subnet_bridge: copy winning miner repo into library

README.md

@@ -10,18 +10,40 @@ tags:
 manako:
   source: winner_fetch
   manifest_element_name: manak0/Detect-fire
-  winner_repo_id: coolroman/firenet
-  winner_revision: e5b4a47a00c97e3541c37a8a54eb9e4fa9c0b13e
-  note: E=0.03037871 (map50=0.600000, size_mb=19.750677)
+  winner_repo_id: meaculpitt/ScoreVision-Fire
+  winner_revision: 71ae3d3e59ced8b330eea5e95710318175bb1342
+  note: E=0.11785877 (map50=0.600000, size_mb=5.090839)
 ---
 
-# Detect-fire-winner
+# ScoreVision-Fire — meaculpitt v2.1
 
-Published winning miner converted into a library element.
+SN44 fire-detection miner for the `manak0/Detect-fire` element.
 
-- Source winner repo: `coolroman/firenet`
-- Source revision: `e5b4a47a00c97e3541c37a8a54eb9e4fa9c0b13e`
-- Manifest element: `manak0/Detect-fire`
-- Element type: `Detect`
-- Objects: fire, smoke, fire extinguisher
-- Runtime model type: `onnxruntime`
+## Pipeline
+- **Architecture**: yolo26n
+- **Resolution**: 1408×768 input → letterbox → 960×960
+- **Preprocessing**: `cv2.dnn.blobFromImage` (fused C++ resize+normalize+transpose)
+- **Inference**: single-pass FP16 ONNX, NMS baked in
+- **Output shape**: `[1, 300, 6]` (xyxy, conf, cls)
+- **Latency**: ~35 ms p95 on RTX 4090 (fits the 50 ms gate)
+
+## Classes (validator GT order, NOT the published class_names.txt order)
+- 0: fire
+- 1: smoke
+- 2: fire extinguisher
+
+Verified by audit of alfred8995/fire001 (scores 1.00) and navierstocks/fire
+(scores 0.96): both use [fire, smoke, fire_extinguisher] and the validator's
+GT order matches. Our model was trained with [fire, fire_ext, smoke]; miner.py
+applies cls_remap=[0,2,1] to translate model output to validator index.
+
+## Training
+- 22,796 training images (validator-synth + Simuletic + D-Fire + z5atr, SHA1 deduped)
+- 2,532 validation images (random 90/10 split, seed=42)
+- 100 epochs, yolo26n, imgsz=960, batch=8, AdamW lr0=0.001 cos_lr
+- CCTV augmentation chain (cctv_aug_patch)
+
+## Benchmarks
+- Broader merged val mAP50: 0.785
+- Validator-distribution synth val mAP50: 0.640 (+24.7 pts above 0.393 baseline)
+- Per-class on synth val: fire=0.523, fire_extinguisher=0.647, smoke=0.749

chute_config.yml

@@ -8,14 +8,22 @@ Image:
 NodeSelector:
   gpu_count: 1
   min_vram_gb_per_gpu: 16
-  max_hourly_price_per_gpu: 2
+  # SN44 chute platform mandates TEE + pro_6000 include for new elements
+  # (verified by crime + beverage deploys 2026-05-04). Cheaper-GPU config
+  # caused repeated 500 ContentTypeError on POST /chutes/.
+  max_hourly_price_per_gpu: 2.00
   include:
-    - pro_6000
+    - "pro_6000"
+  exclude:
+    - "5090"
+    - b200
+    - h200
+    - h20
+    - mi300x
 
 Chute:
-  timeout_seconds: 900
   concurrency: 4
   max_instances: 5
   scaling_threshold: 0.5
-  shutdown_after_seconds: 288000
-  tee: true
+  shutdown_after_seconds: 288000   # 80h idle
+  tee: true

class_names.txt

@@ -0,0 +1,3 @@
+fire
+smoke
+fire extinguisher

miner.py

@@ -1,10 +1,24 @@
-from pathlib import Path
+# build-marker: fire-v2-blob-imgsz960
+"""SN44 fire detection miner — yolo26n single-pass @ imgsz=960.
+
+v2 (2026-05-09): trained on merged 25k pool (validator-synth + D-Fire +
+Simuletic + z5atr). FP16 ONNX, ~5 MB. Single forward pass at imgsz=960
+fits the 50 ms p95 latency gate (~35 ms on 4090, blobFromImage preproc).
+
+SAHI tiling was tested but blew the latency budget (5x preproc/postproc
+overhead). Code preserved at fire/deploy/miner_sahi.py for later experiments.
+
+Classes (validator order from manak0/Detect-fire class_names.txt):
+  0=fire, 1=fire extinguisher, 2=smoke
+
+Single ONNX expected at path_hf_repo/weights.onnx (yolo26n e2e [1,300,6]).
+"""
 import math
+from pathlib import Path
 
 import cv2
 import numpy as np
 import onnxruntime as ort
-from numpy import ndarray
 from pydantic import BaseModel
 
 
@@ -24,270 +38,144 @@ class TVFrameResult(BaseModel):
 
 
 class Miner:
-    def __init__(self,
-        path_hf_repo: Path
-    ) -> None:
-        model_path = path_hf_repo / "weights.onnx"
-        self.class_names = ["fire","smoke","fire extinguisher"]
-        model_class_order = ["fire","smoke","fire extinguisher"]
+    def __init__(self, path_hf_repo) -> None:
+        self.path_hf_repo = Path(path_hf_repo)
+        # Validator's actual GT class order is [fire, smoke, fire extinguisher]
+        # — verified by audit of alfred8995/fire001 (scores 1.00) and
+        # navierstocks/fire (scores 0.96), both using this order. The published
+        # manak0/Detect-fire class_names.txt list [fire, fire_ext, smoke] does
+        # NOT match the actual scoring index.
+        # Our model was trained with [fire, fire_ext, smoke] (cls=1=ext, cls=2=smoke).
+        # cls_remap translates model output index → validator GT index.
+        self.class_names = ["fire", "smoke", "fire extinguisher"]
+        model_class_order = ["fire", "fire extinguisher", "smoke"]
         self.cls_remap = np.array(
-            [self.class_names.index(n) for n in model_class_order], dtype=np.int32
-        )
-        print("ORT version:", ort.__version__)
+            [self.class_names.index(n) for n in model_class_order],
+            dtype=np.int32,
+        )  # → [0, 2, 1]: model cls 0→0, 1→2, 2→1
 
         try:
             ort.preload_dlls()
-            print("✅ onnxruntime.preload_dlls() success")
-        except Exception as e:
-            print(f"⚠️ preload_dlls failed: {e}")
-
-        print("ORT available providers BEFORE session:", ort.get_available_providers())
+        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),
+                str(self.path_hf_repo / "weights.onnx"),
                 sess_options=sess_options,
                 providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
             )
-            print("✅ Created ORT session with preferred CUDA provider list")
-        except Exception as e:
-            print(f"⚠️ CUDA session creation failed, falling back to CPU: {e}")
+        except Exception:
             self.session = ort.InferenceSession(
-                str(model_path),
+                str(self.path_hf_repo / "weights.onnx"),
                 sess_options=sess_options,
                 providers=["CPUExecutionProvider"],
             )
-
-        print("ORT session providers:", self.session.get_providers())
-
-        for inp in self.session.get_inputs():
-            print("INPUT:", inp.name, inp.shape, inp.type)
-
-        for out in self.session.get_outputs():
-            print("OUTPUT:", out.name, out.shape, out.type)
-
         self.input_name = self.session.get_inputs()[0].name
-        self.output_names = [output.name for output in self.session.get_outputs()]
-        self.input_shape = self.session.get_inputs()[0].shape
-
-        # Your export is fixed-size 1280, but we still read actual ONNX input shape first.
-        self.input_height = self._safe_dim(self.input_shape[2], default=960)
-        self.input_width = self._safe_dim(self.input_shape[3], default=960)
-
-        # Tuned for validator scoring: reduce FP (FALSE_POSITIVE pillar),
-        # preserve recall (MAP50, RECALL), improve precision.
-        self.conf_thres = 0.36  # Higher = fewer FP, slightly lower recall
-        self.iou_thres = 0.43   # Lower = suppress duplicate detections (FP)
-        self.max_det = 150      # Cap detections per image
-        self.use_tta = True
-
-        # Box sanity: filter tiny/spurious detections (common FP source)
-        self.min_box_area = 14 * 14  # ~144 px²
-        self.min_side = 8
-        self.max_aspect_ratio = 8.0
-
-        print(f"✅ ONNX model loaded from: {model_path}")
-        print(f"✅ ONNX providers: {self.session.get_providers()}")
-        print(f"✅ ONNX input: name={self.input_name}, shape={self.input_shape}")
-
-    def __repr__(self) -> str:
-        return (
-            f"ONNXRuntime(session={type(self.session).__name__}, "
-            f"providers={self.session.get_providers()})"
-        )
-
-    @staticmethod
-    def _safe_dim(value, default: int) -> int:
-        return value if isinstance(value, int) and value > 0 else default
-
-    def _letterbox(
-        self,
-        image: ndarray,
-        new_shape: tuple[int, int],
-        color=(114, 114, 114),
-    ) -> tuple[ndarray, float, tuple[float, float]]:
-        """
-        Resize with unchanged aspect ratio and pad to target shape.
-        Returns:
-            padded_image,
-            ratio,
-            (pad_w, pad_h)  # half-padding
-        """
-        h, w = image.shape[:2]
-        new_w, new_h = new_shape
-
-        ratio = min(new_w / w, new_h / h)
-        resized_w = int(round(w * ratio))
-        resized_h = int(round(h * ratio))
-
-        if (resized_w, resized_h) != (w, h):
+        self.output_names = [o.name for o in self.session.get_outputs()]
+        self.input_dtype = (np.float16
+                            if 'float16' in self.session.get_inputs()[0].type
+                            else np.float32)
+
+        self.input_h = 960
+        self.input_w = 960
+        self.conf_thres_per_class = np.array([0.20, 0.20, 0.20], dtype=np.float32)
+        self.iou_thresh = 0.5
+        self.cross_iou_thresh = 0.7
+        self.max_det = 100
+        self.min_box_area = 64
+        self.min_side = 6
+        self.max_aspect_ratio = 10.0
+
+        warm = np.zeros((768, 1408, 3), dtype=np.uint8)
+        for _ in range(3):
+            try: self._infer_single(warm)
+            except Exception: break
+
+    def __repr__(self):
+        thr = ",".join(f"{n[:4]}={t:.2f}" for n, t
+                       in zip(self.class_names, self.conf_thres_per_class.tolist()))
+        return (f"FireMiner v2  yolo26n@{self.input_w} single-pass blob  "
+                f"conf=[{thr}]  iou={self.iou_thresh}")
+
+    def _preprocess(self, image_bgr):
+        """Letterbox + cv2.dnn.blobFromImage (fused C++ resize/normalize/transpose)."""
+        h, w = image_bgr.shape[:2]
+        ratio = min(self.input_w / w, self.input_h / h)
+        nw, nh = int(round(w * ratio)), int(round(h * ratio))
+        if (nw, nh) != (w, h):
             interp = cv2.INTER_CUBIC if ratio > 1.0 else cv2.INTER_LINEAR
-            image = cv2.resize(image, (resized_w, resized_h), interpolation=interp)
-
-        dw = new_w - resized_w
-        dh = new_h - resized_h
-        dw /= 2.0
-        dh /= 2.0
-
-        left = int(round(dw - 0.1))
-        right = int(round(dw + 0.1))
-        top = int(round(dh - 0.1))
-        bottom = int(round(dh + 0.1))
-
-        padded = cv2.copyMakeBorder(
-            image,
-            top,
-            bottom,
-            left,
-            right,
-            borderType=cv2.BORDER_CONSTANT,
-            value=color,
-        )
-        return padded, ratio, (dw, dh)
-
-    def _preprocess(
-        self, image: ndarray
-    ) -> tuple[np.ndarray, float, tuple[float, float], tuple[int, int]]:
-        """
-        Preprocess for fixed-size ONNX export:
-        - enhance image quality (CLAHE, denoise, sharpen)
-        - letterbox to model input size
-        - BGR -> RGB
-        - normalize to [0,1]
-        - HWC -> NCHW float32
-        """
-        orig_h, orig_w = image.shape[:2]
-
-        img, ratio, pad = self._letterbox(
-            image, (self.input_width, self.input_height)
+            resized = cv2.resize(image_bgr, (nw, nh), interpolation=interp)
+        else:
+            resized = image_bgr
+        canvas = np.full((self.input_h, self.input_w, 3), 114, dtype=np.uint8)
+        dy = (self.input_h - nh) // 2
+        dx = (self.input_w - nw) // 2
+        canvas[dy:dy+nh, dx:dx+nw] = resized
+        # blobFromImage: fused BGR→RGB (swapRB) + /255 + transpose CHW + add batch dim
+        blob = cv2.dnn.blobFromImage(
+            canvas, scalefactor=1/255.0,
+            size=(self.input_w, self.input_h),
+            mean=(0, 0, 0), swapRB=True, crop=False,
         )
-        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
-        img = img.astype(np.float32) / 255.0
-        img = np.transpose(img, (2, 0, 1))[None, ...]
-        img = np.ascontiguousarray(img, dtype=np.float32)
-
-        return img, ratio, pad, (orig_w, orig_h)
-
-    @staticmethod
-    def _clip_boxes(boxes: np.ndarray, image_size: tuple[int, int]) -> np.ndarray:
-        w, h = image_size
-        boxes[:, 0] = np.clip(boxes[:, 0], 0, w - 1)
-        boxes[:, 1] = np.clip(boxes[:, 1], 0, h - 1)
-        boxes[:, 2] = np.clip(boxes[:, 2], 0, w - 1)
-        boxes[:, 3] = np.clip(boxes[:, 3], 0, h - 1)
-        return boxes
-
-    @staticmethod
-    def _xywh_to_xyxy(boxes: np.ndarray) -> np.ndarray:
-        out = np.empty_like(boxes)
-        out[:, 0] = boxes[:, 0] - boxes[:, 2] / 2.0
-        out[:, 1] = boxes[:, 1] - boxes[:, 3] / 2.0
-        out[:, 2] = boxes[:, 0] + boxes[:, 2] / 2.0
-        out[:, 3] = boxes[:, 1] + boxes[:, 3] / 2.0
-        return out
-
-    def _soft_nms(
-        self,
-        boxes: np.ndarray,
-        scores: np.ndarray,
-        sigma: float = 0.5,
-        score_thresh: float = 0.01,
-    ) -> tuple[np.ndarray, np.ndarray]:
-        """
-        Soft-NMS: Gaussian decay of overlapping scores instead of hard removal.
-        Returns (kept_original_indices, updated_scores).
-        """
-        N = len(boxes)
-        if N == 0:
-            return np.array([], dtype=np.intp), np.array([], dtype=np.float32)
-
-        boxes = boxes.astype(np.float32, copy=True)
-        scores = scores.astype(np.float32, copy=True)
-        order = np.arange(N)
-
-        for i in range(N):
-            max_pos = i + int(np.argmax(scores[i:]))
-            boxes[[i, max_pos]] = boxes[[max_pos, i]]
-            scores[[i, max_pos]] = scores[[max_pos, i]]
-            order[[i, max_pos]] = order[[max_pos, i]]
-
-            if i + 1 >= N:
-                break
-
-            xx1 = np.maximum(boxes[i, 0], boxes[i + 1:, 0])
-            yy1 = np.maximum(boxes[i, 1], boxes[i + 1:, 1])
-            xx2 = np.minimum(boxes[i, 2], boxes[i + 1:, 2])
-            yy2 = np.minimum(boxes[i, 3], boxes[i + 1:, 3])
-            inter = np.maximum(0.0, xx2 - xx1) * np.maximum(0.0, yy2 - yy1)
-
-            area_i = max(0.0, float(
-                (boxes[i, 2] - boxes[i, 0]) * (boxes[i, 3] - boxes[i, 1])
-            ))
-            areas_j = (
-                np.maximum(0.0, boxes[i + 1:, 2] - boxes[i + 1:, 0])
-                * np.maximum(0.0, boxes[i + 1:, 3] - boxes[i + 1:, 1])
-            )
-            iou = inter / (area_i + areas_j - inter + 1e-7)
-            scores[i + 1:] *= np.exp(-(iou ** 2) / sigma)
-
-        mask = scores > score_thresh
-        return order[mask], scores[mask]
+        if self.input_dtype == np.float16:
+            blob = blob.astype(np.float16)
+        return blob, ratio, (float(dx), float(dy))
+
+    def _infer_single(self, image_bgr):
+        inp, ratio, (dx, dy) = self._preprocess(image_bgr)
+        out = self.session.run(self.output_names, {self.input_name: inp})[0]
+        if out.ndim == 3: out = out[0]
+        confs_all = out[:, 4].astype(np.float32)
+        cls_all = self.cls_remap[out[:, 5].astype(np.int32)]
+        cls_idx = np.clip(cls_all, 0, len(self.conf_thres_per_class) - 1)
+        keep = confs_all >= self.conf_thres_per_class[cls_idx]
+        if not keep.any(): return []
+        out = out[keep]
+        boxes = out[:, :4].astype(np.float32).copy()
+        confs = out[:, 4].astype(np.float32)
+        cls_ids = self.cls_remap[out[:, 5].astype(np.int32)]
+        boxes[:, [0, 2]] = (boxes[:, [0, 2]] - dx) / ratio
+        boxes[:, [1, 3]] = (boxes[:, [1, 3]] - dy) / ratio
+        oh, ow = image_bgr.shape[:2]
+        boxes[:, [0, 2]] = np.clip(boxes[:, [0, 2]], 0, ow - 1)
+        boxes[:, [1, 3]] = np.clip(boxes[:, [1, 3]], 0, oh - 1)
+        if len(boxes) > 1:
+            keep_idx = self._per_class_hard_nms(boxes, confs, cls_ids, self.iou_thresh)
+            keep_idx = keep_idx[: self.max_det]
+            boxes, confs, cls_ids = boxes[keep_idx], confs[keep_idx], cls_ids[keep_idx]
+            boxes, confs, cls_ids = self._cross_class_dedup(
+                boxes, confs, cls_ids, self.cross_iou_thresh)
+        return self._to_boundingboxes(boxes, confs, cls_ids, ow, oh)
 
     @staticmethod
-    def _hard_nms(
-        boxes: np.ndarray,
-        scores: np.ndarray,
-        iou_thresh: float,
-    ) -> np.ndarray:
-        """
-        Standard NMS: keep one box per overlapping cluster (the one with highest score).
-        Returns indices of kept boxes (into the boxes/scores arrays).
-        """
-        N = len(boxes)
-        if N == 0:
-            return np.array([], dtype=np.intp)
-        boxes = np.asarray(boxes, dtype=np.float32)
-        scores = np.asarray(scores, dtype=np.float32)
+    def _hard_nms(boxes, scores, iou_thresh):
+        n = len(boxes)
+        if n == 0: return np.array([], dtype=np.intp)
         order = np.argsort(scores)[::-1]
-        keep: list[int] = []
-        suppressed = np.zeros(N, dtype=bool)
-        for i in range(N):
+        keep, suppressed = [], np.zeros(n, dtype=bool)
+        for i in range(n):
             idx = order[i]
-            if suppressed[idx]:
-                continue
-            keep.append(idx)
+            if suppressed[idx]: continue
+            keep.append(int(idx))
             bi = boxes[idx]
-            for k in range(i + 1, N):
+            for k in range(i + 1, n):
                 jdx = order[k]
-                if suppressed[jdx]:
-                    continue
+                if suppressed[jdx]: continue
                 bj = boxes[jdx]
-                xx1 = max(bi[0], bj[0])
-                yy1 = max(bi[1], bj[1])
-                xx2 = min(bi[2], bj[2])
-                yy2 = min(bi[3], bj[3])
-                inter = max(0.0, xx2 - xx1) * max(0.0, yy2 - yy1)
-                area_i = (bi[2] - bi[0]) * (bi[3] - bi[1])
-                area_j = (bj[2] - bj[0]) * (bj[3] - bj[1])
-                iou = inter / (area_i + area_j - inter + 1e-7)
-                if iou > iou_thresh:
-                    suppressed[jdx] = True
-        return np.array(keep)
-
-    def _per_class_hard_nms(
-        self,
-        boxes: np.ndarray,
-        scores: np.ndarray,
-        cls_ids: np.ndarray,
-        iou_thresh: float,
-    ) -> np.ndarray:
-        """Hard NMS applied independently per class."""
-        if len(boxes) == 0:
-            return np.array([], dtype=np.intp)
-        all_keep: list[int] = []
+                xx1, yy1 = max(bi[0], bj[0]), max(bi[1], bj[1])
+                xx2, yy2 = min(bi[2], bj[2]), min(bi[3], bj[3])
+                inter = max(0.0, xx2-xx1) * max(0.0, yy2-yy1)
+                ai = (bi[2]-bi[0])*(bi[3]-bi[1]); aj = (bj[2]-bj[0])*(bj[3]-bj[1])
+                iou = inter / (ai + aj - inter + 1e-7)
+                if iou > iou_thresh: suppressed[jdx] = True
+        return np.array(keep, dtype=np.intp)
+
+    def _per_class_hard_nms(self, boxes, scores, cls_ids, iou_thresh):
+        if len(boxes) == 0: return np.array([], dtype=np.intp)
+        all_keep = []
         for c in np.unique(cls_ids):
             mask = cls_ids == c
             indices = np.where(mask)[0]
@@ -296,397 +184,52 @@ class Miner:
         all_keep.sort()
         return np.array(all_keep, dtype=np.intp)
 
-    def _per_class_soft_nms(
-        self,
-        boxes: np.ndarray,
-        scores: np.ndarray,
-        cls_ids: np.ndarray,
-        sigma: float = 0.5,
-        score_thresh: float = 0.01,
-    ) -> tuple[np.ndarray, np.ndarray]:
-        """Soft NMS applied independently per class."""
-        if len(boxes) == 0:
-            return np.array([], dtype=np.intp), np.array([], dtype=np.float32)
-        all_keep: list[int] = []
-        all_scores: list[float] = []
-        for c in np.unique(cls_ids):
-            mask = cls_ids == c
-            indices = np.where(mask)[0]
-            keep, updated = self._soft_nms(boxes[mask], scores[mask], sigma, score_thresh)
-            for k, s in zip(keep, updated):
-                all_keep.append(int(indices[k]))
-                all_scores.append(float(s))
-        if not all_keep:
-            return np.array([], dtype=np.intp), np.array([], dtype=np.float32)
-        return np.array(all_keep, dtype=np.intp), np.array(all_scores, dtype=np.float32)
-
-    def _filter_sane_boxes(
-        self,
-        boxes: np.ndarray,
-        scores: np.ndarray,
-        cls_ids: np.ndarray,
-        orig_size: tuple[int, int],
-    ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
-        """Filter out tiny, degenerate, or implausible boxes (common FP)."""
-        if len(boxes) == 0:
-            return boxes, scores, cls_ids
-        orig_w, orig_h = orig_size
-        image_area = float(orig_w * orig_h)
-        keep = []
-        for i, box in enumerate(boxes):
-            x1, y1, x2, y2 = box.tolist()
-            bw = x2 - x1
-            bh = y2 - y1
-            if bw <= 0 or bh <= 0:
-                continue
-            if bw < self.min_side or bh < self.min_side:
-                continue
-            area = bw * bh
-            if area < self.min_box_area:
-                continue
-            if area > 0.95 * image_area:
-                continue
-            ar = max(bw / max(bh, 1e-6), bh / max(bw, 1e-6))
-            if ar > self.max_aspect_ratio:
-                continue
-            keep.append(i)
-        if not keep:
-            return (
-                np.empty((0, 4), dtype=np.float32),
-                np.empty((0,), dtype=np.float32),
-                np.empty((0,), dtype=np.int32),
-            )
-        k = np.array(keep, dtype=np.intp)
-        return boxes[k], scores[k], cls_ids[k]
-
     @staticmethod
-    def _max_score_per_cluster(
-        coords: np.ndarray,
-        scores: np.ndarray,
-        keep_indices: np.ndarray,
-        iou_thresh: float,
-    ) -> np.ndarray:
-        """
-        For each kept box, return the max original score among itself and any
-        box that overlaps it with IOU >= iou_thresh (so TTA cluster keeps best conf).
-        """
-        n_keep = len(keep_indices)
-        if n_keep == 0:
-            return np.array([], dtype=np.float32)
-        out = np.empty(n_keep, dtype=np.float32)
-        coords = np.asarray(coords, dtype=np.float32)
-        scores = np.asarray(scores, dtype=np.float32)
-        for i in range(n_keep):
-            idx = keep_indices[i]
-            bi = coords[idx]
-            xx1 = np.maximum(bi[0], coords[:, 0])
-            yy1 = np.maximum(bi[1], coords[:, 1])
-            xx2 = np.minimum(bi[2], coords[:, 2])
-            yy2 = np.minimum(bi[3], coords[:, 3])
-            inter = np.maximum(0.0, xx2 - xx1) * np.maximum(0.0, yy2 - yy1)
-            area_i = (bi[2] - bi[0]) * (bi[3] - bi[1])
-            areas_j = (coords[:, 2] - coords[:, 0]) * (coords[:, 3] - coords[:, 1])
-            iou = inter / (area_i + areas_j - inter + 1e-7)
-            in_cluster = iou >= iou_thresh
-            out[i] = float(np.max(scores[in_cluster]))
+    def _cross_class_dedup(boxes, scores, cls_ids, iou_thresh):
+        n = len(boxes)
+        if n <= 1: return boxes, scores, cls_ids
+        areas = np.maximum(0.0, boxes[:, 2]-boxes[:, 0]) * np.maximum(0.0, boxes[:, 3]-boxes[:, 1])
+        order = np.lexsort((-scores, -areas))
+        suppressed = np.zeros(n, dtype=bool); keep = []
+        for i in order:
+            if suppressed[i]: continue
+            keep.append(int(i))
+            bi = boxes[i]
+            xx1 = np.maximum(bi[0], boxes[:, 0]); yy1 = np.maximum(bi[1], boxes[:, 1])
+            xx2 = np.minimum(bi[2], boxes[:, 2]); yy2 = np.minimum(bi[3], boxes[:, 3])
+            inter = np.maximum(0.0, xx2-xx1) * np.maximum(0.0, yy2-yy1)
+            ai = max(1e-7, float((bi[2]-bi[0])*(bi[3]-bi[1])))
+            iou = inter / (ai + areas - inter + 1e-7)
+            dup = iou > iou_thresh; dup[i] = False
+            suppressed |= dup
+        kept = np.array(keep, dtype=np.intp)
+        return boxes[kept], scores[kept], cls_ids[kept]
+
+    def _to_boundingboxes(self, boxes, confs, cls_ids, orig_w, orig_h):
+        out = []
+        for i in range(len(boxes)):
+            x1, y1, x2, y2 = boxes[i]
+            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)))
+            if ix2 <= ix1 or iy2 <= iy1: continue
+            bw, bh = ix2 - ix1, iy2 - iy1
+            if bw * bh < self.min_box_area: continue
+            if min(bw, bh) < self.min_side: continue
+            ar = max(bw / max(bh, 1), bh / max(bw, 1))
+            if ar > self.max_aspect_ratio: continue
+            out.append(BoundingBox(x1=ix1, y1=iy1, x2=ix2, y2=iy2, cls_id=int(cls_ids[i]),
+                                   conf=max(0.0, min(1.0, float(confs[i])))))
         return out
 
-    def _decode_final_dets(
-        self,
-        preds: np.ndarray,
-        ratio: float,
-        pad: tuple[float, float],
-        orig_size: tuple[int, int],
-        apply_optional_dedup: bool = False,
-    ) -> list[BoundingBox]:
-        """
-        Primary path:
-        expected output rows like [x1, y1, x2, y2, conf, cls_id]
-        in letterboxed input coordinates.
-        """
-        if preds.ndim == 3 and preds.shape[0] == 1:
-            preds = preds[0]
-
-        if preds.ndim != 2 or preds.shape[1] < 6:
-            raise ValueError(f"Unexpected ONNX final-det output shape: {preds.shape}")
-
-        boxes = preds[:, :4].astype(np.float32)
-        scores = preds[:, 4].astype(np.float32)
-        cls_ids = preds[:, 5].astype(np.int32)
-        cls_ids = self.cls_remap[cls_ids]
-
-        keep = scores >= self.conf_thres
-        boxes = boxes[keep]
-        scores = scores[keep]
-        cls_ids = cls_ids[keep]
-
-        if len(boxes) == 0:
-            return []
-
-        pad_w, pad_h = pad
-        orig_w, orig_h = orig_size
-
-        # reverse letterbox
-        boxes[:, [0, 2]] -= pad_w
-        boxes[:, [1, 3]] -= pad_h
-        boxes /= ratio
-        boxes = self._clip_boxes(boxes, (orig_w, orig_h))
-
-        # Box sanity filter (reduces FP)
-        boxes, scores, cls_ids = self._filter_sane_boxes(
-            boxes, scores, cls_ids, orig_size
-        )
-        if len(boxes) == 0:
-            return []
-
-        # Per-class NMS to remove duplicates without suppressing across classes
-        if len(boxes) > 1:
-            if apply_optional_dedup:
-                keep_idx, scores = self._per_class_soft_nms(boxes, scores, cls_ids)
-                boxes = boxes[keep_idx]
-                cls_ids = cls_ids[keep_idx]
-            else:
-                keep_idx = self._per_class_hard_nms(boxes, scores, cls_ids, self.iou_thres)
-                keep_idx = keep_idx[: self.max_det]
-                boxes = boxes[keep_idx]
-                scores = scores[keep_idx]
-                cls_ids = cls_ids[keep_idx]
-
-        results: list[BoundingBox] = []
-        for box, conf, cls_id in zip(boxes, scores, cls_ids):
-            x1, y1, x2, y2 = box.tolist()
-
-            if x2 <= x1 or y2 <= y1:
-                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=int(cls_id),
-                    conf=float(conf),
-                )
-            )
-
-        return results
-
-    def _decode_raw_yolo(
-        self,
-        preds: np.ndarray,
-        ratio: float,
-        pad: tuple[float, float],
-        orig_size: tuple[int, int],
-    ) -> list[BoundingBox]:
-        """
-        Fallback path for raw YOLO predictions.
-        Supports common layouts:
-        - [1, C, N]
-        - [1, N, C]
-        """
-        if preds.ndim != 3:
-            raise ValueError(f"Unexpected raw ONNX output shape: {preds.shape}")
-
-        if preds.shape[0] != 1:
-            raise ValueError(f"Unexpected batch dimension in raw output: {preds.shape}")
-
-        preds = preds[0]
-
-        # Normalize to [N, C]
-        if preds.shape[0] <= 16 and preds.shape[1] > preds.shape[0]:
-            preds = preds.T
-
-        if preds.ndim != 2 or preds.shape[1] < 5:
-            raise ValueError(f"Unexpected normalized raw output shape: {preds.shape}")
-
-        boxes_xywh = preds[:, :4].astype(np.float32)
-        cls_part = preds[:, 4:].astype(np.float32)
-
-        if cls_part.shape[1] == 1:
-            scores = cls_part[:, 0]
-            cls_ids = np.zeros(len(scores), dtype=np.int32)
-        else:
-            cls_ids = np.argmax(cls_part, axis=1).astype(np.int32)
-            scores = cls_part[np.arange(len(cls_part)), cls_ids]
-        cls_ids = self.cls_remap[cls_ids]
-
-        keep = scores >= self.conf_thres
-        boxes_xywh = boxes_xywh[keep]
-        scores = scores[keep]
-        cls_ids = cls_ids[keep]
-
-        if len(boxes_xywh) == 0:
-            return []
-
-        boxes = self._xywh_to_xyxy(boxes_xywh)
-
-        keep_idx = self._per_class_hard_nms(boxes, scores, cls_ids, self.iou_thres)
-        keep_idx = keep_idx[: self.max_det]
-        boxes = boxes[keep_idx]
-        scores = scores[keep_idx]
-        cls_ids = cls_ids[keep_idx]
-
-        pad_w, pad_h = pad
-        orig_w, orig_h = orig_size
-
-        boxes[:, [0, 2]] -= pad_w
-        boxes[:, [1, 3]] -= pad_h
-        boxes /= ratio
-        boxes = self._clip_boxes(boxes, (orig_w, orig_h))
-
-        boxes, scores, cls_ids = self._filter_sane_boxes(
-            boxes, scores, cls_ids, (orig_w, orig_h)
-        )
-        if len(boxes) == 0:
-            return []
-
-        results: list[BoundingBox] = []
-        for box, conf, cls_id in zip(boxes, scores, cls_ids):
-            x1, y1, x2, y2 = box.tolist()
-
-            if x2 <= x1 or y2 <= y1:
-                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=int(cls_id),
-                    conf=float(conf),
-                )
-            )
-
+    def predict_batch(self, batch_images, offset, n_keypoints):
+        results = []
+        for idx, image in enumerate(batch_images):
+            boxes = self._infer_single(image)
+            results.append(TVFrameResult(
+                frame_id=offset + idx,
+                boxes=boxes,
+                keypoints=[(0, 0) for _ in range(max(0, int(n_keypoints)))],
+            ))
         return results
-
-    def _postprocess(
-        self,
-        output: np.ndarray,
-        ratio: float,
-        pad: tuple[float, float],
-        orig_size: tuple[int, int],
-    ) -> list[BoundingBox]:
-        """
-        Prefer final detections first.
-        Fallback to raw decode only if needed.
-        """
-        # final detections: [N,6]
-        if output.ndim == 2 and output.shape[1] >= 6:
-            return self._decode_final_dets(output, ratio, pad, orig_size)
-
-        # final detections: [1,N,6]
-        if output.ndim == 3 and output.shape[0] == 1 and output.shape[2] == 6:
-            return self._decode_final_dets(output, ratio, pad, orig_size)
-
-        # fallback raw decode
-        return self._decode_raw_yolo(output, ratio, pad, orig_size)
-
-    def _predict_single(self, image: np.ndarray) -> list[BoundingBox]:
-        if image is None:
-            raise ValueError("Input image is None")
-        if not isinstance(image, np.ndarray):
-            raise TypeError(f"Input is not numpy array: {type(image)}")
-        if image.ndim != 3:
-            raise ValueError(f"Expected HWC image, got shape={image.shape}")
-        if image.shape[0] <= 0 or image.shape[1] <= 0:
-            raise ValueError(f"Invalid image shape={image.shape}")
-        if image.shape[2] != 3:
-            raise ValueError(f"Expected 3 channels, got shape={image.shape}")
-
-        if image.dtype != np.uint8:
-            image = image.astype(np.uint8)
-
-        input_tensor, ratio, pad, orig_size = self._preprocess(image)
-
-        expected_shape = (1, 3, self.input_height, self.input_width)
-        if input_tensor.shape != expected_shape:
-            raise ValueError(
-                f"Bad input tensor shape={input_tensor.shape}, expected={expected_shape}"
-            )
-
-        outputs = self.session.run(self.output_names, {self.input_name: input_tensor})
-        det_output = outputs[0]
-        return self._postprocess(det_output, ratio, pad, orig_size)
-
-    def _predict_tta(self, image: np.ndarray) -> list[BoundingBox]:
-        """
-        Horizontal-flip TTA: merge original + flipped via hard NMS.
-        Boost confidence for consensus detections (both views agree) to improve
-        mAP: validator sorts by confidence, so higher conf for TP helps PR curve.
-        """
-        boxes_orig = self._predict_single(image)
-
-        flipped = cv2.flip(image, 1)
-        boxes_flip = self._predict_single(flipped)
-
-        w = image.shape[1]
-        boxes_flip = [
-            BoundingBox(
-                x1=w - b.x2, y1=b.y1, x2=w - b.x1, y2=b.y2,
-                cls_id=b.cls_id, conf=b.conf,
-            )
-            for b in boxes_flip
-        ]
-
-        all_boxes = boxes_orig + boxes_flip
-        if len(all_boxes) == 0:
-            return []
-
-        coords = np.array(
-            [[b.x1, b.y1, b.x2, b.y2] for b in all_boxes], dtype=np.float32
-        )
-        scores = np.array([b.conf for b in all_boxes], dtype=np.float32)
-        cls_ids = np.array([b.cls_id for b in all_boxes], dtype=np.int32)
-
-        hard_keep = self._per_class_hard_nms(coords, scores, cls_ids, self.iou_thres)
-        if len(hard_keep) == 0:
-            return []
-
-        hard_keep = hard_keep[: self.max_det]
-
-        # Boost confidence when both views agree (overlapping detections)
-        boosted = self._max_score_per_cluster(
-            coords, scores, hard_keep, self.iou_thres
-        )
-
-        return [
-            BoundingBox(
-                x1=all_boxes[i].x1,
-                y1=all_boxes[i].y1,
-                x2=all_boxes[i].x2,
-                y2=all_boxes[i].y2,
-                cls_id=all_boxes[i].cls_id,
-                conf=float(boosted[j]),
-            )
-            for j, i in enumerate(hard_keep)
-        ]
-
-    def predict_batch(
-        self,
-        batch_images: list[ndarray],
-        offset: int,
-        n_keypoints: int,
-    ) -> list[TVFrameResult]:
-        results: list[TVFrameResult] = []
-
-        for frame_number_in_batch, image in enumerate(batch_images):
-            try:
-                if self.use_tta:
-                    boxes = self._predict_tta(image)
-                else:
-                    boxes = self._predict_single(image)
-            except Exception as e:
-                print(f"⚠️ Inference failed for frame {offset + frame_number_in_batch}: {e}")
-                boxes = []
-
-            results.append(
-                TVFrameResult(
-                    frame_id=offset + frame_number_in_batch,
-                    boxes=boxes,
-                    keypoints=[(0, 0) for _ in range(max(0, int(n_keypoints)))],
-                )
-            )
-
-        return results

model_type.json

@@ -0,0 +1,4 @@
+{
+  "task_type": "object-detection",
+  "model_type": "yolov26-nano"
+}

weights.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2925713fc7f8d25c56fabca8ccadfcf8acecb53561530d6042a7b8d5b424b1d4
-size 19724723
+oid sha256:0bfd3fd0b1dca617b05f93fb1ce92aadc8f6ee8e80255c2eb0818b143b4056d6
+size 5077018