miner.py

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 for Detect-crime.

    Recipe (see recipe.md for derivation):
      * per-class confidence thresholds (king's lever)
      * per-class IoU thresholds (closer pairs allowed for graffiti / spray paint)
      * per-class soft-NMS with sigma=0.5
      * horizontal-flip TTA merged with **weighted box fusion** (coord-averaged)
      * per-class cluster confidence boost (fixes king's cross-class leak)
      * no min-side / min-area filter (small TPs survive)
      * no h/w aspect gate beyond a sane outlier cap (8x)
    """

    class_names = ["balaclava", "hoodie", "glove", "bat", "spray paint", "graffiti"]
    input_size = 1280
    max_det = 300
    max_aspect_ratio = 8.0
    soft_sigma = 0.5

    _conf_thres_array = np.array(
        # balaclava, hoodie, glove, bat, spray paint, graffiti
        [0.50, 0.60, 0.30, 0.20, 0.45, 0.30],
        dtype=np.float32,
    )

    _iou_thres_array = np.array(
        # hoodie/balaclava: clean people, tight NMS
        # glove/bat: small rare objects, mid
        # spray/graffiti: legitimately overlapping marks, loose
        [0.60, 0.60, 0.55, 0.55, 0.45, 0.45],
        dtype=np.float32,
    )

    def __init__(self, path_hf_repo: Path) -> None:
        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 providers available:", ort.get_available_providers())

        sess_options = ort.SessionOptions()
        sess_options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL

        try:
            self.session = ort.InferenceSession(
                str(model_path),
                sess_options=sess_options,
                providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
            )
            print("ORT session: CUDA provider")
        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())

        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 = [o.name for o in self.session.get_outputs()]
        self.input_shape = self.session.get_inputs()[0].shape
        # Detect FP16 vs FP32 input from the session metadata.
        input_type = self.session.get_inputs()[0].type
        self.input_dtype = np.float16 if "float16" in input_type else np.float32
        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: {model_path}")
        print("per-class conf: " + ", ".join(
            f"{n}={t:.2f}" for n, t in zip(self.class_names,
                                           self._conf_thres_array.tolist())))
        print("per-class iou : " + ", ".join(
            f"{n}={t:.2f}" for n, t in zip(self.class_names,
                                           self._iou_thres_array.tolist())))

    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

    # ---------------------------------------------------------------- preproc
    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))
        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=self.input_dtype)
        return img, ratio, pad, (orig_w, orig_h)

    # ---------------------------------------------------------------- helpers
    @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 _iou_matrix(a: np.ndarray, b: np.ndarray) -> np.ndarray:
        if len(a) == 0 or len(b) == 0:
            return np.zeros((len(a), len(b)), dtype=np.float32)
        ax1, ay1, ax2, ay2 = a[:, 0:1], a[:, 1:2], a[:, 2:3], a[:, 3:4]
        bx1, by1, bx2, by2 = b[:, 0], b[:, 1], b[:, 2], b[:, 3]
        ix1 = np.maximum(ax1, bx1)
        iy1 = np.maximum(ay1, by1)
        ix2 = np.minimum(ax2, bx2)
        iy2 = np.minimum(ay2, by2)
        inter = np.maximum(0.0, ix2 - ix1) * np.maximum(0.0, iy2 - iy1)
        area_a = np.maximum(0.0, ax2 - ax1) * np.maximum(0.0, ay2 - ay1)
        area_b = np.maximum(0.0, bx2 - bx1) * np.maximum(0.0, by2 - by1)
        union = area_a + area_b - inter + 1e-7
        return (inter / union).astype(np.float32)

    # ---------------------------------------------------------------- NMS
    def _soft_nms(
        self, boxes: np.ndarray, scores: np.ndarray, sigma: float,
        score_thresh: float = 0.001,
    ) -> tuple[np.ndarray, np.ndarray]:
        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)
            a_i = max(0.0, float(
                (boxes[i, 2] - boxes[i, 0]) * (boxes[i, 3] - boxes[i, 1])
            ))
            a_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 / (a_i + a_j - inter + 1e-7)
            scores[i + 1:] *= np.exp(-(iou ** 2) / sigma)
        mask = scores > score_thresh
        return order[mask], scores[mask]

    def _per_class_soft_nms(
        self, boxes: np.ndarray, scores: np.ndarray, cls_ids: np.ndarray,
    ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
        if len(boxes) == 0:
            return boxes, scores, cls_ids
        out_b: list = []
        out_s: list = []
        out_c: list = []
        for c in np.unique(cls_ids):
            mask = cls_ids == c
            sub_b = boxes[mask]
            sub_s = scores[mask]
            idx, decayed = self._soft_nms(sub_b, sub_s, self.soft_sigma)
            if len(idx) == 0:
                continue
            out_b.append(sub_b[idx])
            out_s.append(decayed)
            out_c.append(np.full(len(idx), c, dtype=cls_ids.dtype))
        if not out_b:
            return (np.empty((0, 4), dtype=np.float32),
                    np.empty((0,), dtype=np.float32),
                    np.empty((0,), dtype=cls_ids.dtype))
        return (np.concatenate(out_b, axis=0),
                np.concatenate(out_s, axis=0),
                np.concatenate(out_c, axis=0))

    # ---------------------------------------------------------------- WBF
    def _weighted_box_fusion(
        self,
        boxes: np.ndarray,
        scores: np.ndarray,
        cls_ids: np.ndarray,
    ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
        """Per-class confidence-weighted box fusion across orig+flip detections."""
        if len(boxes) == 0:
            return boxes, scores, cls_ids
        fused_b: list = []
        fused_s: list = []
        fused_c: list = []
        for c in np.unique(cls_ids):
            mask = cls_ids == c
            sub_b = boxes[mask].astype(np.float32)
            sub_s = scores[mask].astype(np.float32)
            iou_thr = float(self._iou_thres_array[int(c)])
            order = np.argsort(-sub_s)
            sub_b = sub_b[order]
            sub_s = sub_s[order]
            used = np.zeros(len(sub_b), dtype=bool)
            for i in range(len(sub_b)):
                if used[i]:
                    continue
                used[i] = True
                cluster_b = [sub_b[i]]
                cluster_s = [sub_s[i]]
                if i + 1 < len(sub_b):
                    rest = sub_b[i + 1:]
                    ious = self._iou_matrix(sub_b[i:i + 1], rest)[0]
                    for j_offset, iou in enumerate(ious):
                        j = i + 1 + j_offset
                        if used[j]:
                            continue
                        if iou >= iou_thr:
                            used[j] = True
                            cluster_b.append(sub_b[j])
                            cluster_s.append(sub_s[j])
                ws = np.asarray(cluster_s, dtype=np.float32)
                bs = np.asarray(cluster_b, dtype=np.float32)
                w_sum = float(ws.sum())
                if w_sum <= 0:
                    continue
                fused_box = (bs * ws[:, None]).sum(axis=0) / w_sum
                # cluster confidence: max member, slight boost when >1 supporter
                support = len(cluster_s)
                fused_conf = float(ws.max())
                if support > 1:
                    fused_conf = min(1.0, fused_conf * (1.0 + 0.10 * (support - 1)))
                fused_b.append(fused_box)
                fused_s.append(fused_conf)
                fused_c.append(int(c))
        if not fused_b:
            return (np.empty((0, 4), dtype=np.float32),
                    np.empty((0,), dtype=np.float32),
                    np.empty((0,), dtype=cls_ids.dtype))
        return (
            np.asarray(fused_b, dtype=np.float32),
            np.asarray(fused_s, dtype=np.float32),
            np.asarray(fused_c, dtype=cls_ids.dtype),
        )

    # ---------------------------------------------------------------- sanity
    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]:
        if len(boxes) == 0:
            return boxes, scores, cls_ids
        orig_w, orig_h = orig_size
        image_area = float(orig_w * orig_h)
        bw = np.maximum(0.0, boxes[:, 2] - boxes[:, 0])
        bh = np.maximum(0.0, boxes[:, 3] - boxes[:, 1])
        area = bw * bh
        ar = np.where(
            (bw > 0) & (bh > 0),
            np.maximum(bw / np.maximum(bh, 1e-6), bh / np.maximum(bw, 1e-6)),
            np.inf,
        )
        keep = (area > 0) & (area <= 0.95 * image_area) & (ar <= self.max_aspect_ratio)
        return boxes[keep], scores[keep], cls_ids[keep]

    # ---------------------------------------------------------------- decode
    def _per_view_pipeline(
        self, boxes: np.ndarray, scores: np.ndarray, cls_ids: np.ndarray,
        orig_size: tuple[int, int],
    ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
        boxes, scores, cls_ids = self._filter_sane_boxes(boxes, scores, cls_ids, orig_size)
        if len(boxes) == 0:
            return boxes, scores, cls_ids
        if len(boxes) > 1:
            boxes, scores, cls_ids = self._per_class_soft_nms(boxes, scores, cls_ids)
        if len(scores) > self.max_det:
            top = np.argsort(-scores)[: self.max_det]
            boxes, scores, cls_ids = boxes[top], scores[top], cls_ids[top]
        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 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)
        keep = scores >= self._conf_thres_array[cls_ids]
        boxes, scores, cls_ids = boxes[keep], scores[keep], 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, orig_size)
        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 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]
        keep = scores >= self._conf_thres_array[cls_ids]
        boxes_xywh, scores, cls_ids = boxes_xywh[keep], scores[keep], 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, orig_size)
        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)

    # ---------------------------------------------------------------- inference
    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)

        # Per-class weighted box fusion (replaces hard-NMS + cluster-boost)
        boxes_f, scores_f, cls_f = self._weighted_box_fusion(coords, scores, cls_ids)
        if len(boxes_f) == 0:
            return []
        if len(scores_f) > self.max_det:
            top = np.argsort(-scores_f)[: self.max_det]
            boxes_f, scores_f, cls_f = boxes_f[top], scores_f[top], cls_f[top]
        return self._build_results(boxes_f, scores_f, cls_f)

    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)
            except Exception as e:
                print(
                    f"Inference failed for frame "
                    f"{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