model_type: ecd

input_features:
  - name: fixed acidity
    type: number
    preprocessing:
      normalization: zscore
  - name: volatile acidity
    type: number
    preprocessing:
      normalization: zscore
  - name: citric acid
    type: number
    preprocessing:
      normalization: zscore
  - name: residual sugar
    type: number
    preprocessing:
      normalization: zscore
  - name: chlorides
    type: number
    preprocessing:
      normalization: zscore
  - name: free sulfur dioxide
    type: number
    preprocessing:
      normalization: zscore
  - name: total sulfur dioxide
    type: number
    preprocessing:
      normalization: zscore
  - name: density
    type: number
    preprocessing:
      normalization: zscore
  - name: pH
    type: number
    preprocessing:
      normalization: zscore
  - name: sulphates
    type: number
    preprocessing:
      normalization: zscore
  - name: alcohol
    type: number
    preprocessing:
      normalization: zscore

output_features:
  - name: quality
    type: binary

trainer:
  epochs: 20

# NOTE: The Optuna executor requires PR #4090 to be merged, or Ludwig >= 0.14.
# Install: pip install ludwig optuna
hyperopt:
  executor:
    type: optuna
    num_samples: 20
    sampler: auto        # auto selects the best sampler; also: tpe, gp, cmaes, random
    pruner: hyperband    # stop unpromising trials early (MedianPruner also supported)
    storage: sqlite:///optuna_results.db  # enables resumability across runs
  parameters:
    trainer.learning_rate:
      space: loguniform
      lower: 1.0e-5
      upper: 1.0e-2
    trainer.batch_size:
      space: int
      lower: 16
      upper: 256
    trainer.optimizer.type:
      space: choice
      categories:
        - adam
        - adamw
        - radam
        - schedule_free_adamw
  goal: minimize
  metric: validation.combined.loss
  split: validation