# Previous-study DINO baseline: full-field diffusion, not residual diffusion.
config_type: diffusion

model_params:
  model_type: elucidated
  base_dim: 128
  dim_mults: [1, 2, 3, 5, 8, 12]
  channels: 3
  self_condition: true
  learned_variance: false
  learned_sinusoidal_cond: false
  random_fourier_features: false
  sinusoidal_pos_emb_theta: 10000
  learned_sinusoidal_dim: 16
  dropout: 0.0
  attn_heads: 8
  attn_dim_head: 64
  flash_attn: true
  image_size: 128
  num_sample_steps: 32
  sigma_min: 0.002
  sigma_max: 80
  sigma_data: 0.5
  rho: 7
  P_mean: -1.2
  P_std: 1.2
  S_churn: 80
  S_tmin: 0.05
  S_tmax: 50
  S_noise: 1.003
  helmholtz_projection: false
  project_velocity: false
  project_B: false

normalization_params:
  type: minmax
  feature_range: [-1, 1]
  inputs_stats_file: ${MODEL_ROOT}/normalization/diffusion_inputs_stats.npz
  targets_stats_file: ${MODEL_ROOT}/normalization/diffusion_targets_stats.npz

dataset_params:
  train_path: ${FEATURE_ROOT}/train
  val_path: ${FEATURE_ROOT}/val
  test_path: ${FEATURE_ROOT}/test
  prediction_mode: direct
  residual_target: false

dataloader_params:
  train: {batch_size: 64, num_workers: 1, shuffle: true, pin_memory: true}

# Backward-compatible key consumed by the shared diffusion dataloader.
train_loader_params:
  batch_size: 64
  shuffle: true
  num_workers: 1
  pin_memory: true

optimizer_params:
  optimizer_type: adamw
  lr: 1.0e-4
  weight_decay: 1.0e-4
  betas: [0.9, 0.999]
  use_scheduler: true
  scheduler_type: cosine
  T_max: 1000
  eta_min: 5.0e-5

train_params:
  recipe: previous_dino
  epochs: 101
  checkpoint_path: ${OUTPUT_ROOT}/checkpoints/dino_re1000.pt
  load_checkpoint: ""
  validation_interval: 10
  checkpoint_metric: denorm_rel_l2
  clip_grad: true
  clip_grad_max_norm: 1.0