一、數據處理

import traceback
import xml.etree.ElementTree as ET
import os
import shutil
import random
import cv2
import numpy as np
from tqdm import tqdmdef convert_annotation_to_list(xml_filepath, size_width, size_height, classes):in_file = open(xml_filepath, encoding='UTF-8')tree = ET.parse(in_file)root = tree.getroot()# size = root.find('size')# size_width = int(size.find('width').text)# size_height = int(size.find('height').text)yolo_annotations = []# if size_width == 0 or size_height == 0:for obj in root.iter('object'):difficult = obj.find('difficult').textcls = obj.find('name').textif cls not in classes or int(difficult) == 1: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)]# 標注越界修正if b[1] > size_width:b[1] = size_widthif b[3] > size_height:b[3] = size_heighttxt_data = [((b[0] + b[1]) / 2.0) / size_width, ((b[2] + b[3]) / 2.0) / size_height,(b[1] - b[0]) / size_width, (b[3] - b[2]) / size_height]# 標注越界修正if txt_data[0] > 1:txt_data[0] = 1if txt_data[1] > 1:txt_data[1] = 1if txt_data[2] > 1:txt_data[2] = 1if txt_data[3] > 1:txt_data[3] = 1yolo_annotations.append(f"{cls_id} {' '.join([str(round(a, 6)) for a in txt_data])}")in_file.close()return yolo_annotationsdef main():classes = ["red", "green", "yellow", "off"]root = r"/ssd/xiedong/lightyolov5"img_path_1 = os.path.join(root, "Traffic-Lights-Dataset-Domestic/JPEGImages")xml_path_1 = os.path.join(root, "Traffic-Lights-Dataset-Domestic/Annotations")img_path_2 = os.path.join(root, "Traffic-Lights-Dataset-Foreign/JPEGImages")xml_path_2 = os.path.join(root, "Traffic-Lights-Dataset-Foreign/Annotations")dst_yolo_root = os.path.join(root, "Traffic-Lights-Dataset-YOLO")dst_yolo_root_img = os.path.join(dst_yolo_root, "images")os.makedirs(dst_yolo_root_img, exist_ok=True)dst_yolo_root_txt = os.path.join(dst_yolo_root, "labels")os.makedirs(dst_yolo_root_txt, exist_ok=True)index = 0img_path_1_files = os.listdir(img_path_1)xml_path_1_files = os.listdir(xml_path_1)for img_id in tqdm(img_path_1_files):# 右邊的.之前的部分xml_id = img_id.split(".")[0] + ".xml"if xml_id in xml_path_1_files:try:new_name = f"{index:06d}.jpg"img = cv2.imdecode(np.fromfile(os.path.join(img_path_1, img_id), dtype=np.uint8), 1)  # img是矩陣cv2.imwrite(os.path.join(dst_yolo_root_img, new_name), img)new_txt_name = f"{index:06d}.txt"yolo_annotations = convert_annotation_to_list(os.path.join(xml_path_1, img_id[:-4] + ".xml"),img.shape[1],img.shape[0],classes)with open(os.path.join(dst_yolo_root_txt, new_txt_name), 'w') as f:f.write('\n'.join(yolo_annotations))index += 1except:traceback.print_exc()img_path_1_files = os.listdir(img_path_2)xml_path_1_files = os.listdir(xml_path_2)for img_id in tqdm(img_path_1_files):# 右邊的.之前的部分xml_id = img_id.split(".")[0] + ".xml"if xml_id in xml_path_1_files:try:new_name = f"{index:06d}.jpg"img = cv2.imdecode(np.fromfile(os.path.join(img_path_2, img_id), dtype=np.uint8), 1)  # img是矩陣cv2.imwrite(os.path.join(dst_yolo_root_img, new_name), img)new_txt_name = f"{index:06d}.txt"yolo_annotations = convert_annotation_to_list(os.path.join(xml_path_2, img_id[:-4] + ".xml"),img.shape[1],img.shape[0],classes)with open(os.path.join(dst_yolo_root_txt, new_txt_name), 'w') as f:f.write('\n'.join(yolo_annotations))index += 1except:traceback.print_exc()if __name__ == '__main__':main()

二、環境

conda create -n py310_yolo8 python=3.10 -yconda activate py310_yolo8conda install pytorch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 pytorch-cuda=11.8 -c pytorch -c nvidiapip install ultralytics

data.yaml

path: /ssd/xiedong/lightyolov5/Traffic-Lights-Dataset-YOLO/
train: images
val: images
test: # test images (optional)# Classes
names:0: 'red'1: 'green'2: 'yellow'3: 'off'

三、訓練

教程：

https://docs.ultralytics.com/modes/train/#comet

新建訓練代碼文件train.py：

from ultralytics import YOLO# Load a model
model = YOLO("yolov8s.pt")  # load a pretrained model (recommended for training)# Train the model with 2 GPUs
results = model.train(data="data.yaml", epochs=100, imgsz=640, device=[0, 1, 2, 3], batch=128)

開啟訓練：

python -m torch.distributed.run --nproc_per_node 4 train.py

結果會存在這里：

訓練截圖：

數據分布：