tune_keras_rnn_bayesoptim: Automatic cross-validated tuning of recurrent neural networks...

In thfuchs/fcf: Time Series Forecasting using Recurrent Neural Networks by Keras

Description

Tune recurrent neural network with Keras functional API and Bayes Optimization to select best performing model

Usage

tune_keras_rnn_bayesoptim(
  data,
  model_type,
  cv_setting,
  tuning_bounds = list(),
  col_id = NULL,
  col_date = "index",
  col_value = "value",
  save = NULL,
  save_id = NULL
)

Arguments

data	Univariate time series (data.frame) with date and value column, specified in col_date and col_value
model_type	One of "simple", "gru" or "lstm"
cv_setting	list of "periods_train", "periods_val", "periods_test" and "skip_span" for rolling_origin
tuning_bounds	list of tuning parameters - see section "Tuning Bounds"
col_id	Optional ID column in data, default to "ticker"
col_date	Date column in data, default to "index"
col_value	Value column in data, default to "value"
save	Automatically save tuning results? Specify NULL if not or character vector with path to directory for yes
save_id	optional character id for model filename

Value

list of Bayes Optimization results per split

Tuning Bounds

The following parameters are (currently) available for tuning.

• lag_1 (integer(2)) lower bounds - default c(1L, 2L)

• lag_2 (integer(2)) upper bounds - default c(1L, 2L)

• n_units (integer(2)) lower and upper bound for rnn units (cells) - default c(8L, 32L)

• n_epochs (integer(2)) lower and upper bound for epochs - default c(20L, 50L)

• optimizer_type (integer(2)) lower and upper bound for optimizer: 1 = "rmsprop", 2 = "adam", 3 = "adagrad" - default c(1L, 3L)

• dropout = (numeric(2)) lower and upper bound for dropout rate - default c(0, 0.5)

• recurrent_dropout = (numeric(2)) lower and upper bound for recurrent dropout rate - default c(0, 0.5)

• learning_rate = (numeric(2)) lower and upper bound for learning rate - default c(0.001, 0.01) Keep attention to the correct type (numeric length 2 / integer length 2). All bounds are to be set, otherwise default serves as fallback.

See Also

Other RNN tuning with Keras: tune_keras_rnn_eval(), tune_keras_rnn_predict()

Examples

## Not run: 
apple <- tsRNN::DT_apple

cv_setting <- list(
  periods_train = 90,
  periods_val = 10,
  periods_test = 10,
  skip_span = 5
)

bayes <- tune_keras_rnn_bayesoptim(apple, model_type = "simple", cv_setting)
bayes

## End(Not run)

thfuchs/fcf documentation built on April 18, 2021, 1:43 p.m.