Upload folder using huggingface_hub

chute_config.yml

@@ -16,6 +16,7 @@ NodeSelector:
     - b200
     - h200
     - mi300x
+    - pro_6000
 
 Chute:
   timeout_seconds: 900

miner.py

@@ -68,13 +68,13 @@ class Miner:
         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=1280)
-        self.input_width = self._safe_dim(self.input_shape[3], default=1280)
+        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 = 0.33  # Higher = fewer FP, slightly lower recall
-        self.iou_thres = 0.5   # Lower = suppress duplicate detections (FP)
+        self.conf_thres = 0.32  # Higher = fewer FP, slightly lower recall
+        self.iou_thres = 0.55   # Lower = suppress duplicate detections (FP)
         self.max_det = 150     # Cap detections; sports ~20-30 persons
         self.use_tta = True
 
@@ -188,7 +188,7 @@ class Miner:
         boxes: np.ndarray,
         scores: np.ndarray,
         sigma: float = 0.5,
-        score_thresh: float = 0.01,
+        score_thresh: float = 0.2,
     ) -> tuple[np.ndarray, np.ndarray]:
         """
         Soft-NMS: Gaussian decay of overlapping scores instead of hard removal.
@@ -198,6 +198,8 @@ class Miner:
         if N == 0:
             return np.array([], dtype=np.intp), np.array([], dtype=np.float32)
 
+        sigma = max(float(sigma), 1e-6)
+
         boxes = boxes.astype(np.float32, copy=True)
         scores = scores.astype(np.float32, copy=True)
         order = np.arange(N)