Upload folder using huggingface_hub

chute_config.yml

@@ -0,0 +1,20 @@
+Image:
+  from_base: parachutes/python:3.12
+  run_command:
+    - pip install --upgrade setuptools wheel
+    - pip install 'numpy>=1.23' 'onnxruntime-gpu==1.20.1' '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
+  include:
+    - pro_6000
+
+Chute:
+  timeout_seconds: 900
+  concurrency: 4
+  max_instances: 5
+  scaling_threshold: 0.5
+  shutdown_after_seconds: 288000
+  tee: true

miner.py

@@ -0,0 +1,566 @@
+from pathlib import Path
+import math
+
+import cv2
+import numpy as np
+import onnxruntime as ort
+from numpy import ndarray
+from pydantic import BaseModel
+
+
+class BoundingBox(BaseModel):
+    x1: int
+    y1: int
+    x2: int
+    y2: int
+    cls_id: int
+    conf: float
+
+
+class TVFrameResult(BaseModel):
+    frame_id: int
+    boxes: list[BoundingBox]
+    keypoints: list[tuple[int, int]]
+
+
+class Miner:
+    """ONNX Runtime miner. Hard per-class NMS + cross-class dedup + flip TTA."""
+
+    class_names = ["balaclava", "hoodie", "glove", "bat", "spray paint", "graffiti"]
+
+    # FALLBACK order the model emits classes in -- remapped to `class_names`
+    # index by `self.cls_remap` (built in __init__). The authoritative order is
+    # read from the ONNX `names` metadata that Ultralytics embeds at export time
+    # (ships inside weights.onnx), so a retrained model with a different class
+    # order is remapped correctly without code changes. This static list is used
+    # only when that metadata is missing or unparsable.
+    model_class_names = ["balaclava", "bat", "glove", "graffiti", "hoodie", "spray paint"]
+
+    input_size = 1280
+
+    # Test-time augmentation (horizontal-flip ensemble) runs a SECOND forward
+    # pass per frame and roughly DOUBLES latency. This 640 model is built for a
+    # <100 ms single-pass budget, so TTA is OFF by default. Turn it on only with
+    # latency headroom — and note the per-class thresholds below should be
+    # re-swept for whichever mode you deploy, since flipping TTA shifts scores.
+    use_tta = False
+
+    iou_thres = 0.3
+    cross_iou_thresh = 0.8
+    max_det = 150
+    
+    _conf_thres_array = np.array(
+        [0.38, 0.38, 0.22, 0.22, 0.33, 0.20], dtype=np.float32,
+    )
+    _bonus_array = np.array(
+        [0.25, 0.25, 0.12, 0.09, 0.21, 0.06], dtype=np.float32,
+    )
+    
+    def __init__(self, path_hf_repo: Path) -> None:
+        # model_path = path_hf_repo / "weights-all.onnx"
+        model_path = path_hf_repo / "weights.onnx"
+        print("ORT version:", ort.__version__)
+
+        try:
+            ort.preload_dlls()
+            print("preload_dlls success")
+        except Exception as e:
+            print(f"preload_dlls failed: {e}")
+
+        print("ORT available providers BEFORE session:", ort.get_available_providers())
+
+        sess_options = ort.SessionOptions()
+        sess_options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
+        # Pin threads for the 2vCPU/4GB public-track latency gate (p95 <= 100ms).
+        sess_options.intra_op_num_threads = 2
+        sess_options.inter_op_num_threads = 1
+        sess_options.execution_mode = ort.ExecutionMode.ORT_SEQUENTIAL
+
+        try:
+            self.session = ort.InferenceSession(
+                str(model_path),
+                sess_options=sess_options,
+                providers=["CPUExecutionProvider"],
+            )            
+        except Exception as e:
+            print(f"CUDA session creation failed, falling back to CPU: {e}")
+            self.session = ort.InferenceSession(
+                str(model_path),
+                sess_options=sess_options,
+                providers=["CPUExecutionProvider"],
+            )
+
+        print("ORT session providers:", self.session.get_providers())
+
+        # Build cls_remap: for each model-emit index i,
+        #   cls_remap[i] = self.class_names.index(model_class_order[i])
+        # i.e. convert a model-side class id into the output class id that
+        # downstream code (BoundingBox.cls_id, the per-class threshold/bonus
+        # arrays) expects. The model-side order comes from the ONNX metadata
+        # when available, else falls back to the static model_class_names.
+        model_class_order = self._read_model_class_order()
+        if model_class_order is None:
+            model_class_order = list(self.model_class_names)
+            print(f"cls order: no usable ONNX metadata, FALLBACK {model_class_order}")
+        else:
+            print(f"cls order: from ONNX metadata {model_class_order}")
+        self.cls_remap = np.array(
+            [self.class_names.index(n) for n in model_class_order],
+            dtype=np.int32,
+        )
+
+        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
+
+        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)
+
+        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(f"ONNX input size: {self.input_width}x{self.input_height}, use_tta={self.use_tta}")
+        print("per-class conf: " + ", ".join(
+            f"{n}={t:.3f}" for n, t in zip(self.class_names,
+                                           self._conf_thres_array.tolist())))
+
+        self._warmup()
+
+    def _warmup(self, iters: int = 3) -> None:
+        try:
+            dummy = np.zeros((720, 1280, 3), dtype=np.uint8)
+            for _ in range(max(1, iters)):
+                self.predict_batch(batch_images=[dummy], offset=0, n_keypoints=0)
+            print(f"warmup: {iters} dummy predict_batch call(s) done")
+        except Exception as e:
+            print(f"warmup skipped: {e}")
+
+    def _read_model_class_order(self) -> "list[str] | None":
+        """Read the model's class order from Ultralytics ONNX metadata.
+        Returns the class names ordered by model-emit index, or None when the
+        metadata is missing/unparsable or doesn't match `class_names` as a set
+        (in which case the static model_class_names fallback is used)."""
+        try:
+            import ast
+
+            meta = self.session.get_modelmeta().custom_metadata_map
+            names = ast.literal_eval(meta["names"])  # e.g. {0: 'balaclava', ...}
+            if isinstance(names, dict):
+                order = [str(names[i]) for i in sorted(names)]
+            else:
+                order = [str(n) for n in names]
+        except Exception as e:
+            print(f"cls order: could not read ONNX names metadata ({e})")
+            return None
+        if sorted(order) != sorted(self.class_names):
+            print(
+                f"cls order: ONNX names {order} do not match expected classes "
+                f"{self.class_names}; ignoring metadata"
+            )
+            return None
+        return order
+
+    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]]:
+        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)
+        return padded, ratio, (dw, dh)
+
+    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))
+        # Fused scale(1/255) + BGR->RGB swap + HWC->NCHW + contiguous float32 in
+        # one optimized OpenCV call (bit-identical to the cvtColor + astype/255 +
+        # transpose chain, but ~half the preprocess time).
+        blob = cv2.dnn.blobFromImage(img, scalefactor=1.0 / 255.0, swapRB=True)
+        return blob, 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
+
+    @staticmethod
+    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)
+        order = np.argsort(-scores)
+        keep: list[int] = []
+        while len(order) > 0:
+            i = int(order[0])
+            keep.append(i)
+            if len(order) == 1:
+                break
+            rest = order[1:]
+            xx1 = np.maximum(boxes[i, 0], boxes[rest, 0])
+            yy1 = np.maximum(boxes[i, 1], boxes[rest, 1])
+            xx2 = np.minimum(boxes[i, 2], boxes[rest, 2])
+            yy2 = np.minimum(boxes[i, 3], boxes[rest, 3])
+            inter = np.maximum(0.0, xx2 - xx1) * np.maximum(0.0, yy2 - yy1)
+            a_i = (max(0.0, boxes[i, 2] - boxes[i, 0]) *
+                   max(0.0, boxes[i, 3] - boxes[i, 1]))
+            a_r = (np.maximum(0.0, boxes[rest, 2] - boxes[rest, 0]) *
+                   np.maximum(0.0, boxes[rest, 3] - boxes[rest, 1]))
+            iou = inter / (a_i + a_r - inter + 1e-7)
+            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:
+        if len(boxes) == 0:
+            return np.array([], dtype=np.intp)
+        all_keep: list[int] = []
+        for c in np.unique(cls_ids):
+            mask = cls_ids == c
+            indices = np.where(mask)[0]
+            keep = self._hard_nms(boxes[mask], scores[mask], iou_thresh)
+            all_keep.extend(indices[keep].tolist())
+        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]:
+        n = len(boxes)
+        if n <= 1:
+            return boxes, scores, cls_ids
+        boxes = np.asarray(boxes, dtype=np.float32)
+        scores = np.asarray(scores, dtype=np.float32)
+        cls_ids = np.asarray(cls_ids, dtype=np.int32)
+        areas = (np.maximum(0.0, boxes[:, 2] - boxes[:, 0]) *
+                 np.maximum(0.0, boxes[:, 3] - boxes[:, 1]))
+        margins = scores - self._conf_thres_array[cls_ids]
+        order = np.lexsort((-areas, -margins))
+        suppressed = np.zeros(n, dtype=bool)
+        keep: list[int] = []
+        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)
+            a_i = max(1e-7, float((bi[2] - bi[0]) * (bi[3] - bi[1])))
+            iou = inter / (a_i + areas - inter + 1e-7)
+            dup = iou > iou_thresh
+            dup[i] = False
+            suppressed |= dup
+        keep_idx = np.array(keep, dtype=np.intp)
+        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:
+        n = len(post_boxes)
+        if n == 0:
+            return np.empty(0, dtype=np.float32)
+        full_areas = (np.maximum(0.0, full_boxes[:, 2] - full_boxes[:, 0]) *
+                      np.maximum(0.0, full_boxes[:, 3] - full_boxes[:, 1]))
+        out = np.empty(n, dtype=np.float32)
+        for i in range(n):
+            bi = post_boxes[i]
+            xx1 = np.maximum(bi[0], full_boxes[:, 0])
+            yy1 = np.maximum(bi[1], full_boxes[:, 1])
+            xx2 = np.minimum(bi[2], full_boxes[:, 2])
+            yy2 = np.minimum(bi[3], full_boxes[:, 3])
+            inter = np.maximum(0.0, xx2 - xx1) * np.maximum(0.0, yy2 - yy1)
+            a_i = max(0.0, float((bi[2] - bi[0]) * (bi[3] - bi[1])))
+            iou = inter / (a_i + full_areas - inter + 1e-7)
+            cluster = (iou >= iou_thresh) & (full_cls == post_cls[i])
+            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:
+        """Boolean keep-mask: score >= per-class threshold, with a per-class
+        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)
+        thr = self._conf_thres_array[cls_ids]
+        keep = scores >= thr
+        for c in np.unique(cls_ids):
+            b = float(self._bonus_array[c])
+            if b <= 0.0:
+                continue
+            cm = cls_ids == c
+            if keep[cm].any():
+                continue
+            idx = np.where(cm)[0]
+            top = int(idx[int(np.argmax(scores[idx]))])
+            if scores[top] >= self._conf_thres_array[c] - b:
+                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]:
+        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]
+        if len(scores) > self.max_det:
+            top = np.argsort(-scores)[: self.max_det]
+            boxes, scores, cls_ids = boxes[top], scores[top], cls_ids[top]
+        if len(boxes) > 1:
+            boxes, scores, cls_ids = self._cross_class_dedup_op(
+                boxes, scores, cls_ids, self.cross_iou_thresh
+            )
+        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]:
+        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)
+
+        # Remap model cls_ids -> output cls_ids BEFORE the conf filter, so the
+        # per-class threshold/bonus arrays (indexed in `class_names` order) are
+        # applied to the right class.
+        n_model_cls = len(self.model_class_names)
+        vmask = (cls_ids >= 0) & (cls_ids < n_model_cls)
+        boxes, scores, cls_ids = boxes[vmask], scores[vmask], cls_ids[vmask]
+        cls_ids = self.cls_remap[cls_ids]
+
+        keep = self._conf_filter_mask(scores, cls_ids)
+        boxes = boxes[keep]
+        scores = scores[keep]
+        cls_ids = cls_ids[keep]
+        if len(boxes) == 0:
+            return []
+
+        pad_w, pad_h = pad
+        boxes[:, [0, 2]] -= pad_w
+        boxes[:, [1, 3]] -= pad_h
+        boxes /= ratio
+        boxes = self._clip_boxes(boxes, orig_size)
+
+        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]:
+        if preds.ndim != 3 or preds.shape[0] != 1:
+            raise ValueError(f"Unexpected raw ONNX output shape: {preds.shape}")
+        preds = preds[0]
+        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 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]
+
+        # Remap model cls_ids -> output cls_ids BEFORE the conf filter, so the
+        # per-class threshold/bonus arrays (indexed in `class_names` order) are
+        # applied to the right class.
+        n_model_cls = len(self.model_class_names)
+        vmask = (cls_ids >= 0) & (cls_ids < n_model_cls)
+        boxes_xywh, scores, cls_ids = boxes_xywh[vmask], scores[vmask], cls_ids[vmask]
+        cls_ids = self.cls_remap[cls_ids]
+
+        keep = self._conf_filter_mask(scores, cls_ids)
+        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)
+
+        pad_w, pad_h = pad
+        boxes[:, [0, 2]] -= pad_w
+        boxes[:, [1, 3]] -= pad_h
+        boxes /= ratio
+        boxes = self._clip_boxes(boxes, orig_size)
+
+        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]:
+        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:
+            return self._decode_final_dets(output, ratio, pad, orig_size)
+        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[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 = (1, 3, self.input_height, self.input_width)
+        if input_tensor.shape != expected:
+            raise ValueError(
+                f"Bad input tensor shape={input_tensor.shape}, expected={expected}"
+            )
+
+        outputs = self.session.run(self.output_names, {self.input_name: input_tensor})
+        return self._postprocess(outputs[0], ratio, pad, orig_size)
+
+    def _predict_tta(self, image: np.ndarray) -> list[BoundingBox]:
+        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 not all_boxes:
+            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 []
+        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,
+        )
+
+        kept_coords = coords[hard_keep]
+        kept_cls = cls_ids[hard_keep]
+        if len(kept_coords) > 1:
+            kept_coords, boosted, kept_cls = self._cross_class_dedup_op(
+                kept_coords, boosted, kept_cls, self.cross_iou_thresh
+            )
+
+        return [
+            BoundingBox(
+                x1=int(math.floor(kept_coords[j, 0])),
+                y1=int(math.floor(kept_coords[j, 1])),
+                x2=int(math.ceil(kept_coords[j, 2])),
+                y2=int(math.ceil(kept_coords[j, 3])),
+                cls_id=int(kept_cls[j]),
+                conf=float(boosted[j]),
+            )
+            for j in range(len(kept_coords))
+        ]
+
+    def predict_batch(self, batch_images: list[ndarray], offset: int,
+                      n_keypoints: int) -> list[TVFrameResult]:
+        results: list[TVFrameResult] = []
+        predict = self._predict_tta if self.use_tta else self._predict_single
+        for frame_number_in_batch, image in enumerate(batch_images):
+            try:
+                boxes = predict(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

weights.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:895e01c914229861d8b10e40a2d4f10fc487e6524b98cd8217a2e763085e29d7
+size 9809204