yolov8训练及测试（ubuntu18.04、tensorrt、ros)
1 数据集制作
1.1标注数据
Linux/Ubuntu/Mac
至少需要 Python 2.6 （推荐使用 Python 3 或更高版本 及 PyQt5）
Ubuntu Linux (Python 3 + Qt5)
git clone https://gitcode.com/gh_mirrors/la/labelImg.git
sudo apt-get install pyqt5-dev-tools
cd labelImg
sudo pip3 install -r requirements/requirements-linux-python3.txt
make qt5py3
python3 labelImg.py
运行python3 labelImg.py出错， File "/home/wyh/environment_setting/labelImg-master/libs/labelDialog.py", line 37, in __init__ layout.addWidget(bb, alignment=Qt.AlignmentFlag.AlignLeft) AttributeError: type object 'AlignmentFlag' has no attribute 'AlignLeft'
原因：因为 PyQt 或 PySide 的版本问题
解决：如果确定用的时PYQT5，将layout.addWidget(bb, alignment=Qt.AlignmentFlag.AlignLeft)更改为layout.addWidget(bb, alignment=Qt.AlignLeft)
1.2 建立对应的数据文件夹
images:图片数据，labels:标注转换后的yolo的txt文件，xmls:labelimg标注的xml格式数据，class.txt:标签txt文件
1.3 将标注后的xml转为txt
#! /usr/local/bin/ python
# -*- coding: utf-8 -*-
# .xml文件转换成.txt文件import copy
from xml.etree import Element, SubElement, tostring, ElementTree
import xml.etree.ElementTree as ET
import pickle
import os
from os import listdir, getcwd
from os.path import join# 检测目标的类别
classes = ["ore carrier", "passenger ship","container ship", "bulk cargo carrier","general cargo ship", "fishing boat"]CURRENT_DIR = os.path.dirname(os.path.abspath(__file__))def convert(size, box):dw = 1. / size[0]dh = 1. / size[1]x = (box[0] + box[1]) / 2.0 # (x_min + x_max) / 2.0y = (box[2] + box[3]) / 2.0 # (y_min + y_max) / 2.0w = box[1] - box[0] # x_max - x_minh = box[3] - box[2] # y_max - y_minx = x * dww = w * dwy = y * dhh = h * dhreturn (x, y, w, h)def convert_annotation(image_id):# .xml格式文件的地址in_file = open('地址1\%s.xml' % (image_id), encoding='UTF-8')# 生成的.txt格式文件的地址out_file = open('地址2\%s.txt' % (image_id), 'w')tree = ET.parse(in_file)root = tree.getroot()size = root.find('size')w = int(size.find('width').text)h = int(size.find('height').text)for obj in root.iter('object'):cls = obj.find('name').textif cls not in classes:continuecls_id = classes.index(cls)xmlbox = obj.find('bndbox')b = (float(xmlbox.find('xmin').text), float(xmlbox.find('xmax').text), float(xmlbox.find('ymin').text),float(xmlbox.find('ymax').text))bb = convert((w, h), b)out_file.write(str(cls_id) + " " + " ".join([str(a) for a in bb]) + '\n')# .xml格式文件的地址
xml_path = os.path.join(CURRENT_DIR, '地址1/')# xml列表
img_xmls = os.listdir(xml_path)
for img_xml in img_xmls:label_name = img_xml.split('.')[0]print(label_name)convert_annotation(label_name)
将代码中路径更改为对应的路径
2 将yolo数据拆分为train、val、test
import os
import random
import shutildef split_dataset(images_dir, labels_dir, output_dir, split_ratio=(0.8, 0.1, 0.1)):"""将图像和标签数据集划分为训练集、验证集和测试集。:param images_dir: 图像文件夹路径:param labels_dir: 标签文件夹路径:param output_dir: 输出目录路径:param split_ratio: 划分比例 (train, val, test)"""# 确保输出目录存在os.makedirs(output_dir, exist_ok=True)for subdir in ['train', 'val', 'test']:os.makedirs(os.path.join(output_dir, subdir, 'images'), exist_ok=True)os.makedirs(os.path.join(output_dir, subdir, 'labels'), exist_ok=True)# 获取所有图像文件名images = [f for f in os.listdir(images_dir) if f.endswith('.jpg') or f.endswith('.png')]labels = [f.replace('.jpg', '.txt').replace('.png', '.txt') for f in images]# 打乱顺序combined = list(zip(images, labels))random.shuffle(combined)images[:], labels[:] = zip(*combined)# 计算划分点num_train = int(len(images) * split_ratio[0])num_val = int(len(images) * split_ratio[1])# 划分数据集for i, image in enumerate(images):label = labels[i]if i < num_train:subset = 'train'elif i < num_train + num_val:subset = 'val'else:subset = 'test'shutil.copy(os.path.join(images_dir, image), os.path.join(output_dir, subset, 'images', image))shutil.copy(os.path.join(labels_dir, label), os.path.join(output_dir, subset, 'labels', label))# 示例调用
split_dataset('/home/wyh/artrc_catkin/src/artrc_yolov8/datasets/origin_data/images','/home/wyh/artrc_catkin/src/artrc_yolov8/datasets/origin_data/labels','/home/wyh/artrc_catkin/src/artrc_yolov8/datasets/split_data')
运行后如图所示
3 根据数据集添加yaml文件
import yaml
import os
def create_yaml(output_dir, train_dir, val_dir, test_dir, class_names, num_classes):"""创建 YOLOv8 数据集配置文件。:param output_dir: 输出目录路径:param train_dir: 训练集目录路径:param val_dir: 验证集目录路径:param test_dir: 测试集目录路径:param class_names: 类别名称列表:param num_classes: 类别数量"""data = {'train': train_dir,'val': val_dir,'test': test_dir,'nc': num_classes,'names': class_names}with open(os.path.join(output_dir, 'dataset.yaml'), 'w') as f:yaml.dump(data, f, default_flow_style=False)# 示例调用
create_yaml('/home/wyh/artrc_catkin/src/artrc_yolov8/datasets/split_data','/home/wyh/artrc_catkin/src/artrc_yolov8/datasets/split_data/train/images','/home/wyh/artrc_catkin/src/artrc_yolov8/datasets/split_data/val/images','/home/wyh/artrc_catkin/src/artrc_yolov8/datasets/split_data/test/images',['corrosion','craze', 'hide_craze','surface_attach','surface_corrosion','surface_eye','surface_injure','surface_oil','thunderstrike'], 9)
运行结果如下文件：
4 训练数据集
cd ultralytics
yolo task=detect mode=train model=yolov8n.pt data=ultralytics/cfg/datasets/dataset.yaml batch=8 epochs=200 imgsz=640 workers=32 device=0
5 训练后使用
5.1 训练后的各中形式数据转换
5.1.1 将.pt转换为onnx
方式一：利用下述pt_to_onnx.py进行转换
#! /usr/local/bin/ python
# -*- coding: utf-8 -*-
from ultralytics import YOLOmodel = YOLO("best.pt")success = model.export(format="onnx", half=False, dynamic=True, opset=17)print("demo")
cd ultralytics
python pt_to_onnx.py
方式二：命令行操作转换
cd ultralytics
setconda
conda activate yolov8
yolo mode=export model=yolov8n.pt format=onnx dynamic=True #simplify=True
yolo mode=export model=yolov8s.pt format=onnx dynamic=True # 不同模型
5.1.2将.onnx转换为.trt
cd /environment_setting/tensorrt-alpha/data/yolov8
# 生成trt文件
# 640 ../../../TensorRT-8.4.1.5/bin/trtexec为各路径，根据实际情况填写
../../../TensorRT-8.4.1.5/bin/trtexec --onnx=best.onnx --saveEngine=best.trt --buildOnly --minShapes=images:1x3x640x640 --optShapes=images:4x3x640x640 --maxShapes=images:8x3x640x640
../../../TensorRT-8.4.1.5/bin/trtexec --onnx=yolov8s.onnx --saveEngine=yolov8s.trt --buildOnly --minShapes=images:1x3x640x640 --optShapes=images:4x3x640x640 --maxShapes=images:8x3x640x640
../../../TensorRT-8.4.1.5/bin/trtexec --onnx=yolov8m.onnx --saveEngine=yolov8m.trt --buildOnly --minShapes=images:1x3x640x640 --optShapes=images:4x3x640x640 --maxShapes=images:8x3x640x640
5.2 利用pt文件进行检测
#!/home/wyh/.conda/envs/yolov8/bin/python3.8
# -*- coding: utf-8 -*-
import cv2
import torch
import rospy
import numpy as np
from ultralytics import YOLO
from time import time
from std_msgs.msg import Header
from sensor_msgs.msg import Image
from artrc_yolov8.msg import BoundingBox, BoundingBoxesclass Yolo_Dect:def __init__(self):# load parametersweight_path = rospy.get_param('~weight_path', '')image_topic = rospy.get_param('~image_topic', '/camera/color/image_raw')pub_topic = rospy.get_param('~pub_topic', '/yolov8/BoundingBoxes')self.camera_frame = rospy.get_param('~camera_frame', '')conf = rospy.get_param('~conf', '0.5')self.visualize = rospy.get_param('~visualize', 'True')# which device will be usedif (rospy.get_param('/use_cpu', 'true')):self.device = 'cpu'else:self.device = 'cuda'self.model = YOLO(weight_path)self.model.fuse()self.model.conf = confself.color_image = Image()self.getImageStatus = False# Load class colorself.classes_colors = {}# image subscribeself.color_sub = rospy.Subscriber(image_topic, Image, self.image_callback,queue_size=1, buff_size=52428800)# output publishersself.position_pub = rospy.Publisher(pub_topic, BoundingBoxes, queue_size=1)self.image_pub = rospy.Publisher('/yolov8/detection_image', Image, queue_size=1)# Load image and detectself.load_and_detect()def image_callback(self, image):# Existing image callback logicpassdef load_and_detect(self):# Load image from file or a specific sourceimage_path = '/home/wyh/artrc_catkin/src/artrc_yolov8/image/60.jpg' # Replace with your image pathself.color_image = cv2.imread(image_path)if self.color_image is None:rospy.logerr("Failed to load image from path: %s", image_path)returnself.color_image = cv2.cvtColor(self.color_image, cv2.COLOR_BGR2RGB)results = self.model(self.color_image, show=False, conf=0.3)self.dectshow(results, self.color_image.shape[0], self.color_image.shape[1])cv2.waitKey(3)def dectshow(self, results, height, width):# Existing detection logicself.frame = results[0].plot()print(str(results[0].speed['inference']))fps = 1000.0 / results[0].speed['inference']cv2.putText(self.frame, f'FPS: {int(fps)}', (20, 50), cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 255, 0), 2, cv2.LINE_AA)self.boundingBoxes = BoundingBoxes()self.boundingBoxes.header = Header(stamp=rospy.Time.now())self.boundingBoxes.image_header = Header(stamp=rospy.Time.now())# 统计数量class_count = {}total_count = 0for result in results[0].boxes:boundingBox = BoundingBox()boundingBox.xmin = np.int64(result.xyxy[0][0].item())boundingBox.ymin = np.int64(result.xyxy[0][1].item())boundingBox.xmax = np.int64(result.xyxy[0][2].item())boundingBox.ymax = np.int64(result.xyxy[0][3].item())boundingBox.Class = results[0].names[result.cls.item()]boundingBox.probability = result.conf.item()self.boundingBoxes.bounding_boxes.append(boundingBox)if boundingBox.Class in class_count:class_count[boundingBox.Class] += 1else:class_count[boundingBox.Class] = 1total_count += 1print("cl:",boundingBox.Class)self.position_pub.publish(self.boundingBoxes)self.publish_image(self.frame, height, width)print("data",self.boundingBoxes)print("Class Count:", class_count)print("total count:",total_count)# if self.visualize:# cv2.imshow('YOLOv8', self.frame)def publish_image(self, imgdata, height, width):image_temp = Image()header = Header(stamp=rospy.Time.now())header.frame_id = self.camera_frameimage_temp.height = heightimage_temp.width = widthimage_temp.encoding = 'bgr8'image_temp.data = np.array(imgdata).tobytes()image_temp.header = headerimage_temp.step = width * 3self.image_pub.publish(image_temp)def main():rospy.init_node('yolov8_ros', anonymous=True)yolo_dect = Yolo_Dect()rospy.spin()if __name__ == "__main__":main()
5.3 利用.onnx文件进行检测
#!/home/wyh/.conda/envs/yolov8/bin/python3.8
# -*- coding: utf-8 -*-
import onnxruntime as rt
import numpy as np
import cv2
import matplotlib.pyplot as plt# 定义类别标签
CLASS_NAMES = ['corrosion','craze', 'hide_craze','surface_attach','surface_corrosion','surface_eye','surface_injure','surface_oil','thunderstrike'] # 请根据你的模型定义实际的类标签COLOR_MAP = {"label_0": (255, 0, 0), # 红色"label_1": (0, 255, 0), # 绿色"label_2": (0, 0, 255), # 蓝色"label_3": (255, 255, 0), # 黄色"label_4": (255, 0, 255), # 品红色"label_5": (0, 255, 255), # 青色"label_6": (128, 0, 128), # 紫色"label_7": (255, 165, 0), # 橙色"label_8": (128, 128, 128), # 灰色
}def nms(pred, conf_thres, iou_thres): conf = pred[..., 4] > conf_thresbox = pred[conf == True] cls_conf = box[..., 5:]cls = []for i in range(len(cls_conf)):cls.append(int(np.argmax(cls_conf[i])))total_cls = list(set(cls)) output_box = [] for i in range(len(total_cls)):clss = total_cls[i] cls_box = []for j in range(len(cls)):if cls[j] == clss:box[j][5] = clsscls_box.append(box[j][:6])cls_box = np.array(cls_box)box_conf = cls_box[..., 4] box_conf_sort = np.argsort(box_conf) max_conf_box = cls_box[box_conf_sort[len(box_conf) - 1]]output_box.append(max_conf_box) cls_box = np.delete(cls_box, 0, 0) while len(cls_box) > 0:max_conf_box = output_box[len(output_box) - 1] del_index = []for j in range(len(cls_box)):current_box = cls_box[j] interArea = getInter(max_conf_box, current_box) iou = getIou(max_conf_box, current_box, interArea) if iou > iou_thres:del_index.append(j) cls_box = np.delete(cls_box, del_index, 0) if len(cls_box) > 0:output_box.append(cls_box[0])cls_box = np.delete(cls_box, 0, 0)return output_boxdef getIou(box1, box2, inter_area):box1_area = box1[2] * box1[3]box2_area = box2[2] * box2[3]union = box1_area + box2_area - inter_areaiou = inter_area / unionreturn ioudef getInter(box1, box2):box1_x1, box1_y1, box1_x2, box1_y2 = box1[0] - box1[2] / 2, box1[1] - box1[3] / 2, \box1[0] + box1[2] / 2, box1[1] + box1[3] / 2box2_x1, box2_y1, box2_x2, box2_y2 = box2[0] - box2[2] / 2, box2[1] - box1[3] / 2, \box2[0] + box2[2] / 2, box2[1] + box2[3] / 2if box1_x1 > box2_x2 or box1_x2 < box2_x1:return 0if box1_y1 > box2_y2 or box1_y2 < box2_y1:return 0x_list = [box1_x1, box1_x2, box2_x1, box2_x2]x_list = np.sort(x_list)x_inter = x_list[2] - x_list[1]y_list = [box1_y1, box1_y2, box2_y1, box2_y2]y_list = np.sort(y_list)y_inter = y_list[2] - y_list[1]inter = x_inter * y_interreturn inter# 画框并添加标签
def draw(img, xscale, yscale, pred):img_ = img.copy()if len(pred):for detect in pred:label = int(detect[5]) # 获取类别标签label_name = CLASS_NAMES[label] # 通过类索引获取类名detect_coords = [int((detect[0] - detect[2] / 2) * xscale), int((detect[1] - detect[3] / 2) * yscale),int((detect[0] + detect[2] / 2) * xscale), int((detect[1] + detect[3] / 2) * yscale)]# 获取颜色，如果没有对应的颜色，就使用默认颜色color = COLOR_MAP.get(label_name, (255, 255, 255)) # 默认为白色# 绘制矩形框img_ = cv2.rectangle(img_, (detect_coords[0], detect_coords[1]), (detect_coords[2], detect_coords[3]), color, 2)# 绘制标签img_ = cv2.putText(img_, label_name, (detect_coords[0], detect_coords[1]-5), cv2.FONT_HERSHEY_SIMPLEX, 0.5, color, 1)return img_if __name__ == '__main__':height, width = 640, 640img0 = cv2.imread('/home/wyh/artrc_catkin/src/artrc_yolov8/image/60.jpg')x_scale = img0.shape[1] / widthy_scale = img0.shape[0] / heightimg = img0 / 255.img = cv2.resize(img, (width, height))img = np.transpose(img, (2, 0, 1))data = np.expand_dims(img, axis=0)sess = rt.InferenceSession('/home/wyh/artrc_catkin/src/artrc_yolov8/weights/best.onnx')input_name = sess.get_inputs()[0].namelabel_name = sess.get_outputs()[0].namepred = sess.run([label_name], {input_name: data.astype(np.float32)})[0]pred = np.squeeze(pred)pred = np.transpose(pred, (1, 0))pred_class = pred[..., 4:]pred_conf = np.max(pred_class, axis=-1)pred = np.insert(pred, 4, pred_conf, axis=-1)result = nms(pred, 0.3, 0.45)ret_img = draw(img0, x_scale, y_scale, result)# 使用OpenCV显示图像cv2.imshow('Detection Result', ret_img)cv2.waitKey(0) # 等待按键事件cv2.destroyAllWindows() # 关闭所有OpenCV窗口