress_callback:
progress_callback(frac, msg)
logger.info("TEST [%d%%] %s", int(frac * 100), msg)
try:
# Modelle sicherstellen
_progress(0.0, "Lade Modelle...")
source_identifier.load_models()
# Test-Identifier mit geteilten Modellen erstellen
test_id = PetIdentifier.create_with_shared_models(
source=source_identifier,
db_path=tmp_db,
storage_dir=tmp_storage,
)
# Phase 1: Registrierung
total_images = sum(len(imgs) for imgs in test_pets.values())
registered_names = set()
test_images = [] # (path, ground_truth_name)
for pet_name, images in test_pets.items():
reg_images = images[:num_registration]
test_imgs = images[num_registration:]
if not test_imgs:
logger.warning("Tier '%s' hat nur %d Bilder, ueberspringe (brauche >%d)",
pet_name, len(images), num_registration)
continue
_progress(0.05, f"Registriere {pet_name} ({len(reg_images)} Bilder)...")
reg_paths = [str(p) for p in reg_images]
test_id.register(pet_name, reg_paths)
registered_names.add(pet_name)
for img in test_imgs:
test_images.append((img, pet_name))
if not test_images:
raise ValueError("Keine Test-Bilder uebrig nach Registrierung")
# Phase 2: Test
image_results = []
total_test = len(test_images)
for i, (img_path, ground_truth) in enumerate(test_images):
frac = 0.10 + 0.80 * (i / total_test)
_progress(frac, f"Teste {img_path.name} ({i + 1}/{total_test})")
result = _test_single_image(test_id, img_path, ground_truth, registered_names)
image_results.append(result)
# Phase 3: Metriken
_progress(0.92, "Berechne Metriken...")
tp = sum(1 for r in image_results if r.classification == "TP")
fn = sum(1 for r in image_results if r.classification == "FN")
fp = sum(1 for r in image_results if r.classification == "FP")
precision = tp / (tp + fp) if (tp + fp) > 0 else 0.0
recall = tp / (tp + fn) if (tp + fn) > 0 else 0.0
f1 = 2 * precision * recall / (precision + recall) if (precision + recall) > 0 else 0.0
# Per-Pet Ergebnisse
pet_results = {}
for r in image_results:
if r.pet_name not in pet_results:
pet_results[r.pet_name] = PetTestResult(
name=r.pet_name, total=0, tp=0, fn=0, fp=0, scores=[])
pr = pet_results[r.pet_name]
pr.total += 1
pr.scores.append(r.ensemble_score)
if r.classification == "TP":
pr.tp += 1
elif r.classification == "FN":
pr.fn += 1
elif r.classification == "FP":
pr.fp += 1
# Threshold-Sweep
_progress(0.95, "Threshold-Sweep...")
sweep_points, opt_thresh, opt_f1 = _threshold_sweep(image_results)
# Per-Model-Analyse
model_analysis = _per_model_analysis(image_results)
_progress(0.98, "Erstelle Ergebnis...")
run_result = TestRunResult(
timestamp=datetime.now().isoformat(),
threshold=SIMILARITY_THRESHOLD,
num_registration=num_registration,
total_images=total_images,
total_test_images=total_test,
tp=tp, fn=fn, fp=fp,
precision=round(precision, 4),
recall=round(recall, 4),
f1=round(f1, 4),
pet_results=[asdict(pr) for pr in pet_results.values()],
image_results=[asdict(r) for r in image_results],
threshold_sweep=[asdict(p) for p in sweep_points],
model_analysis=[asdict(a) for a in model_analysis],
optimal_threshold=opt_thresh,
optimal_f1=opt_f1,
)
# Ergebnis speichern
_progress(1.0, "Fertig!")
save_results(run_result)
return run_result
finally:
# Temporaeres Verzeichnis aufraeumen
shutil.rmtree(tmp_dir, ignore_errors=True)
# --- Persistenz ---
def save_results(result: TestRunResult) -> Path:
"""Speichert Testergebnis als JSON."""
TEST_DATA_DIR.mkdir(exist_ok=True)
ts = datetime.now().strftime("%Y%m%d_%H%M%S")
path = TEST_DATA_DIR / f"results_{ts}.json"
with open(path, "w") as f:
json.dump(asdict(result), f, indent=2, ensure_ascii=False)
logger.info("Test-Ergebnisse gespeichert: %s", path)
return path
def load_results(path: Path) -> dict:
"""Laedt Testergebnis aus JSON als dict."""
with open(path, "r") as f:
return json.load(f)
def list_result_files() -> list[str]:
"""Listet alle gespeicherten Ergebnis-Dateien."""
if not TEST_DATA_DIR.exists():
return []
files = sorted(TEST_DATA_DIR.glob("results_*.json"), reverse=True)
return [f.name for f in files]