scorevision: push petrol v2 model

chute_config.yml

@@ -0,0 +1,22 @@
+Image:
+  from_base: parachutes/python:3.12
+  run_command:
+  - pip install --upgrade setuptools wheel
+  - pip install 'numpy>=1.23' 'onnxruntime-gpu>=1.16' 'nvidia-cudnn-cu12' 'nvidia-cublas-cu12'
+    'opencv-python-headless>=4.7' 'pillow>=9.5' 'huggingface_hub>=0.19.4' 'pydantic>=2.0'
+    'pyyaml>=6.0' 'aiohttp>=3.9' 'ensemble-boxes>=1.0' 'torch>=2.6,<3.0'
+NodeSelector:
+  gpu_count: 1
+  min_vram_gb_per_gpu: 16
+  max_hourly_price_per_gpu: 0.50
+  exclude:
+  - '5090'
+  - b200
+  - h200
+  - mi300x
+Chute:
+  timeout_seconds: 300
+  concurrency: 4
+  max_instances: 5
+  scaling_threshold: 0.5
+  shutdown_after_seconds: 288000

class_names.txt

@@ -0,0 +1,4 @@
+petrol hose
+petrol pump
+price board
+roof canopy

miner.py

@@ -0,0 +1,370 @@
+from pathlib import Path
+import math
+import logging
+
+import cv2
+import numpy as np
+import onnxruntime as ort
+from numpy import ndarray
+from pydantic import BaseModel
+
+logger = logging.getLogger(__name__)
+
+# ─── Petrol miner v1.1 ───────────────────────────────────────────────
+# Improvements over auto-generated baseline:
+# 1. Fix end-to-end ONNX decode (model outputs [1,300,6] post-NMS)
+# 2. Spatial co-occurrence scoring (pump+canopy boost, isolated suppress)
+# 3. Geometric validation (aspect ratio + size checks per class)
+# ──────────────────────────────────────────────────────────────────────
+
+# Class IDs
+CLS_HOSE = 0
+CLS_PUMP = 1
+CLS_PRICEBOARD = 2
+CLS_CANOPY = 3
+
+# ── Geometric validation thresholds (derived from 2000-label analysis) ──
+# Canopy: wide/flat, aspect(w/h) mean=2.96. Suppress if aspect < 0.8 (too tall)
+CANOPY_MIN_ASPECT = 0.8
+# Pump: roughly square/tall, aspect mean=0.91. Suppress if aspect > 4.0 (too wide)
+PUMP_MAX_ASPECT = 4.0
+# Price board: small. Suppress if area > 15% of image
+PRICEBOARD_MAX_AREA_FRAC = 0.15
+# Hose: variable. Suppress if area < 0.05% of image (tiny FP)
+HOSE_MIN_AREA_FRAC = 0.0005
+
+# ── Spatial co-occurrence boost/suppress amounts ──
+COOCCUR_BOOST_PUMP_CANOPY = 0.05
+COOCCUR_BOOST_PUMP_HOSE = 0.08
+COOCCUR_BOOST_CANOPY_HOSE = 0.05
+COOCCUR_SUPPRESS_ISOLATED = 0.03  # per missing expected neighbor
+# Proximity threshold: normalized distance between box centers
+COOCCUR_PROXIMITY = 0.5  # half of image dimension
+
+# ── Geometric suppress penalty ──
+GEOMETRIC_SUPPRESS_PENALTY = 0.10
+
+
+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:
+    VERSION = "petrol-v1.1"
+
+    def __init__(self, path_hf_repo: Path) -> None:
+        self.path_hf_repo = path_hf_repo
+        self.class_names = ['petrol hose', 'petrol pump', 'price board', 'roof canopy']
+        self.session = ort.InferenceSession(
+            str(path_hf_repo / "weights.onnx"),
+            providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
+        )
+        self.input_name = self.session.get_inputs()[0].name
+        input_shape = self.session.get_inputs()[0].shape
+        self.input_h = int(input_shape[2])
+        self.input_w = int(input_shape[3])
+        self.conf_threshold = 0.25
+        self.iou_threshold = 0.45
+
+        # Detect output format: end-to-end [1,N,6] vs raw [1,C,N]
+        out_shape = self.session.get_outputs()[0].shape
+        # End-to-end: [1, max_dets, 6] where max_dets is small (100-300)
+        # Raw: [1, 4+nc, N] where N is large (8400+)
+        if len(out_shape) == 3 and out_shape[2] == 6 and (out_shape[1] or 0) <= 1000:
+            self._end2end = True
+            logger.info("[init] End-to-end ONNX output detected")
+        else:
+            self._end2end = False
+            logger.info("[init] Raw ONNX output detected")
+
+        logger.info(f"[init] {self.VERSION} loaded, input={self.input_w}x{self.input_h}, "
+                     f"end2end={self._end2end}")
+
+    def __repr__(self) -> str:
+        return f"Petrol Miner {self.VERSION} end2end={self._end2end}"
+
+    # ─── Preprocessing ────────────────────────────────────────────────
+
+    def _preprocess(self, image_bgr: ndarray) -> tuple[np.ndarray, tuple[int, int]]:
+        h, w = image_bgr.shape[:2]
+        rgb = cv2.cvtColor(image_bgr, cv2.COLOR_BGR2RGB)
+        resized = cv2.resize(rgb, (self.input_w, self.input_h))
+        x = resized.astype(np.float32) / 255.0
+        x = np.transpose(x, (2, 0, 1))[None, ...]
+        return x, (h, w)
+
+    # ─── NMS (only needed for raw output format) ─────────────────────
+
+    def _nms(self, dets):
+        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]
+
+    # ─── Decode: handles both end-to-end and raw formats ─────────────
+
+    def _decode_end2end(self, out, orig_h, orig_w):
+        """Decode end-to-end [1, N, 6] output: [x1,y1,x2,y2,conf,cls_id] in input coords."""
+        pred = out[0]  # [N, 6]
+        if pred.ndim != 2 or pred.shape[1] != 6:
+            return []
+
+        confs = pred[:, 4]
+        keep = confs >= self.conf_threshold
+        pred = pred[keep]
+        if pred.shape[0] == 0:
+            return []
+
+        sx = orig_w / float(self.input_w)
+        sy = orig_h / float(self.input_h)
+
+        results = []
+        for i in range(pred.shape[0]):
+            x1 = pred[i, 0] * sx
+            y1 = pred[i, 1] * sy
+            x2 = pred[i, 2] * sx
+            y2 = pred[i, 3] * sy
+            conf = float(pred[i, 4])
+            cls_id = int(pred[i, 5])
+            results.append((x1, y1, x2, y2, conf, cls_id))
+        return results
+
+    def _decode_raw(self, out, orig_h, orig_w):
+        """Decode raw [1, 4+nc, N] or [1, N, 4+nc] output."""
+        pred = out[0]
+        if pred.ndim != 2:
+            return []
+        if pred.shape[0] < pred.shape[1]:
+            pred = pred.T
+        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, confs, cls_ids = boxes[keep], confs[keep], cls_ids[keep]
+        if boxes.shape[0] == 0:
+            return []
+
+        sx = orig_w / float(self.input_w)
+        sy = orig_h / float(self.input_h)
+
+        dets = []
+        for i in range(boxes.shape[0]):
+            cx, cy, bw, bh = boxes[i].tolist()
+            x1 = (cx - bw / 2.0) * sx
+            y1 = (cy - bh / 2.0) * sy
+            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])))
+        return self._nms(dets)
+
+    # ─── Geometric validation ────────────────────────────────────────
+
+    def _geometric_validate(self, dets, orig_h, orig_w):
+        """Suppress detections that fail basic geometric expectations.
+
+        Returns list with penalties applied to conf.
+        - Canopy: must be wide (aspect w/h >= 0.8)
+        - Pump: must not be extremely wide (aspect w/h <= 4.0)
+        - Price board: must be small (area <= 15% of image)
+        - Hose: must not be tiny (area >= 0.05% of image)
+        """
+        img_area = max(orig_h * orig_w, 1)
+        result = []
+        for x1, y1, x2, y2, conf, cls_id in dets:
+            bw = max(x2 - x1, 1)
+            bh = max(y2 - y1, 1)
+            aspect = bw / bh
+            box_area = bw * bh
+            area_frac = box_area / img_area
+            penalty = 0.0
+
+            if cls_id == CLS_CANOPY:
+                if aspect < CANOPY_MIN_ASPECT:
+                    penalty = GEOMETRIC_SUPPRESS_PENALTY
+            elif cls_id == CLS_PUMP:
+                if aspect > PUMP_MAX_ASPECT:
+                    penalty = GEOMETRIC_SUPPRESS_PENALTY
+            elif cls_id == CLS_PRICEBOARD:
+                if area_frac > PRICEBOARD_MAX_AREA_FRAC:
+                    penalty = GEOMETRIC_SUPPRESS_PENALTY
+            elif cls_id == CLS_HOSE:
+                if area_frac < HOSE_MIN_AREA_FRAC:
+                    penalty = GEOMETRIC_SUPPRESS_PENALTY
+
+            new_conf = max(0.0, conf - penalty)
+            if new_conf >= self.conf_threshold:
+                result.append((x1, y1, x2, y2, new_conf, cls_id))
+        return result
+
+    # ─── Spatial co-occurrence scoring ───────────────────────────────
+
+    def _spatial_cooccurrence(self, dets, orig_h, orig_w):
+        """Adjust confidences based on spatial co-occurrence patterns.
+
+        Boosts:
+        - Pump near canopy: both get +0.05
+        - Pump near hose: hose gets +0.08
+        - Canopy near hose: hose gets +0.05
+
+        Suppresses:
+        - Low-conf detection with no neighbors of expected class: -0.03
+          (except price boards, which are 91% solo in training data)
+        """
+        if not dets:
+            return dets
+
+        n = len(dets)
+        adjustments = [0.0] * n
+        diag = math.sqrt(orig_h ** 2 + orig_w ** 2)
+        prox = COOCCUR_PROXIMITY * diag  # absolute pixel distance
+
+        # Precompute centers
+        centers = []
+        for x1, y1, x2, y2, conf, cls_id in dets:
+            centers.append(((x1 + x2) / 2, (y1 + y2) / 2))
+
+        # Build per-class index
+        cls_map = {}
+        for i, (_, _, _, _, _, cls_id) in enumerate(dets):
+            cls_map.setdefault(cls_id, []).append(i)
+
+        def near(i, j):
+            dx = centers[i][0] - centers[j][0]
+            dy = centers[i][1] - centers[j][1]
+            return math.sqrt(dx * dx + dy * dy) < prox
+
+        # Pump + Canopy boost
+        for pi in cls_map.get(CLS_PUMP, []):
+            for ci in cls_map.get(CLS_CANOPY, []):
+                if near(pi, ci):
+                    adjustments[pi] = max(adjustments[pi], COOCCUR_BOOST_PUMP_CANOPY)
+                    adjustments[ci] = max(adjustments[ci], COOCCUR_BOOST_PUMP_CANOPY)
+
+        # Pump + Hose boost (hose gets larger boost)
+        for pi in cls_map.get(CLS_PUMP, []):
+            for hi in cls_map.get(CLS_HOSE, []):
+                if near(pi, hi):
+                    adjustments[hi] = max(adjustments[hi], COOCCUR_BOOST_PUMP_HOSE)
+
+        # Canopy + Hose boost
+        for ci in cls_map.get(CLS_CANOPY, []):
+            for hi in cls_map.get(CLS_HOSE, []):
+                if near(ci, hi):
+                    adjustments[hi] = max(adjustments[hi], COOCCUR_BOOST_CANOPY_HOSE)
+
+        # Suppress isolated low-confidence detections (not price boards)
+        for i, (x1, y1, x2, y2, conf, cls_id) in enumerate(dets):
+            if cls_id == CLS_PRICEBOARD:
+                continue  # price boards are often solo (91% in training)
+            if conf > 0.60:
+                continue  # high confidence — don't suppress
+
+            has_neighbor = False
+            for j in range(n):
+                if i == j:
+                    continue
+                if near(i, j):
+                    has_neighbor = True
+                    break
+            if not has_neighbor:
+                adjustments[i] = min(adjustments[i],
+                                     adjustments[i] - COOCCUR_SUPPRESS_ISOLATED)
+
+        # Apply adjustments
+        result = []
+        for i, (x1, y1, x2, y2, conf, cls_id) in enumerate(dets):
+            new_conf = min(1.0, max(0.0, conf + adjustments[i]))
+            if new_conf >= self.conf_threshold:
+                result.append((x1, y1, x2, y2, new_conf, cls_id))
+        return result
+
+    # ─── Main inference ──────────────────────────────────────────────
+
+    def _infer_single(self, image_bgr: ndarray) -> list[BoundingBox]:
+        inp, (orig_h, orig_w) = self._preprocess(image_bgr)
+        out = self.session.run(None, {self.input_name: inp})[0]
+
+        # Decode based on detected output format
+        if self._end2end:
+            dets = self._decode_end2end(out, orig_h, orig_w)
+        else:
+            dets = self._decode_raw(out, orig_h, orig_w)
+
+        if not dets:
+            return []
+
+        # Post-processing pipeline
+        dets = self._geometric_validate(dets, orig_h, orig_w)
+        dets = self._spatial_cooccurrence(dets, orig_h, orig_w)
+
+        # Convert to BoundingBox
+        out_boxes = []
+        for x1, y1, x2, y2, conf, cls_id in 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)),
+                )
+            )
+        return out_boxes
+
+    def predict_batch(
+        self,
+        batch_images: list[ndarray],
+        offset: int,
+        n_keypoints: int,
+    ) -> list[TVFrameResult]:
+        results = []
+        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,
+                )
+            )
+        return results

model_type.json

@@ -0,0 +1,4 @@
+{
+  "task_type": "object-detection",
+  "model_type": "yolov11-nano"
+}

weights.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:68cbfb0b6c19ae57ba3b724eb4380cdfb61b4558f2aa8722b6f3555ef0b267a2
+size 19155617