onnxruntime miner.py (was ultralytics .pt loader)

miner.py

@@ -1,35 +1,26 @@
-"""miner.py — uploaded to nexu02/ScoreVision HF repo (R18 public).
-
-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):
-  - 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)
+"""miner.py — uploaded to nexu02/ScoreVision HF repo (R17 ONNX migration).
+
+Migrated from .pt → ONNX FP16 to comply with subnet requirement
+(.onnx-only models). Same R17 weights (mAP50 0.928, mAP50-95 0.764) +
+identical inference recipe to keep the #1 dashboard standing.
+
+Inference (same as R17 .pt version):
+  - imgsz=1280, conf=0.50, iou=0.45
+  - hflip TTA (manual: run twice, merge with per-class NMS)
   - cross-class NMS at IoU 0.6
+
+Runtime: onnxruntime-gpu (CUDAExecutionProvider) with CPU fallback.
+FP16 input/weights to fit under 30 MB HF cap (19.3 MB total).
 """
 from pathlib import Path
+import math
+
+import cv2
 import numpy as np
+import onnxruntime as ort
 from numpy import ndarray
 from pydantic import BaseModel
-from ultralytics import YOLO
+
 
 CLASS_NAMES = ["cup", "bottle", "can"]
 
@@ -49,72 +40,235 @@ class TVFrameResult(BaseModel):
     keypoints: list[tuple[int, int]]
 
 
-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
-    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)
-    union = area_a + area_b - inter
-    return inter / union if union > 0 else 0.0
+def _iou_xyxy(a: np.ndarray, b: np.ndarray) -> np.ndarray:
+    """Vectorised IoU between one box (a) and array of boxes (b)."""
+    xx1 = np.maximum(a[0], b[:, 0])
+    yy1 = np.maximum(a[1], b[:, 1])
+    xx2 = np.minimum(a[2], b[:, 2])
+    yy2 = np.minimum(a[3], b[:, 3])
+    inter = np.maximum(0.0, xx2 - xx1) * np.maximum(0.0, yy2 - yy1)
+    a_area = max(0.0, (a[2] - a[0]) * (a[3] - a[1]))
+    b_area = np.maximum(0.0, (b[:, 2] - b[:, 0]) * (b[:, 3] - b[:, 1]))
+    return inter / (a_area + b_area - inter + 1e-7)
+
+
+def _hard_nms(boxes: np.ndarray, scores: np.ndarray, iou_thr: float) -> np.ndarray:
+    """Per-class hard NMS — assumes boxes already filtered to one class."""
+    n = len(boxes)
+    if n == 0:
+        return np.array([], dtype=np.intp)
+    order = np.argsort(-scores)
+    keep = []
+    while len(order) > 0:
+        i = int(order[0])
+        keep.append(i)
+        if len(order) == 1:
+            break
+        rest = order[1:]
+        iou = _iou_xyxy(boxes[i], boxes[rest])
+        order = rest[iou <= iou_thr]
+    return np.array(keep, dtype=np.intp)
 
 
-def _cross_class_nms(boxes: list[BoundingBox], iou_thresh: float = 0.6) -> list[BoundingBox]:
-    if len(boxes) <= 1: return boxes
-    sorted_boxes = sorted(boxes, key=lambda b: -b.conf)
-    kept: list[BoundingBox] = []
-    for b in sorted_boxes:
-        if any(_iou(b, k) >= iou_thresh for k in kept):
+def _per_class_nms(boxes, scores, cls_ids, iou_thr):
+    if len(boxes) == 0:
+        return np.array([], dtype=np.intp)
+    keep_all = []
+    for c in np.unique(cls_ids):
+        m = cls_ids == c
+        idx = np.where(m)[0]
+        k = _hard_nms(boxes[m], scores[m], iou_thr)
+        keep_all.extend(idx[k].tolist())
+    keep_all.sort()
+    return np.array(keep_all, dtype=np.intp)
+
+
+def _cross_class_nms(boxes, scores, cls_ids, iou_thr):
+    """Cross-class NMS — drop overlapping boxes regardless of class."""
+    if len(boxes) <= 1:
+        return np.arange(len(boxes))
+    order = np.argsort(-scores)
+    keep = []
+    suppressed = np.zeros(len(boxes), dtype=bool)
+    for i in order:
+        if suppressed[i]:
             continue
-        kept.append(b)
-    return kept
+        keep.append(int(i))
+        iou = _iou_xyxy(boxes[i], boxes)
+        dup = iou > iou_thr
+        dup[i] = False
+        suppressed |= dup
+    return np.array(sorted(keep), dtype=np.intp)
 
 
 class Miner:
-    IMAGE_SIZE = 1280
-    CONF_THRESH = 0.50
-    IOU_THRESH = 0.45
-    USE_TTA = True
+    """R17 ONNX miner. Same recipe as .pt version: 1280 + flip TTA + cross-class NMS."""
+
+    INPUT_SIZE = 1280
+    CONF_THR = 0.50
+    IOU_THR = 0.45
     CROSS_CLASS_IOU = 0.6
 
     def __init__(self, path_hf_repo: Path) -> None:
-        weights_path = path_hf_repo / "best.pt"
-        if not weights_path.exists():
-            raise FileNotFoundError(f"missing weights at {weights_path}")
-        self.model = YOLO(str(weights_path))
-        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 R18 loaded from {weights_path}")
+        model_path = path_hf_repo / "best.onnx"
+        if not model_path.exists():
+            raise FileNotFoundError(f"missing weights at {model_path}")
+
+        print(f"ORT version: {ort.__version__}")
+        try:
+            ort.preload_dlls()
+        except Exception:
+            pass
+
+        sess_options = ort.SessionOptions()
+        sess_options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
+
+        try:
+            self.session = ort.InferenceSession(
+                str(model_path),
+                sess_options=sess_options,
+                providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
+            )
+        except Exception as e:
+            print(f"CUDA session failed, fallback CPU: {e}")
+            self.session = ort.InferenceSession(
+                str(model_path),
+                sess_options=sess_options,
+                providers=["CPUExecutionProvider"],
+            )
+        print(f"ORT providers: {self.session.get_providers()}")
+        for inp in self.session.get_inputs():
+            print(f"INPUT  {inp.name} shape={inp.shape} dtype={inp.type}")
+        for out in self.session.get_outputs():
+            print(f"OUTPUT {out.name} shape={out.shape} dtype={out.type}")
+
+        self.input_name = self.session.get_inputs()[0].name
+        # FP16 model expects float16 inputs
+        in_type = self.session.get_inputs()[0].type
+        self.input_dtype = np.float16 if "float16" in in_type else np.float32
+        print(f"✅ R17 ONNX loaded, input dtype={self.input_dtype.__name__}")
 
     def __repr__(self) -> str:
-        return (f"YOLO11s_R18(imgsz={self.IMAGE_SIZE}, "
-                f"conf={self.CONF_THRESH}, iou={self.IOU_THRESH}, "
-                f"tta={self.USE_TTA})")
+        return f"R17_ONNX(imgsz={self.INPUT_SIZE}, conf={self.CONF_THR}, iou={self.IOU_THR})"
+
+    def _letterbox(self, img: np.ndarray, size: int):
+        h, w = img.shape[:2]
+        r = min(size / w, size / h)
+        new_w, new_h = int(round(w * r)), int(round(h * r))
+        if (new_w, new_h) != (w, h):
+            interp = cv2.INTER_LINEAR
+            img = cv2.resize(img, (new_w, new_h), interpolation=interp)
+        dw, dh = (size - new_w) / 2.0, (size - new_h) / 2.0
+        top = int(round(dh - 0.1)); bottom = int(round(dh + 0.1))
+        left = int(round(dw - 0.1)); right = int(round(dw + 0.1))
+        padded = cv2.copyMakeBorder(img, top, bottom, left, right,
+                                    borderType=cv2.BORDER_CONSTANT, value=(114, 114, 114))
+        return padded, r, (dw, dh)
+
+    def _preprocess(self, img_bgr: np.ndarray):
+        h, w = img_bgr.shape[:2]
+        padded, r, pad = self._letterbox(img_bgr, self.INPUT_SIZE)
+        rgb = cv2.cvtColor(padded, cv2.COLOR_BGR2RGB)
+        x = rgb.astype(self.input_dtype) / 255.0
+        x = np.transpose(x, (2, 0, 1))[None, ...]
+        return np.ascontiguousarray(x, dtype=self.input_dtype), r, pad, (w, h)
+
+    def _decode_raw(self, raw: np.ndarray, r: float, pad, orig_size):
+        """Decode YOLO11 raw output (1, 7, N) → boxes + scores + class.
+        Output shape: 4 box (xywh) + 3 class scores.
+        """
+        if raw.ndim == 3:
+            raw = raw[0]
+        if raw.shape[0] < raw.shape[1]:
+            raw = raw.T  # → (N, 7)
+        boxes_xywh = raw[:, :4].astype(np.float32)
+        cls_scores = raw[:, 4:].astype(np.float32)
+        cls_ids = np.argmax(cls_scores, axis=1)
+        scores = cls_scores[np.arange(len(cls_scores)), cls_ids]
+
+        keep = scores >= self.CONF_THR
+        if not keep.any():
+            return (np.empty((0, 4)), np.empty(0), np.empty(0, dtype=int))
+        boxes_xywh, scores, cls_ids = boxes_xywh[keep], scores[keep], cls_ids[keep]
+
+        # xywh → xyxy
+        boxes = np.empty_like(boxes_xywh)
+        boxes[:, 0] = boxes_xywh[:, 0] - boxes_xywh[:, 2] / 2
+        boxes[:, 1] = boxes_xywh[:, 1] - boxes_xywh[:, 3] / 2
+        boxes[:, 2] = boxes_xywh[:, 0] + boxes_xywh[:, 2] / 2
+        boxes[:, 3] = boxes_xywh[:, 1] + boxes_xywh[:, 3] / 2
+
+        # Undo letterbox padding/scale
+        pad_w, pad_h = pad
+        boxes[:, [0, 2]] -= pad_w
+        boxes[:, [1, 3]] -= pad_h
+        boxes /= r
+
+        # Clip to original image
+        w, h = orig_size
+        boxes[:, [0, 2]] = np.clip(boxes[:, [0, 2]], 0, w - 1)
+        boxes[:, [1, 3]] = np.clip(boxes[:, [1, 3]], 0, h - 1)
+
+        return boxes, scores, cls_ids
+
+    def _predict_single(self, img_bgr: np.ndarray):
+        x, r, pad, orig = self._preprocess(img_bgr)
+        out = self.session.run(None, {self.input_name: x})[0]
+        return self._decode_raw(out, r, pad, orig)
+
+    def _predict_with_tta(self, img_bgr: np.ndarray):
+        """Predict + horizontal flip TTA, merge with per-class NMS."""
+        boxes1, scores1, cls1 = self._predict_single(img_bgr)
+        flipped = cv2.flip(img_bgr, 1)
+        boxes2, scores2, cls2 = self._predict_single(flipped)
+        if len(boxes2):
+            w = img_bgr.shape[1]
+            new = boxes2.copy()
+            new[:, 0] = w - boxes2[:, 2]
+            new[:, 2] = w - boxes2[:, 0]
+            boxes2 = new
+        if not len(boxes1) and not len(boxes2):
+            return np.empty((0, 4)), np.empty(0), np.empty(0, dtype=int)
+        boxes = np.concatenate([boxes1, boxes2]) if len(boxes1) and len(boxes2) else (boxes1 if len(boxes1) else boxes2)
+        scores = np.concatenate([scores1, scores2]) if len(boxes1) and len(boxes2) else (scores1 if len(scores1) else scores2)
+        cls_ids = np.concatenate([cls1, cls2]) if len(boxes1) and len(boxes2) else (cls1 if len(cls1) else cls2)
+        keep = _per_class_nms(boxes, scores, cls_ids, self.IOU_THR)
+        return boxes[keep], scores[keep], cls_ids[keep]
 
     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,
-        )
+                       n_keypoints: int) -> list[TVFrameResult]:
         out: list[TVFrameResult] = []
         kp_zeros = [(0, 0) for _ in range(max(0, int(n_keypoints)))]
-        for i, r in enumerate(results):
+        for i, image in enumerate(batch_images):
             frame_id = offset + i
-            boxes: list[BoundingBox] = []
-            if r.boxes is not None and r.boxes.data is not None:
-                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
-                    xi1, yi1, xi2, yi2 = int(x1), int(y1), int(x2), int(y2)
-                    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),
+            try:
+                if image is None or image.ndim != 3 or image.shape[2] != 3:
+                    out.append(TVFrameResult(frame_id=frame_id, boxes=[], keypoints=kp_zeros))
+                    continue
+                if image.dtype != np.uint8:
+                    image = image.astype(np.uint8)
+
+                boxes, scores, cls_ids = self._predict_with_tta(image)
+                if len(boxes):
+                    # Cross-class NMS (validator counts cross-class overlap as FP)
+                    keep = _cross_class_nms(boxes, scores, cls_ids, self.CROSS_CLASS_IOU)
+                    boxes, scores, cls_ids = boxes[keep], scores[keep], cls_ids[keep]
+
+                results = []
+                for b, s, c in zip(boxes, scores, cls_ids):
+                    x1, y1, x2, y2 = b
+                    if x2 <= x1 or y2 <= y1:
+                        continue
+                    c_int = int(c)
+                    if c_int < 0 or c_int >= len(CLASS_NAMES):
+                        continue
+                    results.append(BoundingBox(
+                        x1=int(math.floor(x1)), y1=int(math.floor(y1)),
+                        x2=int(math.ceil(x2)),  y2=int(math.ceil(y2)),
+                        cls_id=c_int, conf=float(s),
                     ))
-            boxes = _cross_class_nms(boxes, iou_thresh=self.CROSS_CLASS_IOU)
-            out.append(TVFrameResult(frame_id=frame_id, boxes=boxes, keypoints=kp_zeros))
+                out.append(TVFrameResult(frame_id=frame_id, boxes=results, keypoints=kp_zeros))
+            except Exception as e:
+                print(f"Inference err for frame {frame_id}: {e}")
+                out.append(TVFrameResult(frame_id=frame_id, boxes=[], keypoints=kp_zeros))
         return out