# FLAME on OxHyperSyntheticCH4 (EMIT) — reference training configuration.
#
# Protocol (aligned to the HyperspectralViTs training regime, arXiv:2410.17248):
# deterministic 64x64 / stride-32 grid windows from
# train_filtered_v2_tiled_64_32.csv, plume windows balanced to ~50% of every
# epoch by a with-replacement weighted sampler, BCE with pos_weight capped at
# pw_max, 50 epochs, bf16 autocast. Validation reproduces the test protocol
# (window pooling, logits >= 0, no morphological filtering).
#
# This configuration trains the self-contained physics variant
# (use_mag_in_seg: false): the segmentation head consumes the model's own
# parameter-free physics score, and mag1c (B_magic30_tile.tif / score_divisor,
# clipped to [0, 2]) acts only as a training-time auxiliary target for the
# score head, co-trained at a weak weight that fades linearly over the whole
# run (aux_weight 0.15, decay_frac 1.0). aux_bg_weight adds mild background
# suppression (score -> 0 where mag1c is silent) and pw_max: 3 provides a mild
# recall bootstrap — both guard against the all-negative collapse.
#
# Alternative (mag-in-seg) variant: set use_mag_in_seg: true so the seg head
# consumes mag1c directly at train and inference time; then set aux_weight: 0,
# aux_bg_weight: 0, pw_max: 1.

dataset: emit

model:
  in_channels: 64                # SWIR 2004-2478 nm (EMIT)
  width: 32
  modes1: 16
  modes2: 16
  n_fno_layers: 3
  n_ufno_layers: 3
  dropout_rate: 0.0
  seg_channels: [32, 16]
  seg_kernel_size: 3
  score_divisor: 1750.0          # shared by score normalisation and the mag1c input
  rgb_divisor: 20.0
  norm_type: group
  score_clamp: [-10.0, 10.0]
  use_mag_in_seg: false
  spectrum_path: resources/emit/ch4_spectrum.npy
  centers_path: resources/emit/band_centers.npy
  baseline_stats_path: resources/emit/baseline_stats.pt

train:
  uid: flame_emit
  lr: 5.0e-4
  epochs: 50
  batch_size: 18                 # must be divisible by the number of GPUs
  n_workers: 2
  log_dir: logs
  seed: 42
  patience: 25
  ckpt_interval: 10
  prefetch_factor: 2
  vis_interval: 10
  pretrain_epochs: 0
  aux_weight: 0.15
  decay_frac: 1.0
  aux_bg_weight: 0.1
  grad_clip_norm: 5.0
  seg_loss: bce
  pw_max: 3
  focal_gamma: 0.0
  val_threshold: 0.5
  amp: bf16
  val_window_batch: 256
  find_unused_parameters: true

data:
  root_dir: datasets/oxhyper_synthetic_ch4
  store_dir: datasets/emit_ram64   # flat fp16 memmap store, built automatically
  npy_cache_dir: ~                 # optional (scripts/build_emit_npy_cache.py)
  windows_csv: train_filtered_v2_tiled_64_32.csv
  wv_window: [2004.0, 2478.0]
  patch_size: 64
  stride: 32
  augment: true
  max_train_windows: ~
  max_val_tiles: ~