model:
  num_classes: 80
  input_size: 640
  stem_channels: 64
  backbone_channels: [96, 192, 384, 512]
  backbone_depths: [2, 3, 6, 3]
  hidden_dim: 384
  fpn_depth: 2
  latent_count: 64
  latent_pool_sizes: [12, 6, 3]
  latent_layers: 2
  decoder_layers: 8
  num_queries: 300
  num_heads: 8
  local_points: 4
  dropout: 0.0
  dense_aux: true

loss:
  cost_class: 2.0
  cost_bbox: 5.0
  cost_giou: 2.0
  weight_class: 2.0
  weight_bbox: 5.0
  weight_giou: 2.0
  weight_dense: 1.0
  focal_alpha: 0.25
  focal_gamma: 2.0
  aux_weight: 1.0
  dense_topk: 5

train:
  epochs: 150
  batch_size: 16
  eval_batch_size: 8
  workers: 8
  prefetch_factor: 4
  lr: 0.0002
  backbone_lr: 0.0001
  min_lr_ratio: 0.05
  weight_decay: 0.05
  warmup_steps: 1500
  clip_grad_norm: 0.1
  amp: true
  amp_dtype: bfloat16
  channels_last: false
  compile: false
  ema_decay: 0.9998
  seed: 42
  eval_every: 1
  print_freq: 50

data:
  train_image_dir: train2017
  train_annotations: annotations/instances_train2017.json
  val_image_dir: val2017
  val_annotations: annotations/instances_val2017.json
  hflip_prob: 0.5
  scale_range: [0.65, 1.0]
  mean: [0.485, 0.456, 0.406]
  std: [0.229, 0.224, 0.225]