hyperparameters.yml

# Hyperparameters shared by all model frameworks

base:
  base_lr:   # base learning rate for the model optimizer
    type: base
    hyperparams:
      - !!float 1e-1
      - !!float 1e-2
      - !!float 1e-3
      - !!float 1e-4
  numeric_strat:   # strategy for handling numeric inputs
    type: numeric
    hyperparams:
      - minmax   # scales the numeric value to [0,1] w/ MinMaxScaler
      - standard   # scales using StandardScaler
      - quantiles   # splits the field into 4 quantile buckets for binary-encoding
      - percentiles   # splits the field into 10 percentile buckets for binary-encoding
  datetime_month:   # include categorical month when encoding datetimes
    type: datetime
    hyperparams:
      - True
      - False
  datetime_year:   # include categorical year when encoding datetimes
    type: datetime
    hyperparams:
      - True
      - False
  # datetime_holiday:   # include categorical holiday when encoding datetimes
  #   type: datetime
  #   hyperparams:
  #     - True
  #     - False
  categorical_strat:   # strategy for handling categorical inputs
    type: categorical
    hyperparams:
      - all_binary   # encodes all levels with binary-encoding
      - top10_perc   # encodes top 10% of levels, encodes rest as "other"
      - top50_perc   # encodes top 50% of levels, encodes rest as "other"

tensorflow:
  weight_decay:   # AdamW Weight Decay
    type: base
    hyperparams:
      - !!float 5e-2
      - !!float 25e-3
      - !!float 1e-2
  batch_size:   # training batch size
    type: base
    hyperparams:
      - 64
      - 128
      - 256
  text_max_length:   # maximum number of input words to use for text
    type: text
    hyperparams:
      - 10
      - 20
      - 40
  text_dropout:   # input text dropout for text-encoding
    type: text
    hyperparams:
      - 0.0
      - 0.2
      - 0.5
  text_rnn_type:   # type of RNN for text-encoding
    type: text
    hyperparams:
      - LSTM
      - GRU
  text_rnn_size:   # size of RNN for text-encoding
    type: text
    hyperparams:
      - 32
      - 64
      - 128
  text_embed_size:   # size of Embeddings for text-encoding
    type: text
    hyperparams:
      - 20
      - 50
      - 100
  mlp_blocks:   # number of MLP blocks after feature concatenation
    type: base
    hyperparams:
      - 2
      - 4
      - 6
  mlp_first_size:   # size of the first Dense layer in the MLP blocks
    type: base
    hyperparams:
      - 64
      - 128
      - 256
  mlp_size:   # size of the Dense layers in the MLP blocks
    type: base
    hyperparams:
      - 64
      - 128
      - 256
  mlp_dropout:   # dropout of the Dense layers in the MLP blocks
    type: base
    hyperparams:
      - 0.0
      - 0.2
      - 0.5
  mlp_activation:   # activation function for Dense layers in the MLP blocks
    type: base
    hyperparams:
      - relu
      - selu
  mlp_regularizer:   # l2 regularization value for Dense layers in the MLP blocks
    type: base
    hyperparams:
      - !!null None
      - l2(1e-2)
      - l2(1e-3)
  output_regularizer:   # l2 regularization value for the final output layer
    type: base
    hyperparams:
      - !!null None
      - l2(1e-2)
      - l2(1e-3)
  reg_objective:   # regression objective for the model (can differ from experiment objective)
    type: regression
    hyperparams:
      - mse
      - msle
      - mape
      - poisson

xgboost:
  max_depth:   # maximum depth of tree
    type: base
    hyperparams:
      - 3
      - 5
      - 7
      - 9
  gamma:   # gamma regularization
    type: base
    hyperparams:
      - 1
      - 5
      - 10
  min_child_weight:   # minimum child weight for split
    type: base
    hyperparams:
      - 1
      - 3
      - 5
  subsample:   # proportion of rows to subsample for trees
    type: base
    hyperparams:
      - 0.5
      - 0.8
      - 1.0
  colsample_bytree: # proportion of columns to subsample for trees
    type: base
    hyperparams:
      - 0.3
      - 0.8
      - 1.0
  max_bin: # max bins for hist binning
    type: base
    hyperparams:
      - 128
      - 256
      - 512
  reg_objective:   # regression objective for the model (can differ from experiment objective)
    type: regression
    hyperparams:
      - reg:linear
      - count:poisson