torchts_mlp: MLP model for time series forecasting

In krzjoa/torchts: Time series Models with torch

MLP model for time series forecasting

Description

MLP model for time series forecasting

Usage

torchts_mlp(
  formula,
  data,
  learn_rate = 0.001,
  hidden_units,
  dropout = FALSE,
  timesteps = 20,
  horizon = 1,
  jump = horizon,
  optim = optim_adam(),
  validation = NULL,
  stateful = FALSE,
  batch_size = 1,
  epochs = 10,
  shuffle = TRUE,
  scale = TRUE,
  sample_frac = 1,
  loss_fn = nn_mse_loss(),
  device = NULL
)

Arguments

formula	(formula) A formula describing, how to use the data
data	(data.frame) A input data.frame.
learn_rate	(numeric) Learning rate.
hidden_units	(integer) Number of hidden units.
timesteps	(integer) Number of timesteps used to produce a forecast.
horizon	(integer) Forecast horizon.
jump	(integer) Input window shift.
optim	(function) A function returning a torch optimizer (like optim_adam) or R expression like optim_adam(amsgrad = TRUE). Such expression will be handled and feed with params and lr arguments.
validation	(data.frame or numeric) Validation dataset or percent of TODO.
batch_size	(integer) Batch size.
epochs	(integer) Number of epochs to train the network.
shuffle	(logical) A dataloader argument - shuffle rows or not?
scale	(logical or list)
sample_frac	(numeric) A fraction of time series to be sampled.
loss_fn	(function) A torch loss function.
device	(character) A torch device.

Examples

library(dplyr, warn.conflicts = FALSE)
library(torch)
library(torchts)
library(timetk)

# Preparing a dataset
tiny_m5_sample <-
  tiny_m5 %>%
  filter(item_id == "FOODS_3_586", store_id == "CA_1") %>%
  mutate(value = as.numeric(value))

tk_summary_diagnostics(tiny_m5_sample)
glimpse(tiny_m5_sample)

TIMESTEPS <- 20

data_split <-
  time_series_split(
    tiny_m5_sample, date,
    initial = "4 years",
    assess  = "1 year",
    lag     = TIMESTEPS
  )

# Training
mlp_model <-
  torchts_mlp(
    value ~ date + value + sell_price + wday,
    data = training(data_split),
    hidden_units = 10,
    timesteps = TIMESTEPS,
    horizon   = 1,
    epochs = 10,
    batch_size = 32
  )

# Prediction
cleared_new_data <-
  testing(data_split) %>%
  clear_outcome(date, value, TIMESTEPS)

forecast <-
  mlp_model %>%
  predict(cleared_new_data)

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