Fix CUDA: ldconfig + LD_LIBRARY_PATH, shutdown_after=14400

chute_config.yml

@@ -3,6 +3,7 @@ Image:
   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' 'torch>=2.8'
+    - python3 -c "import os,glob;libs=[d for p in ['/usr/local/lib/python3.12/dist-packages','/usr/lib/python3/dist-packages'] for d in glob.glob(os.path.join(p,'nvidia','*','lib'))];open('/etc/ld.so.conf.d/nvidia-pip.conf','w').write('\n'.join(libs)+'\n');print('Registered CUDA libs:',libs)" && ldconfig
 
 NodeSelector:
   gpu_count: 1
@@ -13,3 +14,4 @@ Chute:
   concurrency: 4
   max_instances: 5
   scaling_threshold: 0.5
+  shutdown_after_seconds: 14400

miner.py

@@ -1,5 +1,13 @@
 from pathlib import Path
 import math
+import os
+import glob
+import site
+
+# Ensure pip-installed NVIDIA CUDA/cuDNN libraries are discoverable
+for sp in site.getsitepackages():
+    for d in glob.glob(os.path.join(sp, 'nvidia', '*', 'lib')):
+        os.environ['LD_LIBRARY_PATH'] = d + ':' + os.environ.get('LD_LIBRARY_PATH', '')
 
 import cv2
 import numpy as np
@@ -38,10 +46,12 @@ class Miner:
         )
         self.input_name = self.session.get_inputs()[0].name
         input_shape = self.session.get_inputs()[0].shape
+        # expected [N, C, H, W]
         self.input_h = int(input_shape[2])
         self.input_w = int(input_shape[3])
-        self.conf_threshold = 0.20
+        self.conf_threshold = 0.15
         self.iou_threshold = 0.3
+        self.use_tta = True
 
     def __repr__(self) -> str:
         return f"ONNX Miner session={type(self.session).__name__} classes={len(self.class_names)}"
@@ -65,48 +75,64 @@ class Miner:
     def _nms(self, dets: list[tuple[float, float, float, float, float, int]]) -> list[tuple[float, float, float, float, float, int]]:
         if not dets:
             return []
+
         boxes = np.array([[d[0], d[1], d[2], d[3]] for d in dets], dtype=np.float32)
         scores = np.array([d[4] for d in dets], dtype=np.float32)
         order = scores.argsort()[::-1]
         keep = []
+
         while order.size > 0:
             i = order[0]
             keep.append(i)
+
             xx1 = np.maximum(boxes[i, 0], boxes[order[1:], 0])
             yy1 = np.maximum(boxes[i, 1], boxes[order[1:], 1])
             xx2 = np.minimum(boxes[i, 2], boxes[order[1:], 2])
             yy2 = np.minimum(boxes[i, 3], boxes[order[1:], 3])
+
             w = np.maximum(0.0, xx2 - xx1)
             h = np.maximum(0.0, yy2 - yy1)
             inter = w * h
+
             area_i = (boxes[i, 2] - boxes[i, 0]) * (boxes[i, 3] - boxes[i, 1])
             area_rest = (boxes[order[1:], 2] - boxes[order[1:], 0]) * (boxes[order[1:], 3] - boxes[order[1:], 1])
             union = np.maximum(area_i + area_rest - inter, 1e-6)
             iou = inter / union
+
             remaining = np.where(iou <= self.iou_threshold)[0]
             order = order[remaining + 1]
+
         return [dets[idx] for idx in keep]
 
-    def _infer_single(self, image_bgr: ndarray) -> list[BoundingBox]:
+    def _decode(self, image_bgr: ndarray) -> list[tuple[float, float, float, float, float, int]]:
+        """Run model and return raw detections before NMS."""
         inp, (orig_h, orig_w) = self._preprocess(image_bgr)
         out = self.session.run(None, {self.input_name: inp})[0]
         pred = self._normalize_predictions(out)
+
         if pred.shape[1] < 5:
             return []
+
         boxes = pred[:, :4]
         cls_scores = pred[:, 4:]
+
         if cls_scores.shape[1] == 0:
             return []
+
         cls_ids = np.argmax(cls_scores, axis=1)
         confs = np.max(cls_scores, axis=1)
         keep = confs >= self.conf_threshold
+
         boxes = boxes[keep]
         confs = confs[keep]
         cls_ids = cls_ids[keep]
+
         if boxes.shape[0] == 0:
             return []
+
         sx = orig_w / float(self.input_w)
         sy = orig_h / float(self.input_h)
+
         dets: list[tuple[float, float, float, float, float, int]] = []
         for i in range(boxes.shape[0]):
             cx, cy, bw, bh = boxes[i].tolist()
@@ -115,30 +141,58 @@ class Miner:
             x2 = (cx + bw / 2.0) * sx
             y2 = (cy + bh / 2.0) * sy
             dets.append((x1, y1, x2, y2, float(confs[i]), int(cls_ids[i])))
-        dets = self._nms(dets)
-        # Filter spurious detections
-        dets = [(x1,y1,x2,y2,conf,cid) for x1,y1,x2,y2,conf,cid in dets
-                if (x2-x1) >= 6 and (y2-y1) >= 6
-                and (x2-x1)*(y2-y1) >= 80
-                and max((x2-x1)/max(y2-y1,1), (y2-y1)/max(x2-x1,1)) <= 10.0]
+
+        return dets
+
+    def _infer_single(self, image_bgr: ndarray) -> list[BoundingBox]:
+        orig_h, orig_w = image_bgr.shape[:2]
+
+        # Original pass
+        all_dets = self._decode(image_bgr)
+
+        # TTA: horizontal flip pass
+        if self.use_tta:
+            flipped = cv2.flip(image_bgr, 1)
+            flip_dets = self._decode(flipped)
+            for x1, y1, x2, y2, conf, cls_id in flip_dets:
+                all_dets.append((orig_w - x2, y1, orig_w - x1, y2, conf, cls_id))
+
+        # NMS
+        all_dets = self._nms(all_dets)
+
         out_boxes: list[BoundingBox] = []
-        for x1, y1, x2, y2, conf, cls_id in dets:
+        for x1, y1, x2, y2, conf, cls_id in all_dets:
             ix1 = max(0, min(orig_w, math.floor(x1)))
             iy1 = max(0, min(orig_h, math.floor(y1)))
             ix2 = max(0, min(orig_w, math.ceil(x2)))
             iy2 = max(0, min(orig_h, math.ceil(y2)))
             out_boxes.append(
-                BoundingBox(x1=ix1, y1=iy1, x2=ix2, y2=iy2,
-                    cls_id=cls_id, conf=max(0.0, min(1.0, conf))))
+                BoundingBox(
+                    x1=ix1,
+                    y1=iy1,
+                    x2=ix2,
+                    y2=iy2,
+                    cls_id=cls_id,
+                    conf=max(0.0, min(1.0, conf)),
+                )
+            )
         return out_boxes
 
     def predict_batch(
-        self, batch_images: list[ndarray], offset: int, n_keypoints: int,
+        self,
+        batch_images: list[ndarray],
+        offset: int,
+        n_keypoints: int,
     ) -> list[TVFrameResult]:
         results: list[TVFrameResult] = []
         for idx, image in enumerate(batch_images):
             boxes = self._infer_single(image)
             keypoints = [(0, 0) for _ in range(max(0, int(n_keypoints)))]
-            results.append(TVFrameResult(
-                frame_id=offset + idx, boxes=boxes, keypoints=keypoints))
+            results.append(
+                TVFrameResult(
+                    frame_id=offset + idx,
+                    boxes=boxes,
+                    keypoints=keypoints,
+                )
+            )
         return results