diffusion/config.yaml

# Canonical ten-regime KH residual diffusion, t=[0,5].
# Train the single-Re KH residual recipe first; its model weights are the warm start.
config_type: diffusion

model_params:
  model_type: elucidated
  base_dim: 128
  dim_mults: [1, 2, 3, 5, 8, 12]
  channels: 4
  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
  field_representation: magnetic
  helmholtz_projection: true
  project_velocity: true
  project_B: true
  projection_mode: full_field_residual
  domain_size_x: 1.0
  domain_size_y: 1.0

normalization_params:
  type: per_re_paired_minmax
  channel_groups: [[0, 1], [2, 3]]
  feature_range: [-1, 1]
  re_values: [80, 200, 400, 650, 1000, 1500, 2050, 2750, 3600, 4500]
  default_re: 1000.0
  inputs_stats_template: ${MODEL_ROOT}/normalization/Re{re}_inputs_stats.npz
  targets_stats_template: ${MODEL_ROOT}/normalization/Re{re}_residual_targets_stats.npz

dataset_params:
  train_path: ${FEATURE_ROOT}/kh/multi_re/phase_t0_5/train
  val_path: ${FEATURE_ROOT}/kh/multi_re/phase_t0_5/val
  test_path: ${FEATURE_ROOT}/kh/multi_re/phase_t0_5/test
  seed: 42
  return_metadata: true
  balanced_re_batches: true
  res_per_batch: 10
  prediction_mode: residual
  residual_target: true
  source_time_range: [0.0, 5.0]
  source_output_dt: 0.02
  source_sub_t: 5
  frames_per_trajectory: 51

dataloader_params:
  train: {batch_size: 64, num_workers: 8, shuffle: true, pin_memory: true}
train_loader_params:
  batch_size: 64
  shuffle: true
  num_workers: 8
  pin_memory: true

optimizer_params:
  optimizer_type: adamw
  lr: 5.0e-5
  weight_decay: 1.0e-4
  betas: [0.9, 0.999]
  param_groups: {enabled: false}
  use_scheduler: true
  scheduler_type: cosine
  T_max: 1000
  eta_min: 2.5e-5

train_params:
  recipe: kh_phase_residual_multi_re
  epochs: 100
  checkpoint_path: ${OUTPUT_ROOT}/checkpoints/kh_multi_re_phase_t0_5.pt
  warm_start_checkpoint: ${OUTPUT_ROOT}/checkpoints/kh_single_re_phase_re1000.pt
  warm_start_mode: model_weights_only
  validation_interval: 5
  checkpoint_metric: denorm_rel_l2
  clip_grad: true
  clip_grad_max_norm: 1.0