Spaces:
Sleeping
Sleeping
File size: 11,970 Bytes
e327f0d | 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 | """
quick_test.py - Hizli end-to-end test scripti.
Egitilmis modelleri val setinden secilen ornek goruntuler uzerinde calistirir,
JSON ciktisi + annotated overlay PNG uretir.
Tek ihtiyacin olan:
cd services/ml
.venv\\Scripts\\activate (Windows)
python quick_test.py
Ne yapar:
1. Mini-test (yolo11n-seg, 2 epoch) ile damage segmentation
2. EfficientNet-B0 ile severity classification (her bbox crop'una)
3. Kural-tabanli severity (referans icin)
4. quick_test_out/ altina <img>.json + <img>.overlay.jpg yazar
Argument istemiyorsan opsiyonel:
python quick_test.py --num 5 --conf 0.15
"""
from __future__ import annotations
import argparse
import json
import random
from pathlib import Path
import cv2
import numpy as np
import torch
import torch.nn.functional as F
from torchvision import transforms
from torchvision.models import efficientnet_b0
from ultralytics import YOLO
ML_DIR = Path(__file__).resolve().parent
def _latest_snapshot() -> Path | None:
bundles = ML_DIR / "runs" / "bundles"
if not bundles.exists():
return None
cands = sorted(
(p for p in bundles.iterdir()
if p.is_dir() and (p / "_SNAPSHOT_FOR_BUILD").is_dir()
and p.name.startswith("full_")),
key=lambda p: p.stat().st_mtime,
reverse=True,
)
return (cands[0] / "_SNAPSHOT_FOR_BUILD") if cands else None
_SNAP = _latest_snapshot()
# Prefer the production snapshot (full_YYYYMMDD_HHMMSS) when present; fall
# back to the legacy mini-test path so this script keeps working in fresh
# checkouts that have only trained the smoke run.
if _SNAP is not None and (_SNAP / "damage_best.pt").exists():
DEFAULT_DAMAGE_W = _SNAP / "damage_best.pt"
else:
DEFAULT_DAMAGE_W = ML_DIR / "runs" / "segment" / "runs" / "arac-hasar" / "mini-test" / "weights" / "best.pt"
if _SNAP is not None and (_SNAP / "severity_best.pt").exists():
DEFAULT_SEVERITY_W = _SNAP / "severity_best.pt"
else:
DEFAULT_SEVERITY_W = ML_DIR / "runs" / "severity" / "best.pt"
DEFAULT_VAL_DIR = ML_DIR / "data" / "cardd_yolo" / "images" / "val"
DEFAULT_OUT = ML_DIR / "quick_test_out"
CARDD_CLASSES = ["dent", "scratch", "crack", "glass_shatter", "lamp_broken", "tire_flat"]
# severity_classifier.py'den birebir kopya (rule-based referans icin)
PART_IMPORTANCE_DEFAULT = 1.0
DAMAGE_TYPE_WEIGHT = {
"scratch": 0.6, "dent": 1.0, "crack": 1.4,
"glass_shatter": 2.0, "lamp_broken": 1.8, "tire_flat": 1.5,
}
def rule_based_severity(damage_type: str, area_ratio: float) -> dict:
weight = DAMAGE_TYPE_WEIGHT.get(damage_type, 1.0)
score = area_ratio * 100.0 * weight * PART_IMPORTANCE_DEFAULT
if score < 1.0:
level = "hafif"
elif score < 4.0:
level = "orta"
else:
level = "agir"
return {"level": level, "score": round(score, 4), "method": "rule_based"}
class SeverityCNN:
"""EfficientNet-B0 severity classifier (train_severity.py ile uyumlu)."""
def __init__(self, ckpt_path: Path, device: str = "cuda"):
self.device = device if torch.cuda.is_available() else "cpu"
ckpt = torch.load(ckpt_path, map_location=self.device, weights_only=False)
self.classes = ckpt["classes"]
self.tr_names = ckpt["tr_names"]
self.img_size = ckpt.get("img_size", 224)
self.train_val_acc = ckpt.get("val_acc", None)
model = efficientnet_b0(weights=None)
in_features = model.classifier[1].in_features
model.classifier[1] = torch.nn.Linear(in_features, len(self.classes))
model.load_state_dict(ckpt["model_state_dict"])
model.to(self.device).eval()
self.model = model
self.tf = transforms.Compose([
transforms.ToPILImage(),
transforms.Resize((self.img_size, self.img_size)),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
])
def predict_bgr(self, bgr_crop: np.ndarray) -> dict:
if bgr_crop is None or bgr_crop.size == 0:
return {"level": "hafif", "confidence": 0.0, "method": "cnn_empty"}
rgb = cv2.cvtColor(bgr_crop, cv2.COLOR_BGR2RGB)
x = self.tf(rgb).unsqueeze(0).to(self.device)
with torch.no_grad():
logits = self.model(x)
probs = F.softmax(logits, dim=1)[0].cpu().numpy()
idx = int(np.argmax(probs))
return {
"level": self.tr_names[idx],
"raw_class": self.classes[idx],
"confidence": float(probs[idx]),
"method": "cnn",
"all_probs": {
self.tr_names[i]: float(p) for i, p in enumerate(probs)
},
}
def crop_padded(image: np.ndarray, bbox, padding: float = 0.15) -> np.ndarray:
h, w = image.shape[:2]
x1, y1, x2, y2 = bbox
bw, bh = x2 - x1, y2 - y1
px, py = int(bw * padding), int(bh * padding)
x1 = max(0, int(x1) - px)
y1 = max(0, int(y1) - py)
x2 = min(w, int(x2) + px)
y2 = min(h, int(y2) + py)
return image[y1:y2, x1:x2]
def annotate(image: np.ndarray, damages: list) -> np.ndarray:
overlay = image.copy()
palette = {
"hafif": (60, 200, 60),
"orta": (40, 180, 230),
"agir": (40, 40, 230),
}
for d in damages:
x1 = int(d["bbox"]["x1"]); y1 = int(d["bbox"]["y1"])
x2 = int(d["bbox"]["x2"]); y2 = int(d["bbox"]["y2"])
sev = d.get("severity_cnn", {}).get("level", "hafif")
color = palette.get(sev, (200, 200, 200))
cv2.rectangle(overlay, (x1, y1), (x2, y2), color, 2)
label = f"{d['type']} | cnn={sev} | rule={d['severity_rule']['level']}"
(tw, th), _ = cv2.getTextSize(label, cv2.FONT_HERSHEY_SIMPLEX, 0.5, 1)
cv2.rectangle(overlay, (x1, max(0, y1 - th - 6)),
(x1 + tw + 6, y1), color, -1)
cv2.putText(overlay, label, (x1 + 3, max(th, y1 - 4)),
cv2.FONT_HERSHEY_SIMPLEX, 0.5, (255, 255, 255), 1, cv2.LINE_AA)
return overlay
def run_one(damage_model: YOLO, sev_cnn: SeverityCNN | None,
img_path: Path, out_dir: Path, conf: float, imgsz: int) -> dict:
image = cv2.imread(str(img_path))
if image is None:
return {"image": str(img_path), "error": "read_failed"}
h, w = image.shape[:2]
pred = damage_model.predict(image, imgsz=imgsz, conf=conf, verbose=False)[0]
damages = []
for i, box in enumerate(pred.boxes):
cls_id = int(box.cls.item())
damage_type = CARDD_CLASSES[cls_id] if cls_id < len(CARDD_CLASSES) else f"cls_{cls_id}"
det_conf = float(box.conf.item())
x1, y1, x2, y2 = box.xyxy[0].tolist()
# NOTE: pred.masks.data is at YOLO's internal resolution (typically
# 160x160 or imgsz/4), not the original image size. To get an
# area-ratio that's comparable to the original frame we either
# resize the mask to (h, w) first, or compute the ratio against the
# mask's own pixel space — the latter is cheaper and equivalent.
if pred.masks is not None and i < len(pred.masks.data):
mask_t = pred.masks.data[i]
mh, mw = mask_t.shape[-2:]
area_pixels = float(mask_t.sum()) # pixels in mask-space
area_ratio = area_pixels / float(mh * mw)
else:
area_pixels = (x2 - x1) * (y2 - y1)
area_ratio = area_pixels / (h * w)
crop = crop_padded(image, (x1, y1, x2, y2))
sev_cnn_pred = sev_cnn.predict_bgr(crop) if sev_cnn else None
sev_rule = rule_based_severity(damage_type, area_ratio)
damages.append({
"id": i,
"type": damage_type,
"detection_confidence": round(det_conf, 4),
"bbox": {"x1": x1, "y1": y1, "x2": x2, "y2": y2},
"area_ratio": round(area_ratio, 6),
"severity_rule": sev_rule,
"severity_cnn": sev_cnn_pred,
})
summary = {
"image": img_path.name,
"width": w,
"height": h,
"damage_count": len(damages),
"classes_detected": sorted({d["type"] for d in damages}),
"max_detection_conf": round(max((d["detection_confidence"] for d in damages),
default=0.0), 4),
}
result = {"summary": summary, "damages": damages}
out_dir.mkdir(parents=True, exist_ok=True)
json_path = out_dir / f"{img_path.stem}.json"
json_path.write_text(json.dumps(result, indent=2, ensure_ascii=False))
overlay = annotate(image, damages)
overlay_path = out_dir / f"{img_path.stem}.overlay.jpg"
cv2.imwrite(str(overlay_path), overlay)
return result
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--damage_weights", type=Path, default=DEFAULT_DAMAGE_W)
ap.add_argument("--severity_weights", type=Path, default=DEFAULT_SEVERITY_W)
ap.add_argument("--val_dir", type=Path, default=DEFAULT_VAL_DIR)
ap.add_argument("--out_dir", type=Path, default=DEFAULT_OUT)
ap.add_argument("--num", type=int, default=3, help="Kac val gorseli denesin")
# Defaults track the production snapshot when present; otherwise we
# assume the legacy mini-test model (imgsz=480, low conf).
_is_prod_snap = _SNAP is not None and (_SNAP / "damage_best.pt").exists()
_default_conf = 0.25 if _is_prod_snap else 0.15
_default_imgsz = 640 if _is_prod_snap else 480
ap.add_argument("--conf", type=float, default=_default_conf,
help="Detection confidence threshold "
"(snapshot: 0.25, mini-test: 0.15)")
ap.add_argument("--imgsz", type=int, default=_default_imgsz,
help="Inference image size (snapshot: 640, mini-test: 480)")
ap.add_argument("--seed", type=int, default=42)
ap.add_argument("--image", type=Path, default=None,
help="Belirli bir gorsel ver, val secimini atla")
args = ap.parse_args()
if not args.damage_weights.exists():
raise FileNotFoundError(
f"Damage agirligi yok: {args.damage_weights}\n"
f"Once `python train.py` tamamla.")
print(f"[+] Damage model: {args.damage_weights}")
damage_model = YOLO(str(args.damage_weights))
sev_cnn = None
if args.severity_weights.exists():
print(f"[+] Severity model: {args.severity_weights}")
sev_cnn = SeverityCNN(args.severity_weights)
print(f" Classes: {sev_cnn.classes} → {sev_cnn.tr_names}")
if sev_cnn.train_val_acc is not None:
print(f" Egitim val_acc: {sev_cnn.train_val_acc:.3f}")
else:
print(f"[!] Severity agirligi yok: {args.severity_weights} (atlandi)")
if args.image:
images = [args.image]
else:
all_imgs = sorted(args.val_dir.glob("*.jpg"))
if not all_imgs:
raise FileNotFoundError(f"Val'da gorsel yok: {args.val_dir}")
random.seed(args.seed)
images = random.sample(all_imgs, min(args.num, len(all_imgs)))
print(f"\n[+] {len(images)} gorsel test ediliyor...")
print(f" out_dir: {args.out_dir}\n")
for img in images:
print(f"--- {img.name} ---")
res = run_one(damage_model, sev_cnn, img, args.out_dir,
conf=args.conf, imgsz=args.imgsz)
s = res["summary"]
print(f" damage_count: {s['damage_count']}, "
f"classes: {s['classes_detected']}, "
f"max_conf: {s['max_detection_conf']}")
for d in res["damages"]:
cnn = d["severity_cnn"]
cnn_str = f"{cnn['level']} ({cnn['confidence']:.2f})" if cnn else "n/a"
print(f" [{d['id']}] {d['type']} conf={d['detection_confidence']:.2f} "
f"area={d['area_ratio']:.4f} "
f"rule={d['severity_rule']['level']} cnn={cnn_str}")
print()
print(f"[OK] Tum sonuclar: {args.out_dir}")
if __name__ == "__main__":
main()
|