scorevision: push artifact
Browse files
miner.py
ADDED
|
@@ -0,0 +1,149 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Turbo Vision miner for manak0/Detect-beverage-detect (cup / can / bottle).
|
| 2 |
+
|
| 3 |
+
CPU-only ONNX inference with 2 intra-op threads (the compliance checker measures
|
| 4 |
+
p95 latency on 2 vCPUs and re-runs the model to check that outputs are reproducible,
|
| 5 |
+
so everything here is deterministic numpy / onnxruntime CPU).
|
| 6 |
+
|
| 7 |
+
Repo layout:
|
| 8 |
+
miner.py this file
|
| 9 |
+
weights.onnx detector, input [1,3,S,S] RGB float 0..1, output [1, 4+3, N]
|
| 10 |
+
(cx, cy, w, h in input pixels, then per-class sigmoid scores)
|
| 11 |
+
config.json {"conf": [cup, can, bottle], "nms_iou": 0.5, "max_det": 100, "gate": 0.0}
|
| 12 |
+
|
| 13 |
+
No `from __future__ import annotations` here: pydantic must see real types when the
|
| 14 |
+
chute imports this file dynamically.
|
| 15 |
+
"""
|
| 16 |
+
import json
|
| 17 |
+
from pathlib import Path
|
| 18 |
+
|
| 19 |
+
import cv2
|
| 20 |
+
import numpy as np
|
| 21 |
+
import onnxruntime as ort
|
| 22 |
+
from numpy import ndarray
|
| 23 |
+
from pydantic import BaseModel
|
| 24 |
+
|
| 25 |
+
|
| 26 |
+
class BoundingBox(BaseModel):
|
| 27 |
+
x1: int
|
| 28 |
+
y1: int
|
| 29 |
+
x2: int
|
| 30 |
+
y2: int
|
| 31 |
+
cls_id: int
|
| 32 |
+
conf: float
|
| 33 |
+
|
| 34 |
+
|
| 35 |
+
class TVFrameResult(BaseModel):
|
| 36 |
+
frame_id: int
|
| 37 |
+
boxes: list[BoundingBox]
|
| 38 |
+
keypoints: list[tuple[int, int]]
|
| 39 |
+
|
| 40 |
+
|
| 41 |
+
def _nms(boxes: ndarray, scores: ndarray, iou: float) -> list[int]:
|
| 42 |
+
order = np.argsort(-scores, kind="stable")
|
| 43 |
+
keep: list[int] = []
|
| 44 |
+
area = (boxes[:, 2] - boxes[:, 0]) * (boxes[:, 3] - boxes[:, 1])
|
| 45 |
+
while order.size:
|
| 46 |
+
i = int(order[0])
|
| 47 |
+
keep.append(i)
|
| 48 |
+
rest = order[1:]
|
| 49 |
+
xx1 = np.maximum(boxes[i, 0], boxes[rest, 0])
|
| 50 |
+
yy1 = np.maximum(boxes[i, 1], boxes[rest, 1])
|
| 51 |
+
xx2 = np.minimum(boxes[i, 2], boxes[rest, 2])
|
| 52 |
+
yy2 = np.minimum(boxes[i, 3], boxes[rest, 3])
|
| 53 |
+
inter = np.maximum(0, xx2 - xx1) * np.maximum(0, yy2 - yy1)
|
| 54 |
+
ov = inter / (area[i] + area[rest] - inter + 1e-7)
|
| 55 |
+
order = rest[ov <= iou]
|
| 56 |
+
return keep
|
| 57 |
+
|
| 58 |
+
|
| 59 |
+
class Miner:
|
| 60 |
+
def __init__(self, path_hf_repo: Path) -> None:
|
| 61 |
+
path_hf_repo = Path(path_hf_repo)
|
| 62 |
+
cfg_path = path_hf_repo / "config.json"
|
| 63 |
+
self.cfg = json.loads(cfg_path.read_text()) if cfg_path.exists() else {}
|
| 64 |
+
self.conf = np.array(self.cfg.get("conf", [0.4, 0.4, 0.4]), dtype=np.float32)
|
| 65 |
+
self.nms_iou = float(self.cfg.get("nms_iou", 0.5))
|
| 66 |
+
self.max_det = int(self.cfg.get("max_det", 100))
|
| 67 |
+
self.gate = float(self.cfg.get("gate", 0.0))
|
| 68 |
+
# per class: if no box passes conf, keep the class's best candidate when it scores >= rescue
|
| 69 |
+
# (1.0 = off). map50 averages over the classes in an image, so the first box of a class is worth most.
|
| 70 |
+
self.rescue = np.array(self.cfg.get("rescue", [1.0, 1.0, 1.0]), dtype=np.float32)
|
| 71 |
+
so = ort.SessionOptions()
|
| 72 |
+
so.intra_op_num_threads = 2
|
| 73 |
+
so.inter_op_num_threads = 1
|
| 74 |
+
so.execution_mode = ort.ExecutionMode.ORT_SEQUENTIAL
|
| 75 |
+
so.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
|
| 76 |
+
self.session = ort.InferenceSession(str(path_hf_repo / "weights.onnx"), so, providers=["CPUExecutionProvider"])
|
| 77 |
+
self.input_name = self.session.get_inputs()[0].name
|
| 78 |
+
self.size = int(self.session.get_inputs()[0].shape[-1])
|
| 79 |
+
self.buf = np.empty((1, 3, self.size, self.size), dtype=np.float32)
|
| 80 |
+
cv2.setNumThreads(1)
|
| 81 |
+
warm = np.zeros((1024, 1024, 3), dtype=np.uint8)
|
| 82 |
+
for _ in range(2):
|
| 83 |
+
self._raw(warm)
|
| 84 |
+
|
| 85 |
+
def __repr__(self) -> str:
|
| 86 |
+
return f"ONNX P3 detector {self.size}px conf={self.conf.tolist()}"
|
| 87 |
+
|
| 88 |
+
def _raw(self, image: ndarray):
|
| 89 |
+
"""letterbox -> ONNX -> (boxes xyxy in original pixels, per-class scores [N,3])"""
|
| 90 |
+
h, w = image.shape[:2]
|
| 91 |
+
s = self.size
|
| 92 |
+
r = min(s / h, s / w)
|
| 93 |
+
nw, nh = int(round(w * r)), int(round(h * r))
|
| 94 |
+
top, left = (s - nh) // 2, (s - nw) // 2
|
| 95 |
+
img = cv2.resize(image, (nw, nh), interpolation=cv2.INTER_LINEAR) if (nw, nh) != (w, h) else image
|
| 96 |
+
self.buf.fill(114 / 255)
|
| 97 |
+
k = np.float32(1 / 255)
|
| 98 |
+
for c in range(3): # BGR -> RGB, HWC -> CHW, 0..1
|
| 99 |
+
np.multiply(img[:, :, 2 - c], k, out=self.buf[0, c, top:top + nh, left:left + nw], dtype=np.float32)
|
| 100 |
+
out = self.session.run(None, {self.input_name: self.buf})[0][0].T # [N, 4+nc]
|
| 101 |
+
xy, wh, sc = out[:, :2], out[:, 2:4], out[:, 4:]
|
| 102 |
+
boxes = np.concatenate([xy - wh / 2, xy + wh / 2], 1)
|
| 103 |
+
boxes[:, [0, 2]] = (boxes[:, [0, 2]] - left) / r
|
| 104 |
+
boxes[:, [1, 3]] = (boxes[:, [1, 3]] - top) / r
|
| 105 |
+
boxes[:, [0, 2]] = boxes[:, [0, 2]].clip(0, w - 1)
|
| 106 |
+
boxes[:, [1, 3]] = boxes[:, [1, 3]].clip(0, h - 1)
|
| 107 |
+
return boxes, sc
|
| 108 |
+
|
| 109 |
+
def candidates(self, image: ndarray, floor: float = 0.05):
|
| 110 |
+
"""All per-class NMS survivors above `floor` (used for threshold tuning)."""
|
| 111 |
+
boxes, sc = self._raw(image)
|
| 112 |
+
cls = sc.argmax(1)
|
| 113 |
+
conf = sc[np.arange(len(sc)), cls]
|
| 114 |
+
m = conf >= floor
|
| 115 |
+
boxes, conf, cls = boxes[m], conf[m], cls[m]
|
| 116 |
+
keep: list[int] = []
|
| 117 |
+
for c in range(sc.shape[1]):
|
| 118 |
+
idx = np.flatnonzero(cls == c)
|
| 119 |
+
if idx.size:
|
| 120 |
+
keep.extend(idx[_nms(boxes[idx], conf[idx], self.nms_iou)].tolist())
|
| 121 |
+
keep = sorted(keep, key=lambda i: -conf[i])
|
| 122 |
+
return boxes[keep], conf[keep], cls[keep]
|
| 123 |
+
|
| 124 |
+
def _detect(self, image: ndarray) -> list[BoundingBox]:
|
| 125 |
+
boxes, conf, cls = self.candidates(image, floor=float(min(self.conf.min(), self.rescue.min(), 0.05)))
|
| 126 |
+
if conf.size == 0 or float(conf.max()) < self.gate:
|
| 127 |
+
return []
|
| 128 |
+
m = conf >= self.conf[cls]
|
| 129 |
+
for c in range(len(self.rescue)):
|
| 130 |
+
if self.rescue[c] < 1.0 and not m[cls == c].any():
|
| 131 |
+
idx = np.flatnonzero((cls == c) & (conf >= self.rescue[c]))
|
| 132 |
+
if idx.size:
|
| 133 |
+
m[idx[0]] = True # candidates are sorted by score, so idx[0] is the class's best
|
| 134 |
+
boxes, conf, cls = boxes[m][: self.max_det], conf[m][: self.max_det], cls[m][: self.max_det]
|
| 135 |
+
q = np.round(boxes).astype(int)
|
| 136 |
+
ok = (q[:, 2] > q[:, 0]) & (q[:, 3] > q[:, 1])
|
| 137 |
+
return [BoundingBox(x1=int(a[0]), y1=int(a[1]), x2=int(a[2]), y2=int(a[3]), cls_id=int(c), conf=float(f))
|
| 138 |
+
for a, c, f in zip(q[ok], cls[ok], conf[ok])]
|
| 139 |
+
|
| 140 |
+
def predict_batch(self, batch_images: list[ndarray], offset: int, n_keypoints: int) -> list[TVFrameResult]:
|
| 141 |
+
results = []
|
| 142 |
+
for i, im in enumerate(batch_images):
|
| 143 |
+
try:
|
| 144 |
+
boxes = self._detect(im)
|
| 145 |
+
except Exception as e: # one bad frame must not fail the batch
|
| 146 |
+
print(f"frame {offset + i} failed: {e}")
|
| 147 |
+
boxes = []
|
| 148 |
+
results.append(TVFrameResult(frame_id=offset + i, boxes=boxes, keypoints=[(0, 0)] * n_keypoints))
|
| 149 |
+
return results
|