R17 miner.py docstring + logging

miner.py

@@ -1,25 +1,25 @@
-"""miner.py — uploaded to nexu02/ScoreVision HF repo (R16 public).
-
-Round 16 (R16): YOLO11s fine-tuned on dataset_v11 = batches 1+2 manually-reviewed
-+ batch 3 raw (587 SAM3-anchored images, validator-distribution).
-
-Training (on RTX PRO 6000 Blackwell, 150 epochs, AdamW auto-lr, batch=32):
-  - val mAP50 = 0.878 (cup 0.890 / bottle 0.855 / can 0.890)
-  - val mAP50-95 = 0.644
-  - precision 0.886 (cup 0.856 / bottle 0.915 / can 0.887)
-  - recall 0.786
-  - close_mosaic=15 final epochs for sharper boxes
-  - patience=30 (didn't trigger; ran full 150)
-
-Inference (kept identical to R5/R6 to preserve calibration):
+"""miner.py — uploaded to nexu02/ScoreVision HF repo (R17 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.
+
+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):
   - 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)
-
-Required by the chute template:
-  - class Miner with __init__(self, path_hf_repo: Path)
-  - predict_batch(batch_images, offset, n_keypoints) -> list[TVFrameResult]
-  - BoundingBox + TVFrameResult pydantic models defined in this file
-  - No imports from anywhere except stdlib + site-packages
 """
 from pathlib import Path
 
@@ -29,8 +29,6 @@ from pydantic import BaseModel
 from ultralytics import YOLO
 
 
-# Class index follows the manifest objects order for Detect-beverage-detect:
-# [cup, bottle, can] → cls_id 0, 1, 2.
 CLASS_NAMES = ["cup", "bottle", "can"]
 
 
@@ -50,10 +48,8 @@ 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)
+    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
     inter = (x2 - x1) * (y2 - y1)
@@ -64,23 +60,15 @@ 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.
-
-    Ultralytics' default NMS only dedupes WITHIN a class. SN44 counts cross-class
-    overlap as a false positive (only one class can be right per object).
-    """
+    """Suppress cross-class overlapping boxes; keep highest-conf when IoU≥thresh."""
     if len(boxes) <= 1:
         return boxes
     sorted_boxes = sorted(boxes, key=lambda b: -b.conf)
     kept: list[BoundingBox] = []
     for b in sorted_boxes:
-        suppressed = False
-        for k in kept:
-            if _iou(b, k) >= iou_thresh:
-                suppressed = True
-                break
-        if not suppressed:
-            kept.append(b)
+        if any(_iou(b, k) >= iou_thresh for k in kept):
+            continue
+        kept.append(b)
     return kept
 
 
@@ -99,19 +87,15 @@ 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 R16 loaded from {weights_path}")
+        print(f"✅ YOLO11s R17 loaded from {weights_path}")
 
     def __repr__(self) -> str:
-        return (f"YOLO11s_R16(imgsz={self.IMAGE_SIZE}, "
+        return (f"YOLO11s_R17(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]:
+    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,
@@ -120,10 +104,8 @@ class Miner:
             augment=self.USE_TTA,
             verbose=False,
         )
-
         out: list[TVFrameResult] = []
         kp_zeros = [(0, 0) for _ in range(max(0, int(n_keypoints)))]
-
         for i, r in enumerate(results):
             frame_id = offset + i
             boxes: list[BoundingBox] = []
@@ -142,5 +124,4 @@ class Miner:
                     ))
             boxes = _cross_class_nms(boxes, iou_thresh=self.CROSS_CLASS_IOU)
             out.append(TVFrameResult(frame_id=frame_id, boxes=boxes, keypoints=kp_zeros))
-
         return out