configuration - Feilian-3D Documentation • Green

9. Configuration System

9.1 config.py Structure

# Data paths
DATA_ROOT = Path("data")
WIND_SPEED_DIR = DATA_ROOT / "wind_speed_filled"
MASK_DIR = DATA_ROOT / "mask_buildings"
 
# Model config
MODEL_CONFIG = {
    "in_channels": 1,
    "out_channels": 1,
    "base_channels": 32,
    "dropout": 0.1,
    "use_batchnorm": True,  # Ignored, using GroupNorm
}
 
# Training config
TRAINING_CONFIG = {
    "epochs": 100,
    "batch_size": 1,
    "learning_rate": 1e-3,
    "weight_decay": 1e-4,
    "val_split": 0.2,
    "use_gradient_loss": True,
    "gradient_loss_weight": 0.1,
    ...
}
 
# Phase-based optimizations
PHASE3_CONFIG = {
    "use_charbonnier_loss": True,
    "improved_gradient_loss": True,
    "use_augmentation": True,
    ...
}

9.2 CLI Overrides

python train.py \
    --epochs 50 \              # Override TRAINING_CONFIG["epochs"]
    --batch-size 2 \
    --base-channels 16 \
    --lr 5e-4 \
    --use-scheduler \          # Boolean flag
    --checkpoint-dir my_ckpts

9.3 Hyperparameter Ranges

Parameter	Default	Range	Notes
`epochs`	100	30-200	Early stopping usually kicks in
`batch_size`	1	1-4	Larger → OOM, gradient accumulation as alternative
`base_channels`	32	16-64	16=fast/small, 32=balanced, 64=slow/large
`lr`	1e-3	1e-4 to 1e-2	With OneCycleLR, max_lr
`weight_decay`	1e-4	1e-5 to 1e-3	AdamW decoupled weight decay
`gradient_weight`	0.1	0-1	λ for gradient loss
`clip_grad`	1.0	0-5	0=disabled, 1=safe, 5=permissive
`patience`	20	10-50	Early stopping patience