dgwon/yolov8s-coco2017-vehicle-detection

# These are the exact same training arguments as used for yolov8m for fair comparison
results = model.train(
    data=yaml_path,
    epochs=100,           # increase for better results
    imgsz=640,
    batch=96,            # let ultralytics decide optimal
    device=0,            # 0 = first GPU
    project=wandb_project_name,
    name=run_name,
    save=True,
    save_period=10,          # checkpoint every 10 epochs
    patience=10,         # early stopping
    workers=16,           # Colab has limited CPU cores

    optimizer="AdamW",
    lr0=3.5e-3,
    lrf=0.05,                # final lr = lr0 * lrf
    cos_lr=True,
    warmup_epochs=3,
    weight_decay=5e-4,
    amp=True,                # automatic mixed precision (BF16 on A100) — free speedup

    close_mosaic=10,
    mosaic=1.0,
    mixup=0.1,
    copy_paste=0.1,          # good for occluded vehicle scenarios

    # Augmentation (helps generalize to varied vehicle appearances)
    fliplr=0.5,
    flipud=0.0,
    hsv_h=0.02,
    hsv_s=0.7,
    hsv_v=0.4,
)