scorevision: push artifact

chute_config.yml

@@ -9,11 +9,9 @@ NodeSelector:
   gpu_count: 1
   min_vram_gb_per_gpu: 16
   max_hourly_price_per_gpu: 2.0
-  exclude:
-  - '5090'
-  - b200
-  - h200
-  - mi300x
+  include:
+  - '4090'
+  - pro_6000
 Chute:
   timeout_seconds: 900
   concurrency: 4

miner.py

@@ -1,6 +1,6 @@
 """
-Score Vision SN44 — Unified miner v3.15 (2026-04-02). Background TRT engine build.
-Dual-model: vehicle (YOLO11m INT8 1280, CUDA) + person (YOLO26s FP16 960 end2end, TRT).
+Score Vision SN44 — Unified miner v3.16 (2026-04-03). YOLO12s + TRT + CLAHE.
+Dual-model: vehicle (YOLO11m INT8 1280, CUDA) + person (YOLO12s FP16 960 end2end, TRT).
 Pose model: YOLOv8n-pose FP16 640 for false-positive filtering + keypoint box refinement.
 Vehicle weights loaded from secondary HF repo (meaculpitt/ScoreVision-Vehicle).
 Person weights loaded from primary HF repo (template downloads automatically).
@@ -13,7 +13,7 @@ Vehicle model (vehicle_weights.onnx):
   Flip TTA always enabled — compensates for higher confidence thresholds.
 
 Person model (person_weights.onnx):
-  YOLO26s FP16 960px end2end [1,300,6]. Single class: 0=person.
+  YOLO12s FP16 960px end2end [1,300,6]. Single class: 0=person.
   Background TRT build: starts on CUDA immediately, builds TRT FP16 engine in background
   thread (~18min on fresh node), swaps to TRT atomically when ready. Cached thereafter.
   SAHI-style tiling: full + 2 adaptive tiles + flip TTA, max-conf NMS merge.
@@ -39,56 +39,120 @@ import logging as _logging
 _cuda_log = _logging.getLogger(__name__)
 
 def _preload_cuda_libs():
-    """Pre-load CUDA + TensorRT libs from pip packages so ORT GPU/TRT providers work."""
+    """Pre-load CUDA + TensorRT libs from pip packages so ORT GPU/TRT providers work.
+
+    Search order for TRT libs (libnvinfer.so, libnvonnxparser.so):
+      1. sys.path entries containing tensorrt_libs/ subdirectory
+      2. site.getsitepackages() + user site-packages for tensorrt_libs/ or tensorrt/
+      3. ctypes.util.find_library('nvinfer') as system-wide fallback
+    If not found, logs clearly and skips TRT — never attempts pip operations.
+    """
     try:
+        import ctypes.util as _ctypes_util
         lib_dirs = []
-        # CUDA libs from nvidia pip packages
+        loaded = set()
+
+        # ── CUDA libs from nvidia pip packages ──
         for mod_name in ['nvidia.cudnn', 'nvidia.cublas', 'nvidia.cuda_runtime',
                          'nvidia.cufft', 'nvidia.curand', 'nvidia.cusolver',
                          'nvidia.cusparse', 'nvidia.nvjitlink']:
             try:
                 mod = __import__(mod_name, fromlist=['__file__'])
                 lib_dir = os.path.join(os.path.dirname(mod.__file__), 'lib')
-                if os.path.isdir(lib_dir):
+                if os.path.isdir(lib_dir) and lib_dir not in lib_dirs:
                     lib_dirs.append(lib_dir)
             except ImportError:
                 pass
 
-        # TensorRT libs — search site-packages for tensorrt_libs directory
-        import sys as _sys_inner
-        _trt_found = False
-        for p in _sys_inner.path:
-            candidate = os.path.join(p, 'tensorrt_libs')
-            if os.path.isdir(candidate):
-                lib_dirs.append(candidate)
-                _trt_found = True
+        # ── TensorRT libs — multi-strategy search ──
+        import sys as _sys
+        _trt_dir = None
+
+        # Strategy 1: sys.path (covers standard pip installs)
+        for p in _sys.path:
+            for subdir in ('tensorrt_libs', 'tensorrt'):
+                candidate = os.path.join(p, subdir)
+                if os.path.isdir(candidate) and _glob.glob(os.path.join(candidate, 'libnvinfer*')):
+                    _trt_dir = candidate
+                    break
+            if _trt_dir:
                 break
-        # Broader search if not found in sys.path
-        if not _trt_found:
-            for base in ['/usr/local/lib', '/usr/lib', os.path.expanduser('~/.local/lib'),
-                         '/home/miner/.local/lib']:
-                for root, dirs, _ in os.walk(base):
-                    if 'tensorrt_libs' in dirs:
-                        lib_dirs.append(os.path.join(root, 'tensorrt_libs'))
-                        _trt_found = True
-                        break
-                    if root.count(os.sep) - base.count(os.sep) > 4:
+
+        # Strategy 2: site-packages directories (covers user installs, venvs)
+        if not _trt_dir:
+            import site
+            search_dirs = list(site.getsitepackages()) if hasattr(site, 'getsitepackages') else []
+            user_site = getattr(site, 'getusersitepackages', lambda: None)()
+            if user_site:
+                search_dirs.append(user_site)
+            # Also check common paths not always in site
+            search_dirs.extend([
+                '/usr/local/lib/python3.12/dist-packages',
+                os.path.expanduser('~/.local/lib/python3.12/site-packages'),
+                '/home/miner/.local/lib/python3.12/site-packages',
+            ])
+            for sp in search_dirs:
+                for subdir in ('tensorrt_libs', 'tensorrt'):
+                    candidate = os.path.join(sp, subdir)
+                    if os.path.isdir(candidate) and _glob.glob(os.path.join(candidate, 'libnvinfer*')):
+                        _trt_dir = candidate
                         break
-                if _trt_found:
+                if _trt_dir:
                     break
 
-        if not lib_dirs:
+        # Strategy 3: ctypes.util.find_library (system-wide LD search)
+        if not _trt_dir:
+            nvinfer_path = _ctypes_util.find_library('nvinfer')
+            if nvinfer_path:
+                _cuda_log.info('TRT found via system library: %s', nvinfer_path)
+                try:
+                    ctypes.CDLL(nvinfer_path, mode=ctypes.RTLD_GLOBAL)
+                    loaded.add('nvinfer')
+                except OSError as e:
+                    _cuda_log.warning('Failed to load system nvinfer: %s', e)
+
+        if _trt_dir:
+            if _trt_dir not in lib_dirs:
+                lib_dirs.append(_trt_dir)
+            _cuda_log.info('TRT libs directory: %s', _trt_dir)
+        elif 'nvinfer' not in loaded:
+            _cuda_log.info('TensorRT libs not found — TRT EP will be unavailable (CUDA EP still works)')
+
+        if not lib_dirs and not loaded:
+            _cuda_log.warning('No CUDA or TRT libs found to preload')
             return
+
+        # Set LD_LIBRARY_PATH for any child processes / dlopen fallbacks
         existing = os.environ.get('LD_LIBRARY_PATH', '')
         os.environ['LD_LIBRARY_PATH'] = ':'.join(lib_dirs + ([existing] if existing else []))
+
+        # Load CUDA libs (glob all .so in nvidia dirs)
         for lib_dir in lib_dirs:
+            if 'tensorrt' in lib_dir:
+                continue  # TRT libs loaded selectively below
             for so in sorted(_glob.glob(os.path.join(lib_dir, 'lib*.so*'))):
                 try:
                     ctypes.CDLL(so, mode=ctypes.RTLD_GLOBAL)
                 except OSError:
                     pass
+
+        # Load TRT libs selectively (only the essentials, not builder resources)
+        if _trt_dir:
+            for lib_name in ['libnvinfer.so', 'libnvinfer_plugin.so', 'libnvonnxparser.so']:
+                matches = _glob.glob(os.path.join(_trt_dir, lib_name + '*'))
+                if matches:
+                    try:
+                        ctypes.CDLL(matches[0], mode=ctypes.RTLD_GLOBAL)
+                        loaded.add(lib_name.split('.')[0])
+                    except OSError as e:
+                        _cuda_log.warning('Failed to load %s: %s', lib_name, e)
+                else:
+                    _cuda_log.info('%s not found in %s', lib_name, _trt_dir)
+
+        if loaded:
+            _cuda_log.info('Preloaded libs: %s', ', '.join(sorted(loaded)))
     except Exception as e:
-        _cuda_log.warning(f'CUDA/TRT preload error: {e}')
+        _cuda_log.warning('CUDA/TRT preload error: %s', e)
 
 _preload_cuda_libs()
 
@@ -1128,6 +1192,14 @@ class Miner:
         inp = np.ascontiguousarray(inp.transpose(2, 0, 1)[np.newaxis])
         return inp, ratio, pl, pt
 
+    def _per_enhance(self, img_bgr):
+        """CLAHE contrast enhancement (clip=12) on LAB L-channel."""
+        lab = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2LAB)
+        l, a, b = cv2.split(lab)
+        clahe = cv2.createCLAHE(clipLimit=12.0, tileGridSize=(8, 8))
+        l = clahe.apply(l)
+        return cv2.cvtColor(cv2.merge([l, a, b]), cv2.COLOR_LAB2BGR)
+
     def _per_decode(self, raw, ratio, pl, pt, oh, ow, conf_thresh):
         pred = raw[0]
         if pred.ndim != 2:
@@ -1676,6 +1748,14 @@ class Miner:
                 all_boxes.append(boxes_flip)
                 all_confs.append(confs_flip)
 
+        # Pass 5: CLAHE enhanced preprocessing pass
+        if time.monotonic() - t_start < PER_RTF_BUDGET / 4:
+            enhanced = self._per_enhance(image_bgr)
+            boxes_enh, confs_enh = self._per_run_pass(enhanced, PER_CONF_LOW)
+            if len(boxes_enh) > 0:
+                all_boxes.append(boxes_enh)
+                all_confs.append(confs_enh)
+
         if not all_boxes:
             return []
 

person_weights.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b23ca9c0a7efd6613753b66033d7c6637b338a79547562687d5e58ba5fb92d7c
-size 19258183
+oid sha256:84e9a6adf84703b2fa13e5e587f1a7ffc8ab69f6d45c2fc7ae71ded6ff9a926f
+size 18860997