architecture: DETR
# pretrain_weights: # rewrite in FocalNet.pretrained in ppdet/modeling/backbones/focalnet.py
pretrain_weights: https://bj.bcebos.com/v1/paddledet/models/pretrained/focalnet_large_lrf_384_fl4_pretrained.pdparams
hidden_dim: 256
use_focal_loss: True

DETR:
  backbone: FocalNet
  transformer: DINOTransformer
  detr_head: DINOHead
  post_process: DETRPostProcess

FocalNet:
  arch: 'focalnet_L_384_22k_fl4'
  out_indices: [1, 2, 3]
  pretrained: https://bj.bcebos.com/v1/paddledet/models/pretrained/focalnet_large_lrf_384_fl4_pretrained.pdparams

DINOTransformer:
  num_queries: 900
  position_embed_type: sine
  num_levels: 4
  nhead: 8
  num_encoder_layers: 6
  num_decoder_layers: 6
  dim_feedforward: 2048
  dropout: 0.0
  activation: relu
  pe_temperature: 20
  pe_offset: 0.0
  num_denoising: 100
  label_noise_ratio: 0.5
  box_noise_scale: 1.0
  learnt_init_query: True

DINOHead:
  loss:
    name: DINOLoss
    loss_coeff: {class: 1, bbox: 5, giou: 2}
    aux_loss: True
    matcher:
      name: HungarianMatcher
      matcher_coeff: {class: 2, bbox: 5, giou: 2}

DETRPostProcess:
  num_top_queries: 300