_BASE_: [
  '../datasets/coco_detection.yml',
  '../runtime.yml',
  '_base_/optimizer_300e.yml',
  '_base_/yolov6_cspbep.yml',
  '_base_/yolov6_reader_high_aug.yml',
]
depth_mult: 1.0
width_mult: 1.0
log_iter: 20
snapshot_epoch: 10
weights: output/yolov6_l_300e_coco/model_final


### reader config
TrainReader:
  batch_size: 32 # default 8 gpus, total bs = 256

EvalReader:
  batch_size: 1


### model config
act: 'silu'
training_mode: "conv_silu" # Note: L use silu

YOLOv6:
  backbone: CSPBepBackbone
  neck: CSPRepBiFPAN
  yolo_head: EffiDeHead_fuseab
  post_process: ~

EffiDeHead_fuseab:
  reg_max: 16
  use_dfl: True
  iou_type: 'giou'
  loss_weight: {cls: 1.0, iou: 2.5, dfl: 0.5, cwd: 10.0}
  distill_weight: {cls: 2.0, dfl: 1.0} # 2:1 in L-relu version, will not work default (self_distill=False)

CSPBepBackbone:
  csp_e: 0.50

CSPRepBiFPAN:
  csp_e: 0.50


### distill config
## Step 1: Training the base model, get about 51.8 mAP

## Step 2: Self-distillation training, get about 52.8 mAP

YOLOv6:
  backbone: CSPBepBackbone
  neck: CSPRepBiFPAN
  yolo_head: EffiDeHead
  post_process: ~

EffiDeHead:
  reg_max: 16
  use_dfl: True

## Please cancel the following comment and train again:
# self_distill: True
# pretrain_weights: output/yolov6_l_300e_coco/model_final.pdparams
# save_dir: output_distill
# weights: output_distill/yolov6_l_300e_coco/model_final