yolov7_tiny_qat_dis.yaml

Global:
  model_dir: ./yolov7-tiny.onnx
  image_path: None      # If image_path is set, it will be trained directly based on unlabeled images, no need to set the COCO dataset path.
  coco_dataset_dir: dataset/coco/
  coco_train_image_dir: train2017
  coco_train_anno_path: annotations/instances_train2017.json
  coco_val_image_dir: val2017
  coco_val_anno_path: annotations/instances_val2017.json
  arch: YOLOv7

Distillation:
  alpha: 1.0
  loss: soft_label

QuantAware:
  onnx_format: true
  activation_quantize_type: 'moving_average_abs_max'
  quantize_op_types:
  - conv2d
  - depthwise_conv2d

TrainConfig:
  train_iter: 5000
  eval_iter: 1000
  learning_rate:
    type: CosineAnnealingDecay 
    learning_rate: 0.00003
    T_max: 8000
  optimizer_builder:
    optimizer:
      type: SGD
    weight_decay: 0.00004