# Base config. Shared optimization settings live here (the paper's shared HPs);
# per-task architecture + solver settings come from configs/experiment/*.yaml,
# selected with `+experiment=<name>`.
defaults:
  - _self_

task: maze            # maze | mnist | sudoku  (selects the task hook)
model_type: sheaf     # sheaf | mpnn
dtype: float32        # registered reproduction runs forbid mixed precision

wandb:
  project: sheaf-admm
  entity: null        # null -> your default wandb entity
  name: null          # null -> wandb auto-name
  group: null         # e.g. group seeds of one experiment together
  tags: []
  mode: disabled      # online | offline | disabled

training:
  seed: 42            # paper sweeps seeds {42, 123, 456}
  lr: 3.0e-4
  weight_decay: 1.0e-6
  epochs: 50
  batch_size: 128
  warmup_steps: 200   # linear warmup -> constant LR
  grad_clip: 1.0      # global-norm clip
  ema_decay: 0.999    # EMA of params, used at eval
  exit_on_nan: true
  val_interval: 5

  # ADMM horizon (Sheaf): K_train iterations at train, K_eval at eval.
  K_train: 40
  K_eval: 100
  loss_window: 4              # average CE over the final w iterates
  train_iters_dist: fixed     # fixed | uniform (Maze resamples K ~ U[train_iters_min, K_train])
  train_iters_min: 15

  # MPNN baseline message-passing rounds (the K analog for the recurrent MPNN).
  mpnn_train_rounds: 40
  mpnn_eval_rounds: 100

data:
  dir: ???            # set per experiment
  train_split: train
  val_splits: [test]
  loader: puzzle      # puzzle (maze/sudoku) | image (mnist)

# kwargs forwarded to the task hook (sheaf_admm.training.make_task)
task_cfg: {}

# sheaf_admm.models.ModelConfig fields (set per experiment)
model: {}