# GridBased training -- 12 keV, MOENCH040 @150 V  (WP3, monochromatic gate G2)

experiment:
  name: "12keV_Moench040"
  results_dir: "Results"

data:
  folder: /mnt/sls_det_storage/moench_data/MLXID/Samples/GridBased/12keV_Moench040_150V
  train_shards: [0, 11]        # inclusive
  val_shards: [0, 1]
  max_train_frames: null       # null = all
  max_val_frames: 8000
  input_scale: 12.0            # keV; frames are divided by this
  energy_scale: 12.0           # keV; energy target is E / energy_scale
  augment: true                # 8-fold D4, a true symmetry of the problem
  batch_size_train: 256
  batch_size_val: 512
  num_workers: 8

model:
  channels: [32, 64, 64, 64]   # 4 x conv3x3 -> receptive-field radius R = 4 px
  predict_energy: true
  norm: "none"                 # see models.py: 'group' breaks strict locality
  prior_prob: 0.01             # focal-loss bias init

loss:
  w_obj: 1.0
  w_off: 1.0
  w_ene: 0.1
  focal_alpha: 0.25
  focal_gamma: 2.0
  smooth_l1_beta_offset: 0.01  # px = 0.25 um, as for the cluster-based models
  smooth_l1_beta_energy: 0.02

training:
  epochs: 150
  lr: 1.0e-3
  weight_decay: 0.0
  scheduler_factor: 0.5
  scheduler_patience: 4
  amp: true
  seed: 0
  checkpoint_epochs: [25, 50, 75, 100, 125, 150]

eval:
  threshold: 0.3           # from the ROC of run 260830_00: eff plateaus below 0.3
  merge_distance: 0.1        # px; position-space duplicate suppression
  match_radius: 1.0            # px; 25 um
  pixel_size_um: 25.0