configs/config.yaml

# =============================================================================
# Military Base Change Detection — Master Configuration
# =============================================================================

# --- Project paths ---
project:
  name: "military-base-change-detection"
  seed: 42

# --- Colab / runtime settings ---
colab:
  enabled: false
  drive_root: "/content/drive/MyDrive/change-detection"
  checkpoint_dir: "/content/drive/MyDrive/change-detection/checkpoints"
  log_dir: "/content/drive/MyDrive/change-detection/logs"
  output_dir: "/content/drive/MyDrive/change-detection/outputs"
  data_dir: "/content/drive/MyDrive/change-detection/processed_data"

# --- Local paths (used when colab.enabled is false) ---
paths:
  raw_data: "./raw_data"
  processed_data: "./processed_data"
  checkpoint_dir: "./checkpoints"
  log_dir: "./logs"
  output_dir: "./outputs"

# --- Dataset ---
dataset:
  name: "levir-cd"                 # levir-cd | whu-cd
  original_size: 1024
  patch_size: 256
  num_workers: 4
  pin_memory: true
  # ImageNet normalization
  mean: [0.485, 0.456, 0.406]
  std: [0.229, 0.224, 0.225]

# --- Augmentation (train only) ---
augmentation:
  enabled: true
  horizontal_flip: 0.5
  vertical_flip: 0.5
  random_rotate_90: 0.5
  color_jitter:
    brightness: 0.2
    contrast: 0.2
    saturation: 0.1
    hue: 0.05

# --- Model selection ---
model:
  name: "unet_pp"                  # siamese_cnn | unet_pp | changeformer

# --- Model-specific configs ---
siamese_cnn:
  backbone: "resnet18"
  pretrained: true

unet_pp:
  encoder_name: "resnet34"
  pretrained: true
  deep_supervision: false

changeformer:
  embed_dims: [64, 128, 320, 512]  # MiT-B1 style
  num_heads: [1, 2, 5, 8]
  mlp_ratios: [8, 8, 4, 4]
  depths: [2, 2, 2, 2]
  pretrained_backbone: true

# --- Training ---
training:
  epochs: 100                      # 200 for changeformer
  optimizer: "adamw"
  learning_rate: 1.0e-4
  weight_decay: 0.01
  scheduler: "cosine"
  warmup_epochs: 5
  grad_clip_max_norm: 1.0
  gradient_accumulation_steps: 1   # set to 2 for changeformer on T4
  amp: true                        # mixed precision
  early_stopping:
    enabled: true
    patience: 15
    metric: "f1"
    mode: "max"
  log_interval: 10                 # log every N batches
  vis_interval: 5                  # visualize predictions every N epochs

# --- Loss ---
loss:
  name: "bce_dice"                 # bce_dice | focal
  bce_dice:
    bce_weight: 0.5
    dice_weight: 0.5
  focal:
    alpha: 0.25
    gamma: 2.0

# --- Evaluation ---
evaluation:
  threshold: 0.5
  metrics:
    - f1
    - iou
    - precision
    - recall
    - oa

# --- GPU-specific batch sizes (auto-detected on Colab) ---
# model_name -> { gpu_type -> batch_size }
batch_sizes:
  siamese_cnn:
    T4: 16
    V100: 16
    default: 8
  unet_pp:
    T4: 8
    V100: 12
    default: 4
  changeformer:
    T4: 4
    V100: 6
    default: 2

# --- Per-model learning rates ---
learning_rates:
  siamese_cnn: 1.0e-3
  unet_pp: 1.0e-4
  changeformer: 6.0e-5

# --- Per-model epoch counts ---
epoch_counts:
  siamese_cnn: 100
  unet_pp: 100
  changeformer: 200

# --- Gradio demo ---
gradio:
  server_port: 7860
  share: false
  default_model: "changeformer"
  default_checkpoint: "checkpoints/changeformer_best.pth"