subnet_bridge: copy winning miner repo into library

README.md

@@ -10,8 +10,17 @@ tags:
 manako:
   source: winner_fetch
   manifest_element_name: manak0/Detect-fire
-  winner_repo_id: navierstocks/wash
-  winner_revision: f0cb290789b770cdd918dc49c0c219c3c01bc70b
+  winner_repo_id: SuperBitDev/fire1
+  winner_revision: 0fbc341ae743ebec0d2a1d48bdaaaaa0d6ad9338
 ---
 
-## YOLO26 ONNX detector
+# Detect-fire-winner
+
+Published winning miner converted into a library element.
+
+- Source winner repo: `SuperBitDev/fire1`
+- Source revision: `0fbc341ae743ebec0d2a1d48bdaaaaa0d6ad9338`
+- Manifest element: `manak0/Detect-fire`
+- Element type: `Detect`
+- Objects: fire, smoke, fire extinguisher
+- Runtime model type: `onnxruntime`

chute_config.yml

@@ -2,13 +2,12 @@ Image:
   from_base: parachutes/python:3.12
   run_command:
     - pip install --upgrade setuptools wheel
-    - pip install 'numpy>=1.23' 'onnxruntime-gpu[cuda,cudnn]>=1.16' 'opencv-python>=4.7' 'pillow>=9.5' 'huggingface_hub>=0.19.4' 'pydantic>=2.0' 'pyyaml>=6.0' 'aiohttp>=3.9'
-    - pip install torch torchvision
+    - pip install 'numpy>=1.23' 'onnxruntime-gpu>=1.16' 'opencv-python>=4.7' 'pillow>=9.5' 'huggingface_hub>=0.19.4' 'pydantic>=2.0' 'pyyaml>=6.0' 'aiohttp>=3.9'
+    - pip install torch==2.8.0 torchvision==0.23.0 torchaudio==2.8.0 --index-url https://download.pytorch.org/whl/cu128
 
 NodeSelector:
   gpu_count: 1
   min_vram_gb_per_gpu: 16
-  max_hourly_price_per_gpu: 2
   include:
     - pro_6000
 

miner.py

@@ -24,25 +24,56 @@ class TVFrameResult(BaseModel):
 
 
 class Miner:
-    """ONNX Runtime miner. Hard per-class NMS + cross-class dedup + flip TTA, with per-class conf bonus rescue."""
+    """ONNX Runtime miner for fire / smoke / fire_extinguisher detection.
 
-    class_names = ["broom", "drainage gate", "nozzle", "track"]
-    input_size = 1280
-    iou_thres = 0.4
+    Strategy (ported from offense miner):
+      - per-class confidence threshold with per-class rescue bonus
+      - per-class hard NMS, then cross-class dedup
+      - horizontal-flip TTA with full-set cluster score boost
+    Plus fire001 specifics: class remap, sanity-box filter, TTA toggle.
+    """
+
+    class_names = ["fire", "smoke", "fire extinguisher"]
+    # Order the model emits classes in -- remapped to `class_names` index.
+    _model_class_order = ["fire", "fire extinguisher", "smoke"]
+
+    iou_thres = 0.55
     cross_iou_thresh = 0.8
-    max_det = 300
-    _conf_thres_array = np.array([0.55, 0.6, 0.7, 0.5], dtype=np.float32)
-    _bonus_array = np.array([0.05, 0.5, 0.0, 0.2], dtype=np.float32)
+    max_det = 150
+
+    # Per-class confidence thresholds. Higher = fewer FP for that class.
+    # Indexed by class_names order: [fire, smoke, fire_extinguisher].
+    _conf_thres_array = np.array(
+        [0.6, 0.4, 0.3], dtype=np.float32
+    )
+    # Per-class rescue bonus. If a class has ZERO boxes passing the threshold
+    # in a frame, its top-1 candidate is admitted when its score is at least
+    # (threshold - bonus). Fire and smoke get a small bonus (variable
+    # appearance); fire extinguisher does not (distinctive object, leave FP
+    # control strict).
+    _bonus_array = np.array(
+        [0, 0.1, 0.15], dtype=np.float32
+    )
+
+    # Box sanity filter (fire001-specific FP reduction): drop tiny / degenerate
+    # / image-spanning / extreme aspect ratio boxes.
+    min_box_area = 14 * 14
+    min_side = 8
+    max_aspect_ratio = 8.0
 
     def __init__(self, path_hf_repo: Path) -> None:
         model_path = path_hf_repo / "weights.onnx"
+        self.cls_remap = np.array(
+            [self.class_names.index(n) for n in self._model_class_order],
+            dtype=np.int32,
+        )
         print("ORT version:", ort.__version__)
 
         try:
             ort.preload_dlls()
-            print("preload_dlls success")
+            print("✅ onnxruntime.preload_dlls() success")
         except Exception as e:
-            print(f"preload_dlls failed: {e}")
+            print(f"⚠️ preload_dlls failed: {e}")
 
         print("ORT available providers BEFORE session:", ort.get_available_providers())
 
@@ -55,9 +86,9 @@ class Miner:
                 sess_options=sess_options,
                 providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
             )
-            print("Created ORT session with preferred CUDA provider list")
+            print("✅ Created ORT session with preferred CUDA provider list")
         except Exception as e:
-            print(f"CUDA session creation failed, falling back to CPU: {e}")
+            print(f"⚠️ CUDA session creation failed, falling back to CPU: {e}")
             self.session = ort.InferenceSession(
                 str(model_path),
                 sess_options=sess_options,
@@ -75,15 +106,19 @@ class Miner:
         self.output_names = [output.name for output in self.session.get_outputs()]
         self.input_shape = self.session.get_inputs()[0].shape
 
-        self.input_height = self._safe_dim(self.input_shape[2], default=self.input_size)
-        self.input_width = self._safe_dim(self.input_shape[3], default=self.input_size)
+        self.input_height = self._safe_dim(self.input_shape[2], default=1280)
+        self.input_width = self._safe_dim(self.input_shape[3], default=1280)
 
-        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}")
+        self.use_tta = True
+
+        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}")
         print("per-class conf: " + ", ".join(
-            f"{n}={t:.3f}" for n, t in zip(self.class_names,
-                                           self._conf_thres_array.tolist())))
+            f"{n}={t:.3f}" for n, t in zip(
+                self.class_names, self._conf_thres_array.tolist()
+            )
+        ))
 
     def __repr__(self) -> str:
         return (
@@ -95,32 +130,44 @@ class Miner:
     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]]:
+    def _letterbox(
+        self,
+        image: ndarray,
+        new_shape: tuple[int, int],
+        color=(114, 114, 114),
+    ) -> tuple[ndarray, float, tuple[float, float]]:
         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):
             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) / 2.0
         dh = (new_h - resized_h) / 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)
+
+        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]]:
+    def _preprocess(
+        self, image: ndarray
+    ) -> tuple[np.ndarray, float, tuple[float, float], tuple[int, int]]:
         orig_h, orig_w = image.shape[:2]
-        img, ratio, pad = self._letterbox(image, (self.input_width, self.input_height))
+        img, ratio, pad = self._letterbox(
+            image, (self.input_width, self.input_height)
+        )
         img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
         img = img.astype(np.float32) / 255.0
         img = np.transpose(img, (2, 0, 1))[None, ...]
@@ -146,8 +193,9 @@ class Miner:
         return out
 
     @staticmethod
-    def _hard_nms(boxes: np.ndarray, scores: np.ndarray,
-                  iou_thresh: float) -> np.ndarray:
+    def _hard_nms(
+        boxes: np.ndarray, scores: np.ndarray, iou_thresh: float
+    ) -> np.ndarray:
         n = len(boxes)
         if n == 0:
             return np.array([], dtype=np.intp)
@@ -172,9 +220,13 @@ class Miner:
             order = rest[iou <= iou_thresh]
         return np.array(keep, dtype=np.intp)
 
-    def _per_class_hard_nms(self, boxes: np.ndarray, scores: np.ndarray,
-                            cls_ids: np.ndarray, iou_thresh: float
-                            ) -> np.ndarray:
+    def _per_class_hard_nms(
+        self,
+        boxes: np.ndarray,
+        scores: np.ndarray,
+        cls_ids: np.ndarray,
+        iou_thresh: float,
+    ) -> np.ndarray:
         if len(boxes) == 0:
             return np.array([], dtype=np.intp)
         all_keep: list[int] = []
@@ -186,9 +238,20 @@ class Miner:
         all_keep.sort()
         return np.array(all_keep, dtype=np.intp)
 
-    def _cross_class_dedup_op(self, boxes: np.ndarray, scores: np.ndarray,
-                              cls_ids: np.ndarray, iou_thresh: float
-                              ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
+    def _cross_class_dedup_op(
+        self,
+        boxes: np.ndarray,
+        scores: np.ndarray,
+        cls_ids: np.ndarray,
+        iou_thresh: float,
+    ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
+        """Remove near-duplicate boxes across classes.
+
+        Order candidates by (score - per_class_threshold) margin, then by area;
+        keep the highest, suppress every other box with IoU > iou_thresh.
+        This suppresses the case where the same physical object is detected
+        as multiple classes (e.g. fire vs smoke on the same flames).
+        """
         n = len(boxes)
         if n <= 1:
             return boxes, scores, cls_ids
@@ -220,12 +283,20 @@ class Miner:
         return boxes[keep_idx], scores[keep_idx], cls_ids[keep_idx]
 
     @staticmethod
-    def _max_score_per_cluster(post_boxes: np.ndarray,
-                               post_cls: np.ndarray,
-                               full_boxes: np.ndarray,
-                               full_scores: np.ndarray,
-                               full_cls: np.ndarray,
-                               iou_thresh: float) -> np.ndarray:
+    def _max_score_per_cluster(
+        post_boxes: np.ndarray,
+        post_cls: np.ndarray,
+        full_boxes: np.ndarray,
+        full_scores: np.ndarray,
+        full_cls: np.ndarray,
+        iou_thresh: float,
+    ) -> np.ndarray:
+        """For each kept (post-NMS) box, return the max score over the FULL
+        candidate set among same-class boxes with IoU >= iou_thresh.
+
+        Used after horizontal-flip TTA: a high-confidence flipped detection
+        can raise the score of the corresponding original detection.
+        """
         n = len(post_boxes)
         if n == 0:
             return np.empty(0, dtype=np.float32)
@@ -245,10 +316,11 @@ class Miner:
             out[i] = float(np.max(full_scores[cluster])) if np.any(cluster) else 0.0
         return out
 
-    def _conf_filter_mask(self, scores: np.ndarray,
-                          cls_ids: np.ndarray) -> np.ndarray:
+    def _conf_filter_mask(
+        self, scores: np.ndarray, cls_ids: np.ndarray
+    ) -> np.ndarray:
         """Boolean keep-mask: score >= per-class threshold, with a per-class
-        rescue — if a class has zero boxes passing, admit its top-1 candidate
+        rescue -- if a class has zero boxes passing, admit its top-1 candidate
         when its score >= (per-class threshold - per-class bonus)."""
         if len(scores) == 0:
             return np.zeros(0, dtype=bool)
@@ -267,9 +339,52 @@ class Miner:
                 keep[top] = True
         return keep
 
-    def _per_view_pipeline(self, boxes: np.ndarray, scores: np.ndarray,
-                           cls_ids: np.ndarray
-                           ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
+    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]:
+        """Drop tiny / degenerate / image-spanning / extreme-AR boxes (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]
+
+    def _per_view_pipeline(
+        self,
+        boxes: np.ndarray,
+        scores: np.ndarray,
+        cls_ids: np.ndarray,
+    ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
+        """Per-view post-processing pipeline: per-class NMS -> cap -> cross-class dedup."""
         if len(boxes) > 1:
             keep = self._per_class_hard_nms(boxes, scores, cls_ids, self.iou_thres)
             boxes, scores, cls_ids = boxes[keep], scores[keep], cls_ids[keep]
@@ -282,9 +397,35 @@ class Miner:
             )
         return boxes, scores, cls_ids
 
-    def _decode_final_dets(self, preds: np.ndarray, ratio: float,
-                           pad: tuple[float, float],
-                           orig_size: tuple[int, int]) -> list[BoundingBox]:
+    @staticmethod
+    def _build_results(
+        boxes: np.ndarray, scores: np.ndarray, cls_ids: np.ndarray
+    ) -> list[BoundingBox]:
+        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_final_dets(
+        self,
+        preds: np.ndarray,
+        ratio: float,
+        pad: tuple[float, float],
+        orig_size: tuple[int, int],
+    ) -> list[BoundingBox]:
+        """Final-detection output path: rows shaped [x1, y1, x2, y2, conf, cls_id]."""
         if preds.ndim == 3 and preds.shape[0] == 1:
             preds = preds[0]
         if preds.ndim != 2 or preds.shape[1] < 6:
@@ -293,6 +434,7 @@ class Miner:
         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 = self._conf_filter_mask(scores, cls_ids)
         boxes = boxes[keep]
@@ -307,12 +449,23 @@ class Miner:
         boxes /= ratio
         boxes = self._clip_boxes(boxes, orig_size)
 
+        boxes, scores, cls_ids = self._filter_sane_boxes(
+            boxes, scores, cls_ids, orig_size
+        )
+        if len(boxes) == 0:
+            return []
+
         boxes, scores, cls_ids = self._per_view_pipeline(boxes, scores, cls_ids)
         return self._build_results(boxes, scores, cls_ids)
 
-    def _decode_raw_yolo(self, preds: np.ndarray, ratio: float,
-                         pad: tuple[float, float],
-                         orig_size: tuple[int, int]) -> list[BoundingBox]:
+    def _decode_raw_yolo(
+        self,
+        preds: np.ndarray,
+        ratio: float,
+        pad: tuple[float, float],
+        orig_size: tuple[int, int],
+    ) -> list[BoundingBox]:
+        """Fallback raw-YOLO output path: per-anchor class logits."""
         if preds.ndim != 3 or preds.shape[0] != 1:
             raise ValueError(f"Unexpected raw ONNX output shape: {preds.shape}")
         preds = preds[0]
@@ -329,6 +482,7 @@ class Miner:
         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 = self._conf_filter_mask(scores, cls_ids)
         boxes_xywh = boxes_xywh[keep]
@@ -344,32 +498,22 @@ class Miner:
         boxes /= ratio
         boxes = self._clip_boxes(boxes, orig_size)
 
+        boxes, scores, cls_ids = self._filter_sane_boxes(
+            boxes, scores, cls_ids, orig_size
+        )
+        if len(boxes) == 0:
+            return []
+
         boxes, scores, cls_ids = self._per_view_pipeline(boxes, scores, cls_ids)
         return self._build_results(boxes, scores, cls_ids)
 
-    @staticmethod
-    def _build_results(boxes: np.ndarray, scores: np.ndarray,
-                       cls_ids: np.ndarray) -> list[BoundingBox]:
-        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 _postprocess(self, output: np.ndarray, ratio: float,
-                     pad: tuple[float, float],
-                     orig_size: tuple[int, int]) -> list[BoundingBox]:
+    def _postprocess(
+        self,
+        output: np.ndarray,
+        ratio: float,
+        pad: tuple[float, float],
+        orig_size: tuple[int, int],
+    ) -> list[BoundingBox]:
         if output.ndim == 2 and output.shape[1] >= 6:
             return self._decode_final_dets(output, ratio, pad, orig_size)
         if output.ndim == 3 and output.shape[0] == 1 and output.shape[2] == 6:
@@ -383,6 +527,8 @@ class Miner:
             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:
@@ -399,6 +545,17 @@ class Miner:
         return self._postprocess(outputs[0], ratio, pad, orig_size)
 
     def _predict_tta(self, image: np.ndarray) -> list[BoundingBox]:
+        """Horizontal-flip TTA.
+
+        Strategy:
+          1. Predict on original and on flipped image.
+          2. Map flipped boxes back to original coordinates.
+          3. Per-class hard NMS on the union.
+          4. For each kept box, compute the max same-class score across the
+             FULL union (not just the post-NMS subset) -- this lets a high-
+             confidence flipped detection raise a borderline original one.
+          5. Cross-class dedup to suppress same-physical-object multi-class.
+        """
         boxes_orig = self._predict_single(image)
         flipped = cv2.flip(image, 1)
         boxes_flip = self._predict_single(flipped)
@@ -426,6 +583,7 @@ class Miner:
         if len(hard_keep) > self.max_det:
             top = np.argsort(-scores[hard_keep])[: self.max_det]
             hard_keep = hard_keep[top]
+
         boosted = self._max_score_per_cluster(
             coords[hard_keep], cls_ids[hard_keep],
             coords, scores, cls_ids, self.iou_thres,
@@ -450,14 +608,24 @@ class Miner:
             for j in range(len(kept_coords))
         ]
 
-    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: list[TVFrameResult] = []
         for frame_number_in_batch, image in enumerate(batch_images):
             try:
-                boxes = self._predict_tta(image)
+                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}")
+                print(
+                    f"⚠️ Inference failed for frame "
+                    f"{offset + frame_number_in_batch}: {e}"
+                )
                 boxes = []
             results.append(
                 TVFrameResult(

pyproject.toml

@@ -5,13 +5,14 @@ requires-python = ">=3.9"
 
 dependencies = [
   "numpy>=1.23",
-  "onnxruntime[cuda,cudnn]>=1.16",
+  "onnxruntime>=1.16",
   "opencv-python>=4.7",
   "pillow>=9.5",
   "huggingface_hub>=0.19.4",
   "pydantic>=2.0",
   "pyyaml>=6.0",
   "aiohttp>=3.9",
-  "torch",
-  "torchvision",
+  "torch==2.8.0",
+  "torchvision==0.23.0",
+  "torchaudio==2.8.0",
 ]

readme.md

@@ -1,17 +0,0 @@
----
-tags:
-- element_type:detect
-- model:onnxruntime
-- subnet:winner
-- object:fire
-- object:smoke
-- object:fire extinguisher
-
-manako:
-  source: winner_fetch
-  manifest_element_name: manak0/Detect-fire
-  winner_repo_id: navierstocks/wash
-  winner_revision: f0cb290789b770cdd918dc49c0c219c3c01bc70b
----
-
-## YOLO26 ONNX detector

weights.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e576f1b4805a0b6f026626e781859a256293f14a0faec0f81b26a550919d2911
-size 19300989
+oid sha256:df7c65835fbffb2923b4b76531d5a343544ffd6018cdca67859f830f27d74b09
+size 19407317