lcnn.yaml

# configurations for LCNN model

# Parameters for training model
train_params:
  device: cuda
  train_batch_size: 1
  valid_batch_size: 1
  test_batch_size: 1
  predict_batch_size: 1
  max_epochs: 100
  # Max time used for training (days:hours:mins:secs)
  max_time: "00:71:45:00"
  # Validate after every n batches
  val_check_interval: 1.0
  num_workers: 10
  gpus: 1
  savefile: "lcnn-diff-rmse-30lt-20062022"
  # many leadtimes
  verif_leadtimes: 6
  # number of batches to validate on
  val_batches: 5000000
  # number of baches to train on (per epoch)
  train_batches: 5000000
  early_stopping:
    monitor: "val_loss"
    patience: 5
  lr_scheduler:
    name: "reduce_lr_on_plateau"
    kwargs:
      mode: "min"
      factor: 0.1
      patience: 3

model:
  lr: 1e-04
  rainnet:
    input_shape: [4, 512, 512]
    kernel_size: 3
    mode: "regression"
    conv_shape:
      [
        ["1", [4, 64]],
        ["2", [64, 128]],
        ["3", [128, 256]],
        ["4", [256, 512]],
        ["5", [512, 1024]],
        ["6", [1536, 512]],
        ["7", [768, 256]],
        ["8", [384, 128]],
        ["9", [192, 64]],
      ]

  loss:
    name: "rmse"
    kwargs:
      alpha: 0.5
    discount_rate: 0.0
  train_leadtimes: 6
  apply_differencing: true
  display: 250
prediction:
  predict_leadtimes: 36
  euler_transform_nworkers: 6
  # Extrapolation keyword arguments
  extrap_kwargs:
    # Interpolation order, options
    # 1 (linear).
    # 0 (nearest neighbor)
    # 3 (cubic)
    interp_order: 3
    allow_nonfinite_values: true
prediction_output:
  # Output directory
  output_dir: /fmi/scratch/project_2005001/nowcasts/lcnn/lcnn_diff_rmse_30lt_20062022
  # Output filename format (can contain {common_time} to change time)
  filename: lcnn_diff_rmse_30lt_20062022_36.h5
  # where to save predictions in the HDF5 file
  group_format: "{common_time:%Y-%m-%d %H:%M:%S}/lcnn-diff-rmse-30lt-20062022"
  # Attributes of the dataset in the HDF5 file
  what_attrs:
    quantity: DBZH
    gain: 0.5
    offset: -32
    nodata: 255
    undetect: 0