# Venice-H1 default configuration
# Matches paper specification exactly (Sections 3.4, 3.5, 4.2)
# OdaxAI Research — nicolo.savioli@odaxai.com

model:
  query_feat_dim: 256       # D: DeRIS query embedding dimension
  hidden_dim: 512           # Hd: Transformer hidden size
  n_layers: 3               # L: Transformer encoder layers
  n_heads: 8                # A: attention heads
  n_queries: 10             # N: candidate queries per sample
  dropout: 0.1              # 10% dropout
  use_grid: true            # multi-scale grid signatures (675-dim)
  tau: 0.05                 # Failure Gate threshold (canonical operating point)

training:
  batch_size: 512
  epochs: 20                # 20 epochs on cached features
  lr: 5.0e-4               # AdamW learning rate
  weight_decay: 1.0e-4
  scheduler: cosine
  warmup_epochs: 3          # 3-epoch warmup
  seed: 42
  fp16: true                # mixed precision

  # Loss: L = L_gate + lambda * L_gain
  lambda_gain: 5.0          # λ=5 (paper Section 3.5)
  focal_gamma: 2.0          # focal BCE γ=2
  auto_wpos: true           # automatic positive weight for class imbalance

data:
  # Pre-extracted feature caches (.pt files)
  # Produced by scripts/extract_features.py
  # Total: ~3.2 GB for 126k training samples
  train_splits:
    - data/cached_train_refcoco_unc_feats.pt
    - data/cached_train_refcoco+_unc_feats.pt
    - data/cached_train_refcocog_umd_feats.pt

  val_splits:
    - data/cached_val_refcoco_unc_feats.pt
    - data/cached_val_refcoco+_unc_feats.pt
    - data/cached_val_refcocog_umd_feats.pt

  test_splits:
    - data/cached_testA_refcoco_unc_feats.pt
    - data/cached_testB_refcoco_unc_feats.pt
    - data/cached_testA_refcoco+_unc_feats.pt
    - data/cached_testB_refcoco+_unc_feats.pt
    - data/cached_test_refcocog_umd_feats.pt

output:
  checkpoint_dir: checkpoints/
  log_dir: runs/venice_h1/