# Multimodal theme classification (classification/classify.py)
# 5-fold CV; for every modality combination every grid configuration is trained in every fold,
# the one with the best inner-validation macro-F1 is chosen, and its test macro-F1 is reported.
# Override any value from the command line, e.g.
#   python classification/classify.py --set data.lyrics=lyrics_phasea --set output_dir=results/classification/phasea

data:
  metadata: data/multimodal/metadata.csv      # songs, labels, 'fold', 'lyrics_group'
  embeddings: data/multimodal/embeddings.npz  # vocal / instrument / mix + lyrics variants
  lyrics: lyrics_phaseb_cv     # lyrics_phaseb_cv: Phase B model trained without the test fold
                               # lyrics_phasea:    Phase A (TSDAE) model, never saw a label

output_dir: results/classification/phaseb

modalities: [lyrics, vocal, instrument, mix]
combos: all          # all = every non-empty subset (15); or a list, e.g. [lyrics, lyrics+vocal]
folds: all           # all, or a list of fold ids, e.g. [0]
seed: 42
inner_val_splits: 6  # 1/6 of the non-test songs form the validation split of each fold
late_fusion: true    # also score the average of the best single-modality models

training:
  max_epochs: 60
  min_epochs: 10               # never stop before this
  patience: 10                 # epochs without a validation gain > min_delta
  min_delta: 0.001
  lr_scheduler_patience: 5     # ReduceLROnPlateau on validation macro-F1, factor 0.5
  batch_size: 64
  weight_decay: 0.0001

model:
  d_model: 256                 # width of every modality projection
  dropout: 0.3
  attention_heads: 4           # xattn
  aux_loss_weight: 0.5         # weight of the per-modality auxiliary losses (aux: true)

grid:
  methods: [mlp, gmu, xattn]             # for combinations of 2+ modalities
  single_modality_methods: [mlp, xattn]  # xattn only for single audio modalities
  lr: [0.0001, 0.0003]
  soft_labels: [false, true]             # train on final_* soft scores instead of the label
  modality_dropout: [0.0, 0.3]           # 2+ modalities only
  aux_loss: [false, true]                # mlp / gmu with 2+ modalities only