yolo/yolov5.py

import torch
import numpy

from towhee import register
from towhee.operator import NNOperator

@register(output_schema=['boxes', 'classes', 'scores'])
class Yolov5(NNOperator):
    def __init__(self):
        super().__init__()
        model_path = str(Path(__file__).parent / 'models/yolov5s.pt')
        self._model = torch.hub.load('ultralytics/yolov5', 'custom', model_path)
        # self._model = torch.hub.load('ultralytics/yolov5', 'yolov5s', pretrained=True)

    def __call__(self, img: numpy.ndarray):
        # Get object detection results with YOLOv5 model
        results = self._model(img)

        boxes = [re[0:4] for re in results.xyxy[0]]
        boxes = [list(map(int, box)) for box in boxes]
        classes = list(results.pandas().xyxy[0].name)
        scores = list(results.pandas().xyxy[0].confidence)
        return [(b, c, s) for b, c, s in zip(boxes, classes, scores)]