LSTM: Instantiate and train an LSTM model

In dhopp1/nowcastLSTM: LSTMs for Nowcasting

Instantiate and train an LSTM model

Description

Instantiate and train an LSTM model

Usage

LSTM(
  data,
  target_variable,
  n_timesteps,
  fill_na_func = "mean",
  fill_ragged_edges_func = "mean",
  n_models = 1,
  train_episodes = 200,
  batch_size = 30,
  decay = 0.98,
  n_hidden = 20,
  n_layers = 2,
  dropout = 0,
  criterion = "''",
  optimizer = "''",
  optimizer_parameters = list(lr = 0.01),
  seeds = c(),
  python_model_name = "model"
)

Arguments

data	n x m+1 dataframe with a column of type Date
target_variable	string with the column name of the target variable
n_timesteps	how many historical periods to consider when training the model. For example if the original data is monthly, n_steps=12 would consider data for the last year.
fill_na_func	function to replace within-series NAs. Options are c("mean", "median", "ARMA").
fill_ragged_edges_func	function to replace NAs in ragged edges (data missing at end of series). Options are c("mean", "median", "ARMA"). Note ARMA filling will be significantly slower as models have to be estimated for each variable to fill ragged edges.
n_models	integer, number of models to train and take the average of for more robust estimates
train_episodes	integer, number of epochs/episodes to train the model
batch_size	integer, number of observations per training batch
decay	double, learning rate decay
n_hidden	integer, number of hidden states in the network
n_layers	integer, number of LSTM layers in the network
dropout	double, dropout rate between the LSTM layers
criterion	torch loss criterion, defaults to MAE. For E.g. MSE, pass "torch.nn.MSELoss()"
optimizer	torch optimizer, defaults to Adam. For a different one, pass e.g. "torch.optim.SGD"
optimizer_parameters	parameters for optimizer, including learning rate. Pass as a named list, e.g. list("lr"=0.01, "weight_decay"=0.001)
seeds	c(int), list of integers, what to seed the initial weights for reproducibility. Must be list of same length as n_models parameter
python_model_name	what the model will be called in the python session. Relevant if more than one model is being trained for simultaneous use. For clarity, should be the same as the name of the R object the model is being saved to.

Value

trained LSTM model

dhopp1/nowcastLSTM documentation built on May 7, 2024, 9:30 a.m.