Upload miner.py with huggingface_hub

miner.py

@@ -0,0 +1,314 @@
+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:
+    def __init__(self, path_hf_repo: Path) -> None:
+        model_path = path_hf_repo / "weights.onnx"
+        self.class_names = ["person"]
+
+        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"],
+            )
+        except Exception:
+            self.session = ort.InferenceSession(
+                str(model_path),
+                sess_options=sess_options,
+                providers=["CPUExecutionProvider"],
+            )
+
+        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
+        self.input_height = self._safe_dim(self.input_shape[2], 1280)
+        self.input_width = self._safe_dim(self.input_shape[3], 1280)
+
+        # Tuned for MAP50 (65%) + FALSE_POSITIVE (35%) scoring
+        # Lower conf = more recall = higher MAP50, but more FP
+        # Balance: slightly aggressive recall since MAP50 weight > FP weight
+        self.conf_thres = 0.40
+        self.conf_high = 0.55
+        self.iou_thres = 0.50
+        self.tta_match_iou = 0.45
+        self.max_det = 200
+        self.use_tta = True
+
+        # Box sanity filters
+        self.min_box_area = 12 * 12
+        self.min_w = 6
+        self.min_h = 6
+        self.max_aspect_ratio = 7.0
+        self.max_box_area_ratio = 0.85
+
+        print(f"Model loaded: {model_path}, providers={self.session.get_providers()}")
+
+    def __repr__(self) -> str:
+        return f"ONNXRuntime(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)
+        rw, rh = int(round(w * ratio)), int(round(h * ratio))
+        if (rw, rh) != (w, h):
+            interp = cv2.INTER_CUBIC if ratio > 1.0 else cv2.INTER_LINEAR
+            image = cv2.resize(image, (rw, rh), interpolation=interp)
+        dw, dh = (new_w - rw) / 2.0, (new_h - rh) / 2.0
+        padded = cv2.copyMakeBorder(
+            image, int(round(dh - 0.1)), int(round(dh + 0.1)),
+            int(round(dw - 0.1)), int(round(dw + 0.1)),
+            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).astype(np.float32) / 255.0
+        img = np.ascontiguousarray(np.transpose(img, (2, 0, 1))[None, ...], dtype=np.float32)
+        return img, ratio, pad, (orig_w, orig_h)
+
+    @staticmethod
+    def _clip_boxes(boxes: np.ndarray, size: tuple[int, int]) -> np.ndarray:
+        w, h = 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(b: np.ndarray) -> np.ndarray:
+        o = np.empty_like(b)
+        o[:, 0] = b[:, 0] - b[:, 2] / 2
+        o[:, 1] = b[:, 1] - b[:, 3] / 2
+        o[:, 2] = b[:, 0] + b[:, 2] / 2
+        o[:, 3] = b[:, 1] + b[:, 3] / 2
+        return o
+
+    @staticmethod
+    def _hard_nms(boxes: np.ndarray, scores: np.ndarray, iou_thresh: float) -> np.ndarray:
+        if len(boxes) == 0:
+            return np.array([], dtype=np.intp)
+        order = np.argsort(scores)[::-1]
+        keep = []
+        while len(order) > 0:
+            i = 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, xx2 - xx1) * np.maximum(0, yy2 - yy1)
+            area_i = max(0, (boxes[i, 2] - boxes[i, 0]) * (boxes[i, 3] - boxes[i, 1]))
+            area_r = np.maximum(0, boxes[rest, 2] - boxes[rest, 0]) * np.maximum(0, boxes[rest, 3] - boxes[rest, 1])
+            iou = inter / (area_i + area_r - inter + 1e-7)
+            order = rest[iou <= iou_thresh]
+        return np.array(keep, dtype=np.intp)
+
+    @staticmethod
+    def _box_iou_one_to_many(box: np.ndarray, boxes: np.ndarray) -> np.ndarray:
+        xx1 = np.maximum(box[0], boxes[:, 0])
+        yy1 = np.maximum(box[1], boxes[:, 1])
+        xx2 = np.minimum(box[2], boxes[:, 2])
+        yy2 = np.minimum(box[3], boxes[:, 3])
+        inter = np.maximum(0, xx2 - xx1) * np.maximum(0, yy2 - yy1)
+        a = max(0, (box[2] - box[0]) * (box[3] - box[1]))
+        b = np.maximum(0, boxes[:, 2] - boxes[:, 0]) * np.maximum(0, boxes[:, 3] - boxes[:, 1])
+        return inter / (a + b - inter + 1e-7)
+
+    def _filter_sane(self, boxes, scores, cls_ids, orig_size):
+        if len(boxes) == 0:
+            return boxes, scores, cls_ids
+        ow, oh = orig_size
+        area_img = float(ow * oh)
+        keep = []
+        for i, box in enumerate(boxes):
+            bw, bh = box[2] - box[0], box[3] - box[1]
+            if bw <= 0 or bh <= 0 or bw < self.min_w or bh < self.min_h:
+                continue
+            area = bw * bh
+            if area < self.min_box_area or area > self.max_box_area_ratio * area_img:
+                continue
+            if max(bw / max(bh, 1e-6), bh / max(bw, 1e-6)) > 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 _decode_raw_yolo(self, preds, ratio, pad, orig_size):
+        if preds.ndim == 3 and preds.shape[0] == 1:
+            preds = preds[0]
+        if preds.shape[0] <= 16 and preds.shape[1] > preds.shape[0]:
+            preds = preds.T
+
+        boxes_xywh = preds[:, :4].astype(np.float32)
+        tail = preds[:, 4:]
+
+        if tail.shape[1] == 1:
+            scores = tail[:, 0]
+            cls_ids = np.zeros(len(scores), dtype=np.int32)
+        else:
+            cls_ids = np.argmax(tail, axis=1).astype(np.int32)
+            scores = tail[np.arange(len(tail)), cls_ids]
+
+        # person only (class 0)
+        mask = (cls_ids == 0) & (scores >= self.conf_thres)
+        boxes_xywh, scores, cls_ids = boxes_xywh[mask], scores[mask], cls_ids[mask]
+        if len(boxes_xywh) == 0:
+            return []
+
+        boxes = self._xywh_to_xyxy(boxes_xywh)
+        boxes[:, [0, 2]] -= pad[0]
+        boxes[:, [1, 3]] -= pad[1]
+        boxes /= ratio
+        boxes = self._clip_boxes(boxes, orig_size)
+        boxes, scores, cls_ids = self._filter_sane(boxes, scores, cls_ids, orig_size)
+        if len(boxes) == 0:
+            return []
+
+        keep = self._hard_nms(boxes, scores, self.iou_thres)[:self.max_det]
+        return [BoundingBox(
+            x1=int(math.floor(boxes[i, 0])), y1=int(math.floor(boxes[i, 1])),
+            x2=int(math.ceil(boxes[i, 2])), y2=int(math.ceil(boxes[i, 3])),
+            cls_id=0, conf=float(scores[i]))
+            for i in keep if boxes[i, 2] > boxes[i, 0] and boxes[i, 3] > boxes[i, 1]]
+
+    def _decode_final_dets(self, preds, ratio, pad, orig_size):
+        if preds.ndim == 3 and preds.shape[0] == 1:
+            preds = preds[0]
+        boxes = preds[:, :4].astype(np.float32)
+        scores = preds[:, 4].astype(np.float32)
+        cls_ids = preds[:, 5].astype(np.int32)
+
+        mask = (cls_ids == 0) & (scores >= self.conf_thres)
+        boxes, scores, cls_ids = boxes[mask], scores[mask], cls_ids[mask]
+        if len(boxes) == 0:
+            return []
+
+        boxes[:, [0, 2]] -= pad[0]
+        boxes[:, [1, 3]] -= pad[1]
+        boxes /= ratio
+        boxes = self._clip_boxes(boxes, orig_size)
+        boxes, scores, cls_ids = self._filter_sane(boxes, scores, cls_ids, orig_size)
+        if len(boxes) == 0:
+            return []
+
+        keep = self._hard_nms(boxes, scores, self.iou_thres)[:self.max_det]
+        return [BoundingBox(
+            x1=int(math.floor(boxes[i, 0])), y1=int(math.floor(boxes[i, 1])),
+            x2=int(math.ceil(boxes[i, 2])), y2=int(math.ceil(boxes[i, 3])),
+            cls_id=0, conf=float(scores[i]))
+            for i in keep if boxes[i, 2] > boxes[i, 0] and boxes[i, 3] > boxes[i, 1]]
+
+    def _postprocess(self, output, ratio, pad, orig_size):
+        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.dtype != np.uint8:
+            image = image.astype(np.uint8)
+        tensor, ratio, pad, orig_size = self._preprocess(image)
+        outputs = self.session.run(self.output_names, {self.input_name: tensor})
+        return self._postprocess(outputs[0], ratio, pad, orig_size)
+
+    def _merge_tta(self, boxes_orig, boxes_flip):
+        if not boxes_orig and not boxes_flip:
+            return []
+
+        co = np.array([[b.x1, b.y1, b.x2, b.y2] for b in boxes_orig], dtype=np.float32) if boxes_orig else np.empty((0, 4), dtype=np.float32)
+        so = np.array([b.conf for b in boxes_orig], dtype=np.float32) if boxes_orig else np.empty(0, dtype=np.float32)
+        cf = np.array([[b.x1, b.y1, b.x2, b.y2] for b in boxes_flip], dtype=np.float32) if boxes_flip else np.empty((0, 4), dtype=np.float32)
+        sf = np.array([b.conf for b in boxes_flip], dtype=np.float32) if boxes_flip else np.empty(0, dtype=np.float32)
+
+        acc_b, acc_s = [], []
+
+        for i in range(len(co)):
+            if so[i] >= self.conf_high:
+                acc_b.append(co[i]); acc_s.append(so[i])
+            elif len(cf) > 0:
+                ious = self._box_iou_one_to_many(co[i], cf)
+                j = int(np.argmax(ious))
+                if ious[j] >= self.tta_match_iou:
+                    acc_b.append(co[i]); acc_s.append(max(so[i], sf[j]))
+
+        for i in range(len(cf)):
+            if sf[i] < self.conf_high:
+                continue
+            if len(co) == 0:
+                acc_b.append(cf[i]); acc_s.append(sf[i]); continue
+            if np.max(self._box_iou_one_to_many(cf[i], co)) < self.tta_match_iou:
+                acc_b.append(cf[i]); acc_s.append(sf[i])
+
+        if not acc_b:
+            return []
+
+        boxes = np.array(acc_b, dtype=np.float32)
+        scores = np.array(acc_s, dtype=np.float32)
+        keep = self._hard_nms(boxes, scores, self.iou_thres)[:self.max_det]
+
+        return [BoundingBox(
+            x1=int(math.floor(boxes[i, 0])), y1=int(math.floor(boxes[i, 1])),
+            x2=int(math.ceil(boxes[i, 2])), y2=int(math.ceil(boxes[i, 3])),
+            cls_id=0, conf=float(scores[i])) for i in keep]
+
+    def _predict_tta(self, image: np.ndarray) -> list[BoundingBox]:
+        boxes_orig = self._predict_single(image)
+        flipped = cv2.flip(image, 1)
+        boxes_flip_raw = 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_raw]
+        return self._merge_tta(boxes_orig, boxes_flip)
+
+    def predict_batch(self, batch_images: list[ndarray], offset: int, n_keypoints: int) -> list[TVFrameResult]:
+        results = []
+        for i, image in enumerate(batch_images):
+            try:
+                boxes = self._predict_tta(image) if self.use_tta else self._predict_single(image)
+            except Exception as e:
+                print(f"Inference failed frame {offset + i}: {e}")
+                boxes = []
+            results.append(TVFrameResult(
+                frame_id=offset + i, boxes=boxes,
+                keypoints=[(0, 0) for _ in range(max(0, int(n_keypoints)))]))
+        return results