train.yaml

MODEL:
  nms_size: 10
  nms_thresh: 0.1

  batch_size: 4
  num_epoch: 150

  device: cuda:0
  load_pre_trained_model: false
  pretrained_path: ./save/SuperRetina.pth

  model_save_path: ./save/new_model.pth
  model_save_epoch: 1

DATASET:
  model_image_width: 768
  model_image_height: 768
  dataset_path: ./data/Lab
  train_split_file: eccv22_train.txt
  val_split_file: eccv22_val.txt
  auxiliary: ./data/Auxiliary

PKE:
  pke_start_epoch: 0  # the epoch to start PKE learn
  geometric_thresh: 0.5
  content_thresh: 0.7
  gaussian_kernel_size: 13  # used to generate heatmap
  gaussian_sigma: 2

  pke_show: true
  pke_show_epoch: 5
  pke_show_list: []  # if pke_show_list == [], then randomly select one sample to show


# value_map is used to record history learned points. The value of the position of the newly added keypoint will increase.
VALUE_MAP:
  area: 8  # if the neighbor of the keypoint have non-zero values, the values will increase by `value_increase_area`
  value_increase_point: 5  # if the neighbor of the keypoint is all zero, the keypoint position's value will increase by `value_increase_point`
  value_increase_area: 1
  value_decay: 1  # if the history keypoints don't appear in this epoch, the correspoding value will decay
  is_value_map_save: false  # if false, running in RAM, otherwise, the value map will store as a temp file in `value_map_save_dir`
  value_map_save_dir: ./data/lab_values