experiment:
  name: a6_full
  output_root: outputs/dsifn/a6_full
  seed: 42
hardware:
  device: cuda:1
  gpu_ids:
  - 1
  mixed_precision: true
efficiency:
  amp: true
  amp_dtype: fp16
  channels_last: false
  gradient_checkpointing: false
  compile: false
  compile_mode: reduce-overhead
  fast_mode: false
  channels_multiplier: 1.0
profiling:
  enabled: false
  warmup_iters: 20
  profile_iters: 100
  log_interval: 10
dataloader:
  num_workers: 8
  pin_memory: true
  persistent_workers: true
  prefetch_factor: 4
  drop_last: true
logging:
  train_metrics_every_iter: false
  save_visualizations: false
  log_interval: 20
optimizer:
  grad_clip_norm: 0.5
dataset:
  name: DSIFN-CD
  root: /storage2/ChangeDetection/MV/Datasets/DSIFN-CD/DSIFN
  mode: binary
  task_type: binary_change
  image_size: 256
  num_workers: 8
  pin_memory: true
  persistent_workers: true
  prefetch_factor: 4
  val_ratio: 0.2
  augmentation: true
  batch_size: 8
  augmentation_ops:
  - horizontal_flip
  - vertical_flip
  cache_images_in_ram: false
  cache_masks_in_ram: false
model:
  mode: dual
  backbone: mambavision
  variant: small
  decoder: adaptive_rf
  pretrained: true
  output_mode: binary
  num_classes: 1
  freeze_backbone: false
  drbi:
    enabled: true
  signed_diff:
    enabled: true
  cram_lite:
    enabled: true
    alpha: 0.5
    apply_stages:
    - 0
    - 1
    - 2
    attention_type: spatial
  arf_fpn:
    enabled: true
  boundary_refine:
    enabled: true
  enable_semantic_heads: false
difference:
  enabled: true
  type: region_boundary
  out_channels: 256
  use_depthwise: true
  gate_hidden_ratio: 0.25
  pre_norm: true
  use_absdiff: true
  use_signed_diff: true
  use_product: false
  product_scale: 0.25
  use_region_gate: true
  use_boundary_gate: true
  region_gate_min: 0.2
  region_gate_max: 0.8
  boundary_gate_min: 0.0
  boundary_gate_max: 0.4
  return_debug: false
decoder:
  type: adaptive_rf
  channels: 256
  dilation_rates:
  - 1
  - 2
  - 4
  - 8
  use_boundary_residual: true
  residual_scale: 0.1
  use_depthwise: true
  aux_weight: 0.4
training:
  max_iterations: 50000
  batch_size: 8
  lr: 5.0e-05
  optimizer: AdamW
  scheduler: cosine
  warmup_iterations: 2500
  validate_every: 5000
  log_every: 20
  gradient_clip: 0.5
  weight_decay: 0.01
  skip_nan_steps: true
  nan_diag_every: 50
  non_blocking_transfer: true
  gradient_checkpointing: false
  use_ema: true
  ema_decay: 0.999
  overwrite_output_dir: false
  allow_resume_for_ablation: false
validation:
  batch_size: 4
  save_samples: true
  sample_count: 16
  split: val
evaluation:
  split: test
  threshold: 0.5
  threshold_sweep:
    enabled: true
    values:
    - 0.3
    - 0.35
    - 0.4
    - 0.45
    - 0.5
    - 0.55
    - 0.6
  threshold_select_metric: F1_1
  use_ema: true
  inference_mode: patch
  crop_size: 256
  overlap: 0.25
  use_tta: false
  save_predictions: false
  save_visualizations: false
  memory_efficient: true
  threshold_list:
  - 0.3
  - 0.35
  - 0.4
  - 0.45
  - 0.5
  - 0.55
  - 0.6
  tta_augmentations:
  - original
  - hflip
  - vflip
  - rot90
  use_cached_predictions: false
loss:
  type: bce_dice
  final:
    type: bce_dice
    bce_weight: 1.0
    dice_weight: 1.0
  coarse:
    enabled: true
    weight: 0.4
  boundary:
    enabled: true
    weight: 0.1
    type: bce_dice
    target_type: sobel
    edge_width: 3
  bce_weight: 1.0
  dice_weight: 1.0
  focal_weight: 0.0
  boundary_weight: 0.1
  focal_gamma: 1.5
metrics:
  allowed:
  - Pre
  - Rec
  - F1
  - IoU
  - OA
  average_mode: global
ema:
  enabled: true
  decay: 0.999
  store_on_cpu: true
resume:
  enabled: false
  checkpoint_path: null
  strict: true
checkpoint:
  path: null
  monitor: f1
  mode: max
  save_best_only: true
  selection_metric: F1
  save_best: true
  save_last: true
  save_every: null
  save_latest: true
  latest_every: 5000
boundary_metrics:
  enabled: true
  boundary_width: 3
  tolerance: 2
debug:
  enabled: false
  output_root: debug
  ablation_trace: false
  name: memory_debug
  steps: 3
  batch_size: 8
  image_size: 256
  use_amp: true
  profile_torch_ops: true
  profile_one_step_only: true
  compare_modes:
  - baseline_forward_only
  - train_forward_loss_backward
  - train_with_return_features_false
  - train_with_return_features_true
  - train_with_ema_disabled
  - train_with_ema_enabled
  - train_with_tta_disabled
  - eval_no_grad
  - eval_with_tta
  - decoder_baseline
  - decoder_refinement
  - decoder_adaptive_rf
  save_memory_summary: true
  seed: 42
task: binary_cd
eval:
  split: test
  threshold: 0.5
  threshold_sweep:
    enabled: true
    values:
    - 0.3
    - 0.35
    - 0.4
    - 0.45
    - 0.5
    - 0.55
    - 0.6
  threshold_select_metric: F1_1
  use_ema: true
  inference_mode: patch
  crop_size: 256
  overlap: 0.25
  use_tta: false
  save_predictions: false
  save_visualizations: false
  memory_efficient: true
  threshold_list:
  - 0.3
  - 0.35
  - 0.4
  - 0.45
  - 0.5
  - 0.55
  - 0.6
  tta_augmentations:
  - original
  - hflip
  - vflip
  - rot90
  use_cached_predictions: false
post_training: {}
benchmark:
  model_name: a6_full
  output_dir: outputs/benchmark_runs/summary
  datasets:
  - DSIFN-CD
  - WHU-CD
  eval_split: test
  main_dataset: DSIFN-CD
  checkpoints: {}
_meta:
  config_path: /storage2/ChangeDetection/MV/MambaRefine-CD/configs/ablations/dsifn/a6_full.yaml
  config_fingerprint: a38e4e9372c18e99