__include__: [
  '../dataset/coco.yml',  # Base dataset configuration. Replace with custom.yml when using a custom dataset
  'ecdet.yml',                      
]

ViTAdapter:
  name: ecvitt
  embed_dim: 192
  num_heads: 3
  interaction_indexes: [10, 11]     # Indices of transformer blocks used for feature interaction/fusion
  weights_path: ecvits/ecvitt.pth   # Pretrained backbone. Automatically downloaded on first use.
  skip_load_backbone: False         # If True, the backbone will be initialized from scratch (no pretrained weights)

optimizer:
  type: AdamW
  params: 
    - # Backbone parameters excluding normalization layers and bias
      params: '^(?=.*backbone)(?!.*(?:norm|bn|bias)).*$'
      lr: 0.000025

    - # Backbone normalization layers (norm/bn) and bias parameters
      params: '^(?=.*backbone)(?=.*(?:norm|bn|bias)).*$'
      lr: 0.000025
      weight_decay: 0.0

    - # Non-backbone normalization layers and bias parameters
      params: '^(?!.*\.backbone)(?=.*(?:norm|bn|bias)).*$'
      weight_decay: 0.0

  lr: 0.0005                # Base learning rate for non-backbone parameters
  betas: [0.9, 0.999]      # AdamW beta coefficients
  weight_decay: 0.0001     # Weight decay applied to parameters except norm/bias

epochs: 74                 # Total training epochs. COCO 6× schedule (12 epochs = 1×) plus 2 extra epochs without augmentation
warmup_iter: 2000         # Number of iterations used for learning rate warmup
lr_gamma: 0.5             # Learning rate decay factor during LR scheduling

eval_spatial_size: [640, 640]  # Input resolution for training/evaluation (height, width). Use [1280,1280] for high-resolution training

train_dataloader:
  total_batch_size: 32      # Global batch size across all GPUs (adjust depending on GPU memory)
  dataset:
    transforms:
      mosaic_epoch: 36        # Apply Mosaic augmentation until this epoch. Recommended to set this to half of stop_epoch
      mosaic_prob: 0.75       # Probability of applying Mosaic augmentation
      stop_epoch: 72          # Disable all augmentations after this epoch (last 2 epochs without augmentation)

  collate_fn:
    mixup_prob: 0.75          # Probability of applying MixUp augmentation
    mixup_epoch: 36           # Apply MixUp augmentation until this epoch. Recommended to set this to half of stop_epoch

__include__: [
  '../dataset/coco.yml',  # Base dataset configuration. Replace with custom.yml when using a custom dataset
  '../ecdet/ecdet_s.yml',           # Inherit detection model configuration
  'ecseg.yml',                      # Segmentation-specific configuration
]

train_dataloader:  # Add only the parameters you wish to override
...