# logdir = os.path.join(output, task_name)
output: ./output_dir/
task_name: base_exp

dataset:
  train:
    rootA: 'flow_model/data/train/weighted_real'
    rootB: 'flow_model/data/train/maps_sample/subject04_crisp_v_180.npy'
    width: 512
    height: 256
    scale_l: 0.8
    scale_h: 1.0
    transform: [] #['h_flip', 'v_flip', 'crop', 'normalize', 'random_resized_crop']
    random_pair: True
    return_name: False
    batch_size: 1

  test:
    rootA: 'flow_model/data/test/weighted_real'
    rootB: 'flow_model/data/test/maps_sample/subject18_crisp_v_180.npy'
    width: 512
    height: 256
    scale_l: 0.8
    scale_h: 1.0
    transform: [] #['h_flip', 'v_flip', 'crop', 'normalize']
    random_pair: False
    return_name: True
    batch_size: 16

lr: 0.0001
epochs: 120
max_iter: 300000
print_freq: 450
save_freq: 450
resume: True
load_path: 'flow_model/checkpoint_for_resume/0.ckpt.pth.tar'

network:
  configurable: False #[True, False]
  pad_size: 10
  in_channel: 3
  out_channels: [30, 120] #[30, 120], [12, 60, 120], [30, 120, 480], [30, 120, 480, 1920]
  weight_type: 'learned' #['fixed', 'sigmoid', 'softmax', 'attention', 'learned']

loss:
  vgg_encoder: 'flow_model/model/losses/vgg_model/vgg_normalised.pth'
  k: 0.7
  weight: 0.7

lr_scheduler:
  type: cosine
  eta_min: 0.0000000