kaggle中的2D目标检测训练trick总结

数据增强

• 主要是mixup和cutmix，详见https://docs.ultralytics.com/zh/guides/yolo-data-augmentation/#copy-paste-mode-copy_paste_mode

你可以在yolo中这样使用

model = YOLO("yolo11m.pt")
data_yaml = "/kaggle/working/yolo_params.yaml"

model.train(
    data=data_yaml,
    epochs=10,                
    batch=4,                   
    imgsz=1056,
    patience=150,               
    optimizer='SGD',
    momentum=0.937,          
    lr0=0.001,                
    weight_decay=0.0005,       
    cos_lr=True,               
    save_period=1,             
    workers=8,
    # Augmentations
    close_mosaic=20,
    hsv_h=0.015,
    hsv_s=0.7,
    hsv_v=0.4,
    flipud=0.5,
    fliplr=0.5,
    translate=0.1,
    scale=0.5,
    shear=0.01
)

模型训练

• self-supervised learning

后处理

• Weighted Boxes Fusion