# GridBased training -- polychromatic 5-30 keV, MOENCH014 @150 V
#
# Same architecture and the same losses as the 12 keV configuration; what
# changes is that the energy head now carries information, so it gets a real
# weight and a scale that is not tied to a beam energy.

experiment:
  name: "poly5to30keV_Moench014"
  results_dir: "Results"

data:
  folder: /mnt/sls_det_storage/moench_data/MLXID/Samples/GridBased/poly5to30keV_Moench014_150V
  train_shards: [0, 11]
  val_shards: [0, 1]
  max_train_frames: null
  max_val_frames: 8000
  input_scale: 15.0            # keV, mid-range; frames are divided by this
  energy_scale: 15.0           # keV, the energy target is E / energy_scale
  augment: true
  batch_size_train: 256
  batch_size_val: 512
  num_workers: 8

model:
  channels: [32, 64, 128, 128] # still 4 x conv3x3, so R = 4 px; wider, because
                               # the cloud size now has to be disentangled from
                               # the energy rather than mapping straight to depth
  predict_energy: true
  norm: "none"
  prior_prob: 0.01

loss:
  w_obj: 1.0
  w_off: 1.0
  w_ene: 1.0                   # 0.1 at a single energy, where the head learned a
                               # constant; here it carries the spectrum
  focal_alpha: 0.25
  focal_gamma: 2.0
  smooth_l1_beta_offset: 0.01
  smooth_l1_beta_energy: 0.03  # ~ the 0.39 keV noise floor of a 3x3 sum, / 15 keV

training:
  epochs: 300
  lr: 1.0e-3
  weight_decay: 0.0
  scheduler_factor: 0.5
  scheduler_patience: 6
  amp: true
  seed: 0
  checkpoint_epochs: [50, 100, 150, 200, 250, 300]

eval:
  threshold: 0.3
  merge_distance: 0.1
  match_radius: 1.0
  pixel_size_um: 25.0