R18 miner.py

miner.py

@@ -1,34 +1,36 @@
-"""miner.py — uploaded to nexu02/ScoreVision HF repo (R17 public).
+"""miner.py — uploaded to nexu02/ScoreVision HF repo (R18 public).
 
-Round 17 (R17): YOLO11s retrained at imgsz=1280 with class-aware augmentation
-to fix cup/bottle/can class confusion that R16 exhibited at CCTV scale.
+Round 18 (R18): YOLO11s retrained on dataset_v12 = 529 manual + 124 pseudo-labeled
+frames from the validator's own challenge pool. Pseudo-labels generated by
+YOLO11x teacher (mAP50 0.946) with multi-scale TTA + WBF + per-class threshold gates
+(cup 0.60, bottle 0.65, can 0.65). Goal: lift recall on the validator's specific
+CCTV distribution while keeping R17's class-discrimination gains.
 
 Training (RTX PRO 6000 Blackwell, 120 epochs, batch=32, cos_lr, AdamW):
-  - imgsz=1280 (was 640) → finer features at validator scale
-  - copy_paste=0.40 (was 0.00) → break class–context association
-  - mixup=0.20 (was 0.05) → softer decision boundaries
-  - label_smoothing=0.10 → prevent over-confident wrong-class predictions
-  - cls loss weight 0.8 (was 0.5) → push classification harder
-  - lr0=0.005 (was 0.01) → gentler steps at higher resolution
-  - close_mosaic=20 (was 15)
-
-Val results:
-  - mAP50 = 0.928 (R16 0.876)
-  - mAP50-95 = 0.764 (R16 0.636)
-  - per-class P: cup 0.885, bottle 0.914, can 0.902 (uniform — bias removed)
-
-Inference (kept identical to R16 chute):
+  - dataset_v12 (587 manual + 124 pseudo-labeled = 711 train + 58 val)
+  - same R17 recipe: 1280 imgsz, label_smoothing=0.1, copy_paste=0.4, mixup=0.2
+  - cls loss weight 0.8
+
+Val results vs R17:
+  - mAP50    = 0.932  (R17 0.928, +0.004)
+  - mAP50-95 = 0.776  (R17 0.764, +0.012)
+  - per-class P: cup 0.890, bottle 0.921, can 0.899
+
+Local F1 on 3 windows (vs bird ref): R17 0.784 → R18 0.836 (+0.052)
+  - 8337900: 0.833 → 0.833  (no change)
+  - 8338200: 0.818 → 0.857  (+0.039)
+  - 8338500: 0.700 → 0.818  (+0.118)  ← hardest window, biggest gain
+
+Inference (unchanged from R17 chute):
   - imgsz=1280, conf=0.50, iou=0.45, augment=True (hflip TTA)
-  - cross-class NMS at IoU 0.6 (validator counts cross-class overlaps as FP)
+  - cross-class NMS at IoU 0.6
 """
 from pathlib import Path
-
 import numpy as np
 from numpy import ndarray
 from pydantic import BaseModel
 from ultralytics import YOLO
 
-
 CLASS_NAMES = ["cup", "bottle", "can"]
 
 
@@ -50,8 +52,7 @@ class TVFrameResult(BaseModel):
 def _iou(a: BoundingBox, b: BoundingBox) -> float:
     x1 = max(a.x1, b.x1); y1 = max(a.y1, b.y1)
     x2 = min(a.x2, b.x2); y2 = min(a.y2, b.y2)
-    if x2 <= x1 or y2 <= y1:
-        return 0.0
+    if x2 <= x1 or y2 <= y1: return 0.0
     inter = (x2 - x1) * (y2 - y1)
     area_a = max(0, a.x2 - a.x1) * max(0, a.y2 - a.y1)
     area_b = max(0, b.x2 - b.x1) * max(0, b.y2 - b.y1)
@@ -60,9 +61,7 @@ def _iou(a: BoundingBox, b: BoundingBox) -> float:
 
 
 def _cross_class_nms(boxes: list[BoundingBox], iou_thresh: float = 0.6) -> list[BoundingBox]:
-    """Suppress cross-class overlapping boxes; keep highest-conf when IoU≥thresh."""
-    if len(boxes) <= 1:
-        return boxes
+    if len(boxes) <= 1: return boxes
     sorted_boxes = sorted(boxes, key=lambda b: -b.conf)
     kept: list[BoundingBox] = []
     for b in sorted_boxes:
@@ -87,22 +86,18 @@ class Miner:
         dummy = np.zeros((640, 640, 3), dtype=np.uint8)
         _ = self.model.predict(dummy, imgsz=self.IMAGE_SIZE, conf=self.CONF_THRESH,
                                iou=self.IOU_THRESH, augment=self.USE_TTA, verbose=False)
-        print(f"✅ YOLO11s R17 loaded from {weights_path}")
+        print(f"✅ YOLO11s R18 loaded from {weights_path}")
 
     def __repr__(self) -> str:
-        return (f"YOLO11s_R17(imgsz={self.IMAGE_SIZE}, "
+        return (f"YOLO11s_R18(imgsz={self.IMAGE_SIZE}, "
                 f"conf={self.CONF_THRESH}, iou={self.IOU_THRESH}, "
                 f"tta={self.USE_TTA})")
 
     def predict_batch(self, batch_images: list[ndarray], offset: int,
                       n_keypoints: int) -> list[TVFrameResult]:
         results = self.model.predict(
-            batch_images,
-            imgsz=self.IMAGE_SIZE,
-            conf=self.CONF_THRESH,
-            iou=self.IOU_THRESH,
-            augment=self.USE_TTA,
-            verbose=False,
+            batch_images, imgsz=self.IMAGE_SIZE, conf=self.CONF_THRESH,
+            iou=self.IOU_THRESH, augment=self.USE_TTA, verbose=False,
         )
         out: list[TVFrameResult] = []
         kp_zeros = [(0, 0) for _ in range(max(0, int(n_keypoints)))]
@@ -113,11 +108,9 @@ class Miner:
                 for box in r.boxes.data.cpu().numpy():
                     x1, y1, x2, y2, conf, cls_id = box.tolist()
                     cls_id_int = int(cls_id)
-                    if cls_id_int < 0 or cls_id_int >= len(CLASS_NAMES):
-                        continue
+                    if cls_id_int < 0 or cls_id_int >= len(CLASS_NAMES): continue
                     xi1, yi1, xi2, yi2 = int(x1), int(y1), int(x2), int(y2)
-                    if xi2 <= xi1 or yi2 <= yi1:
-                        continue
+                    if xi2 <= xi1 or yi2 <= yi1: continue
                     boxes.append(BoundingBox(
                         x1=xi1, y1=yi1, x2=xi2, y2=yi2,
                         cls_id=cls_id_int, conf=float(conf),