Configuration Recipes

Common edits to config.yml.

Disable the CNN (run LSEQM only)

Set blend_alpha to 1.0:

deep_learning:
  blend_alpha: 1.0

The CNN still trains but its output is multiplied by zero in the blend. To skip training entirely, comment out the CNN steps in notebooks/02_lseqmdl_bias_correction.ipynb.

Disable the confidence mask (uniform blend)

station_density:
  use_confidence_mask: false

The blend then uses a flat blend_alpha everywhere. Useful for an ablation: compare LSEQM+DL with the mask off and on to quantify what the mask is doing.

Change the GPD threshold

statistical_params:
  gpd_threshold_percentile: 90   # was 80

Higher thresholds give a cleaner separation between the gamma body and the GPD tail but reduce the number of exceedances per dekad. There is no “too few exceedances” error - below n_splits_gpd_crossvalidate exceedances the fit silently drops to a single non-cross-validated GPD fit. To keep the K-Fold average, lower the threshold or raise wet_day_threshold so that dry days are excluded.

Switch to non-interactive (batch) mode

runtime:
  interactive: false
  existing_file_action: "skip"      # or "overwrite", "abort"
  existing_model_action: "use_existing"

In skip mode, a re-run will not regenerate files that already exist. Use this when iterating on a single dekad - flip to overwrite for the cell you are changing.

Tighten the upper cap

statistical_params:
  upper_cap_threshold_percentile: 99.0   # was 99.9

Clips ~1% of the tail. Use if you see unrealistic spikes after the GPD step (rare but possible when exceedances are very heavy-tailed).

Run on Colab without a local checkout

directories:
  main_dir: "/content/hybrid-bias-correction"
  input_dir: "{main_dir}/data/example_bali"
  output_dir: "{main_dir}/data/example_bali/output"

The Bali config already has these values commented in; uncomment the Colab line and comment the WSL line.

Use a denser CNN

deep_learning:
  num_filters_1: 64        # was 32
  num_filters_2: 128       # was 64
  dense_layer_size: 256    # was 128

This nearly quadruples the parameter count of the Flatten -> Dense layer. Only worth doing if the val_loss plateau on the default network is well above your noise floor.