scorevision: push artifact

miner.py

@@ -1,3 +1,13 @@
+"""
+Score Vision SN44 — VehicleDetect miner. v2 (2026-03-25).
+
+Model: YOLO11n ONNX, 4 classes trained as:
+  0 = car, 1 = bus, 2 = truck, 3 = motorcycle
+
+Official submission order (remapped in MODEL_TO_OUT):
+  0 = bus, 1 = car, 2 = truck, 3 = motorcycle
+"""
+
 from pathlib import Path
 import math
 
@@ -8,6 +18,17 @@ from numpy import ndarray
 from pydantic import BaseModel
 
 
+# ── Model class index → submission class index ───────────────────────────────
+# Trained order: car=0, bus=1, truck=2, motorcycle=3
+# Official order: bus=0, car=1, truck=2, motorcycle=3
+MODEL_TO_OUT: dict[int, int] = {0: 1, 1: 0, 2: 2, 3: 3}
+OUT_NAMES = ["bus", "car", "truck", "motorcycle"]
+
+IMG_SIZE = 640
+CONF_THRESH = 0.55   # sweep-optimised: max composite (0.60×mAP + 0.40×FP_score)
+IOU_THRESH = 0.45
+
+
 class BoundingBox(BaseModel):
     x1: int
     y1: int
@@ -25,134 +46,117 @@ class TVFrameResult(BaseModel):
 
 class Miner:
     """
-    Auto-generated by subnet_bridge from a Manako element repo.
-    This miner is intentionally self-contained for chute import restrictions.
+    VehicleDetect miner for SN44. Loaded by turbovision template at startup.
     """
 
     def __init__(self, path_hf_repo: Path) -> None:
         self.path_hf_repo = path_hf_repo
-        self.class_names = ['person']
         self.session = ort.InferenceSession(
             str(path_hf_repo / "weights.onnx"),
             providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
         )
         self.input_name = self.session.get_inputs()[0].name
-        input_shape = self.session.get_inputs()[0].shape
-        # expected [N, C, H, W]
-        self.input_h = int(input_shape[2])
-        self.input_w = int(input_shape[3])
-        self.conf_threshold = 0.70  # sweep-optimised: max composite 0.65×mAP+0.35×FP_score
-        self.iou_threshold = 0.45
+        self.conf_threshold = CONF_THRESH
+        self.iou_threshold = IOU_THRESH
 
     def __repr__(self) -> str:
-        return f"ONNX Miner session={type(self.session).__name__} classes={len(self.class_names)}"
-
-    def _preprocess(self, image_bgr: ndarray) -> tuple[np.ndarray, tuple[int, int]]:
-        h, w = image_bgr.shape[:2]
-        rgb = cv2.cvtColor(image_bgr, cv2.COLOR_BGR2RGB)
-        resized = cv2.resize(rgb, (self.input_w, self.input_h))
-        x = resized.astype(np.float32) / 255.0
-        x = np.transpose(x, (2, 0, 1))[None, ...]
-        return x, (h, w)
-
-    def _normalize_predictions(self, raw: np.ndarray) -> np.ndarray:
-        # Common ultralytics export shapes:
-        # - [1, C, N] where C=4+num_classes
-        # - [1, N, C]
-        pred = raw[0]
-        if pred.ndim != 2:
-            raise ValueError(f"Unexpected prediction shape: {raw.shape}")
-        if pred.shape[0] < pred.shape[1]:
-            pred = pred.transpose(1, 0)
-        return pred
+        return f"VehicleDetect Miner session={type(self.session).__name__}"
+
+    def _letterbox(self, img: ndarray) -> tuple[np.ndarray, float, int, int]:
+        h, w = img.shape[:2]
+        r = min(IMG_SIZE / h, IMG_SIZE / w)
+        new_w, new_h = int(round(w * r)), int(round(h * r))
+        img_r = cv2.resize(img, (new_w, new_h), interpolation=cv2.INTER_LINEAR)
+        dw, dh = IMG_SIZE - new_w, IMG_SIZE - new_h
+        pad_l, pad_t = dw // 2, dh // 2
+        img_p = cv2.copyMakeBorder(
+            img_r, pad_t, dh - pad_t, pad_l, dw - pad_l,
+            cv2.BORDER_CONSTANT, value=(114, 114, 114),
+        )
+        return img_p, r, pad_l, pad_t
 
-    def _nms(self, dets: list[tuple[float, float, float, float, float, int]]) -> list[tuple[float, float, float, float, float, int]]:
-        if not dets:
-            return []
+    def _preprocess(self, image_bgr: ndarray) -> tuple[np.ndarray, float, int, int]:
+        img_p, ratio, pad_l, pad_t = self._letterbox(image_bgr)
+        img_rgb = cv2.cvtColor(img_p, cv2.COLOR_BGR2RGB)
+        inp = img_rgb.astype(np.float32) / 255.0
+        inp = np.ascontiguousarray(inp.transpose(2, 0, 1)[np.newaxis])
+        return inp, ratio, pad_l, pad_t
 
-        boxes = np.array([[d[0], d[1], d[2], d[3]] for d in dets], dtype=np.float32)
-        scores = np.array([d[4] for d in dets], dtype=np.float32)
+    def _nms(self, boxes: np.ndarray, scores: np.ndarray) -> list[int]:
+        if not len(boxes):
+            return []
+        x1, y1, x2, y2 = boxes[:, 0], boxes[:, 1], boxes[:, 2], boxes[:, 3]
+        areas = (x2 - x1) * (y2 - y1)
         order = scores.argsort()[::-1]
-        keep = []
-
-        while order.size > 0:
+        keep: list[int] = []
+        while len(order):
             i = order[0]
-            keep.append(i)
-
-            xx1 = np.maximum(boxes[i, 0], boxes[order[1:], 0])
-            yy1 = np.maximum(boxes[i, 1], boxes[order[1:], 1])
-            xx2 = np.minimum(boxes[i, 2], boxes[order[1:], 2])
-            yy2 = np.minimum(boxes[i, 3], boxes[order[1:], 3])
-
-            w = np.maximum(0.0, xx2 - xx1)
-            h = np.maximum(0.0, yy2 - yy1)
-            inter = w * h
-
-            area_i = (boxes[i, 2] - boxes[i, 0]) * (boxes[i, 3] - boxes[i, 1])
-            area_rest = (boxes[order[1:], 2] - boxes[order[1:], 0]) * (boxes[order[1:], 3] - boxes[order[1:], 1])
-            union = np.maximum(area_i + area_rest - inter, 1e-6)
-            iou = inter / union
-
-            remaining = np.where(iou <= self.iou_threshold)[0]
-            order = order[remaining + 1]
-
-        return [dets[idx] for idx in keep]
+            keep.append(int(i))
+            xx1 = np.maximum(x1[i], x1[order[1:]])
+            yy1 = np.maximum(y1[i], y1[order[1:]])
+            xx2 = np.minimum(x2[i], x2[order[1:]])
+            yy2 = np.minimum(y2[i], y2[order[1:]])
+            inter = np.maximum(0, xx2 - xx1) * np.maximum(0, yy2 - yy1)
+            iou = inter / (areas[i] + areas[order[1:]] - inter + 1e-7)
+            order = order[1:][iou <= self.iou_threshold]
+        return keep
 
     def _infer_single(self, image_bgr: ndarray) -> list[BoundingBox]:
-        inp, (orig_h, orig_w) = self._preprocess(image_bgr)
-        out = self.session.run(None, {self.input_name: inp})[0]
-        pred = self._normalize_predictions(out)
+        orig_h, orig_w = image_bgr.shape[:2]
+        inp, ratio, pad_l, pad_t = self._preprocess(image_bgr)
+        raw = self.session.run(None, {self.input_name: inp})[0]
 
-        if pred.shape[1] < 5:
-            return []
+        # Output: [1, 8, 8400] → pred: [8, 8400] → [8400, 8]
+        pred = raw[0]
+        if pred.shape[0] < pred.shape[1]:
+            pred = pred.T  # [8400, 8]
 
-        boxes = pred[:, :4]
-        cls_scores = pred[:, 4:]
-
-        if cls_scores.shape[1] == 0:
-            return []
+        bboxes_cx = pred[:, :4]   # cx, cy, w, h in letterboxed coords
+        cls_scores = pred[:, 4:]  # [8400, 4]
 
         cls_ids = np.argmax(cls_scores, axis=1)
         confs = np.max(cls_scores, axis=1)
-        keep = confs >= self.conf_threshold
-
-        boxes = boxes[keep]
-        confs = confs[keep]
-        cls_ids = cls_ids[keep]
+        mask = confs >= self.conf_threshold
 
-        if boxes.shape[0] == 0:
+        if not mask.any():
             return []
 
-        sx = orig_w / float(self.input_w)
-        sy = orig_h / float(self.input_h)
+        bboxes_cx = bboxes_cx[mask]
+        confs = confs[mask]
+        cls_ids = cls_ids[mask]
 
-        dets: list[tuple[float, float, float, float, float, int]] = []
-        for i in range(boxes.shape[0]):
-            cx, cy, bw, bh = boxes[i].tolist()
-            x1 = (cx - bw / 2.0) * sx
-            y1 = (cy - bh / 2.0) * sy
-            x2 = (cx + bw / 2.0) * sx
-            y2 = (cy + bh / 2.0) * sy
-            dets.append((x1, y1, x2, y2, float(confs[i]), int(cls_ids[i])))
+        # cx,cy,w,h → x1,y1,x2,y2 (in letterboxed image coords)
+        cx, cy, bw, bh = bboxes_cx[:, 0], bboxes_cx[:, 1], bboxes_cx[:, 2], bboxes_cx[:, 3]
+        lx1 = cx - bw / 2
+        ly1 = cy - bh / 2
+        lx2 = cx + bw / 2
+        ly2 = cy + bh / 2
 
-        dets = self._nms(dets)
+        # Unletterbox back to original image coords
+        x1 = np.clip((lx1 - pad_l) / ratio, 0, orig_w)
+        y1 = np.clip((ly1 - pad_t) / ratio, 0, orig_h)
+        x2 = np.clip((lx2 - pad_l) / ratio, 0, orig_w)
+        y2 = np.clip((ly2 - pad_t) / ratio, 0, orig_h)
+        boxes = np.stack([x1, y1, x2, y2], axis=1)
 
         out_boxes: list[BoundingBox] = []
-        for x1, y1, x2, y2, conf, cls_id in dets:
-            ix1 = max(0, min(orig_w, math.floor(x1)))
-            iy1 = max(0, min(orig_h, math.floor(y1)))
-            ix2 = max(0, min(orig_w, math.ceil(x2)))
-            iy2 = max(0, min(orig_h, math.ceil(y2)))
-            out_boxes.append(
-                BoundingBox(
-                    x1=ix1,
-                    y1=iy1,
-                    x2=ix2,
-                    y2=iy2,
-                    cls_id=cls_id,
+        for model_cls in range(4):
+            cls_mask = cls_ids == model_cls
+            if not cls_mask.any():
+                continue
+            keep = self._nms(boxes[cls_mask], confs[cls_mask])
+            sub_cls = MODEL_TO_OUT[model_cls]
+            for k in keep:
+                box = boxes[cls_mask][k]
+                conf = float(confs[cls_mask][k])
+                out_boxes.append(BoundingBox(
+                    x1=max(0, min(orig_w, math.floor(box[0]))),
+                    y1=max(0, min(orig_h, math.floor(box[1]))),
+                    x2=max(0, min(orig_w, math.ceil(box[2]))),
+                    y2=max(0, min(orig_h, math.ceil(box[3]))),
+                    cls_id=sub_cls,
                     conf=max(0.0, min(1.0, conf)),
-                )
-            )
+                ))
         return out_boxes
 
     def predict_batch(
@@ -165,11 +169,9 @@ class Miner:
         for idx, image in enumerate(batch_images):
             boxes = self._infer_single(image)
             keypoints = [(0, 0) for _ in range(max(0, int(n_keypoints)))]
-            results.append(
-                TVFrameResult(
-                    frame_id=offset + idx,
-                    boxes=boxes,
-                    keypoints=keypoints,
-                )
-            )
+            results.append(TVFrameResult(
+                frame_id=offset + idx,
+                boxes=boxes,
+                keypoints=keypoints,
+            ))
         return results