lagged_mlp: General interface to MLP networks with lagged variables

In krzjoa/torchts: Time series Models with torch

General interface to MLP networks with lagged variables

Description

General interface to MLP networks with lagged variables

Usage

lagged_mlp(
  mode = "regression",
  timesteps = NULL,
  horizon = 1,
  learn_rate = 0.01,
  epochs = 50,
  hidden_units = NULL,
  dropout = NULL,
  batch_size = 32,
  scale = TRUE,
  shuffle = FALSE,
  jump = 1,
  sample_frac = 1
)

Arguments

mode	(character) Model mode, default: 'regression'.
timesteps	(integer) Number of timesteps to look back.
horizon	(integer) Forecast horizon.
learn_rate	(numeric or dials::learn_rate) Learning rate.
epochs	(integer or dials::epochs) Number of epochs.
hidden_units	(integer) Number of hidden units.
dropout	(logical or dials::dropout) Flag to use dropout.
batch_size	(integer) Batch size.
scale	(logical) Scale input features.
shuffle	(logical) Shuffle examples during the training (default: FALSE).
jump	(integer) Input window shift.
sample_frac	(numeric) A percent of subsamples used for training.

Details

This is a parsnip API to the lagged feed-forward networks (aka MLP). For now the only available engine is torchts_mlp.

Categorical features

Categorical features are detected automatically - if a column of your input data (defined in the formula) is logical, character, factor or integer.

Empty model

Neural networks, unlike many other models (e.g. linear models) can return values before any training epoch ended. It's because every neural networks model starts with "random" parameters, which are gradually tuned in the following iterations according to the Gradient Descent algorithm.

If you'd like to get a non-trained model, simply set epochs = 0. You still have to "fit" the model to stick the standard parsnip's API procedure.

Examples

library(torchts)
library(parsnip)
library(dplyr, warn.conflicts = FALSE)
library(rsample)

# Univariate time series
tarnow_temp <-
 weather_pl %>%
 filter(station == "TARNÓW") %>%
 select(date, temp = tmax_daily)

data_split <- initial_time_split(tarnow_temp)

mlp_model <-
   lagged_mlp(
     timesteps = 20,
     horizon = 1,
     epochs = 10,
     hidden_units = 32
   )

mlp_model <-
   mlp_model %>%
   fit(temp ~ date, data = training(data_split))

krzjoa/torchts documentation built on June 24, 2022, 5:30 a.m.