timeseries_dataset_from_array: Creates a dataset of sliding windows over a timeseries provided as array

Description

Creates a dataset of sliding windows over a timeseries provided as array

Usage

timeseries_dataset_from_array(
  data,
  targets,
  sequence_length,
  sequence_stride = 1L,
  sampling_rate = 1L,
  batch_size = 128L,
  shuffle = FALSE,
  ...,
  seed = NULL,
  start_index = NULL,
  end_index = NULL
)

Value

A tf.data.Dataset instance. If targets was passed, the dataset yields batches of two items: (batch_of_sequences, batch_of_targets). If not, the dataset yields only batch_of_sequences.

Arguments

data: array or eager tensor containing consecutive data points (timesteps). The first axis is expected to be the time dimension.
targets: Targets corresponding to timesteps in data. targets[i] should be the target corresponding to the window that starts at index i (see example 2 below). Pass NULL if you don't have target data (in this case the dataset will only yield the input data).
sequence_length: Length of the output sequences (in number of timesteps).
sequence_stride: Period between successive output sequences. For stride s, output samples would start at index data[i], data[i + s], data[i + (2 * s)], etc.
sampling_rate: Period between successive individual timesteps within sequences. For rate r, timesteps data[i], data[i + r], ... data[i + sequence_length] are used for create a sample sequence.
batch_size: Number of timeseries samples in each batch (except maybe the last one).
shuffle: Whether to shuffle output samples, or instead draw them in chronological order.
...: For backwards and forwards compatibility, ignored presently.
seed: Optional int; random seed for shuffling.
start_index, end_index: Optional int (1 based); data points earlier than start_index or later then end_index will not be used in the output sequences. This is useful to reserve part of the data for test or validation.

Example 1

Consider indices 0:99. With sequence_length=10, sampling_rate=2, sequence_stride=3, shuffle=FALSE, the dataset will yield batches of sequences composed of the following indices:

First sequence:  0  2  4  6  8 10 12 14 16 18
Second sequence: 3  5  7  9 11 13 15 17 19 21
Third sequence:  6  8 10 12 14 16 18 20 22 24
...
Last sequence:   78 80 82 84 86 88 90 92 94 96

In this case the last 3 data points are discarded since no full sequence can be generated to include them (the next sequence would have started at index 81, and thus its last step would have gone over 99).

Example 2

Temporal regression.

Consider an array data of scalar values, of shape (steps). To generate a dataset that uses the past 10 timesteps to predict the next timestep, you would use:

steps <- 100
# data is integer seq with some noise
data <- array(1:steps + abs(rnorm(steps, sd = .25)))
inputs_data <- head(data, -10) # drop last 10
targets <- tail(data, -10)    # drop first 10
dataset <- timeseries_dataset_from_array(
  inputs_data, targets, sequence_length=10)
library(tfdatasets)
dataset_iterator <- as_iterator(dataset)
repeat {
  batch <- iter_next(dataset_iterator)
  if(is.null(batch)) break
  c(input, target) %<-% batch
  stopifnot(exprs = {
    # First sequence: steps [1-10]
    # Corresponding target: step 11
    all.equal(as.array(input[1, ]), data[1:10])
    all.equal(as.array(target[1]), data[11])
    all.equal(as.array(input[2, ]), data[2:11])
    all.equal(as.array(target[2]), data[12])
    all.equal(as.array(input[3, ]), data[3:12])
    all.equal(as.array(target[3]), data[13])
  })
}

Example 3

Temporal regression for many-to-many architectures.

Consider two arrays of scalar values X and Y, both of shape (100). The resulting dataset should consist of samples with 20 timestamps each. The samples should not overlap. To generate a dataset that uses the current timestamp to predict the corresponding target timestep, you would use:

X <- seq(100)
Y <- X*2
sample_length <- 20
input_dataset <- timeseries_dataset_from_array(
  X, NULL, sequence_length=sample_length, sequence_stride=sample_length)
target_dataset <- timeseries_dataset_from_array(
  Y, NULL, sequence_length=sample_length, sequence_stride=sample_length)
library(tfdatasets)
dataset_iterator <-
  zip_datasets(input_dataset, target_dataset) %>%
  as_array_iterator()
while(!is.null(batch <- iter_next(dataset_iterator))) {
  c(inputs, targets) %<-% batch
  stopifnot(
    all.equal(inputs[1,], X[1:sample_length]),
    all.equal(targets[1,], Y[1:sample_length]),
    # second sample equals output timestamps 20-40
    all.equal(inputs[2,], X[(1:sample_length) + sample_length]),
    all.equal(targets[2,], Y[(1:sample_length) + sample_length])
  )
}

Example

int_sequence <- seq(20)
dummy_dataset <- timeseries_dataset_from_array(
  data = head(int_sequence, -3), # drop last 3
  targets = tail(int_sequence, -3), # drop first 3
  sequence_length = 3,
  start_index = 3,
  end_index = 9,
  batch_size = 2
)
library(tfdatasets)
dummy_dataset_iterator <- as_array_iterator(dummy_dataset)
repeat {
  batch <- iter_next(dummy_dataset_iterator)
  if (is.null(batch)) # iterator exhausted
    break
  c(inputs, targets) %<-% batch
  for (r in 1:nrow(inputs))
    cat(sprintf("input: [ %s ]  target: %s\n",
                paste(inputs[r,], collapse = " "), targets[r]))
  cat("---------------------------\n") # demark batchs
}

Will give output like:

input: [ 3 4 5 ]  target: 6
input: [ 4 5 6 ]  target: 7
---------------------------
input: [ 5 6 7 ]  target: 8
input: [ 6 7 8 ]  target: 9
---------------------------
input: [ 7 8 9 ]  target: 10

Details

This function takes in a sequence of data-points gathered at equal intervals, along with time series parameters such as length of the sequences/windows, spacing between two sequence/windows, etc., to produce batches of timeseries inputs and targets.