data:
  augment: true
  batch_size: 8
  block_buffer_deg: 0.02
  block_salt: cryosentinel-blocks-v1
  block_size_deg: 0.25
  copy_paste:
    min_donor_water_pixels: 32
    p: 0.3
    sub_sample_frac: 1.0
  data_dir: /data/multimodal_chips_v3
  hard_neg_sampler:
    positive_to_negative_ratio: 3.0
    replacement: true
  hard_neg_water_frac_threshold: 0.002
  include_unknown_in_train: true
  mosaic:
    jitter: 0.2
    p: 0.15
  multi_scale:
    p: 0.5
    scale_max: 1.25
    scale_min: 0.75
  norm_stats_path: /data/multimodal_chips_v3/dataset_stats_v3.json
  num_workers: 8
  persistent_workers: true
  pin_memory: true
  regions:
  - hindu_kush
  - pamir
  - karakoram
  - western_himalaya
  - central_himalaya
  - eastern_himalaya
  - hengduan_nyainqentanglha
  - tibetan_plateau
  - hma_other
  - tien_shan_full
  - ile_alatau
  - zhetysu_alatau
  rot90: false
  shard_cache_size: 4
  spectral_jitter:
    eps_dem: 0.02
    eps_s1: 0.1
    eps_s2: 0.05
    p: 0.5
  use_block_split: true
  years:
  - 2017
  - 2021
  - 2022
  - 2023
model:
  backbone: terramind_v1_large
  backbone_lr_mult: 0.1
  backbone_pretrained: true
  compile_model: true
  decoder: UperNetDecoder
  decoder_kwargs:
    decoder_channels: 256
    decoder_scale_modules: true
  dice_variant: generalized
  focal_alpha: 0.25
  focal_gamma: 2.0
  freeze_backbone_layers: 2
  label_smoothing: 0.1
  llrd_decay: 0.75
  loss_type: mega
  lovasz_per_image: true
  lr: 0.0001
  modalities:
  - S2L2A
  - S1GRD
  - DEM
  ohem_keep_ratio: 0.7
  ohem_min_kept: 4096
  pos_weight_max: 200.0
  test_tta: true
  tversky_alpha: 0.3
  tversky_beta: 0.7
  val_tta: false
  w_bce: 0.5
  w_boundary: 0.3
  w_dice: 1.0
  w_focal: 1.0
  w_lovasz: 0.3
  w_tversky: 0.5
  weight_decay: 0.0001
output_dir: /ckpts
run_name: terramind_v3_smoke
run_test: true
seed: 42
trainer:
  accelerator: gpu
  accumulate_grad_batches: 2
  check_val_every_n_epoch: 1
  devices: 1
  early_stopping_patience: 8
  ema:
    apply_at_test: true
    apply_at_validation: true
    cpu_shadow: true
    decay: 0.9995
    skip_buffers: true
  gradient_clip_val: 1.0
  log_every_n_steps: 25
  max_epochs: 30
  max_steps: 200
  precision: bf16-mixed