All of the augmentation methods expect a numpy array x of size (batch, time_steps, channel) where batch is the size of the dataset or batch, time_steps is the number of time steps, and channel is the number of dimensions. Even if 1D time series are used, channel should still be 1.
import utils.augmentation as augAdding jittering, or noise, to the time series.
aug.jitter(x, sigma=0.03)x : 3D numpy array
Numpy array of time series in format (batch, time_steps, channel).
sigma : float
Standard deviation of the distribution to be added.
3D numpy array
Numpy array of generated data of equal size of the input x.
Scaling each time series by a constant amount.
aug.scaling(x, sigma=0.1)x : 3D numpy array
Numpy array of time series in format (batch, time_steps, channel).
sigma : float
Standard deviation of the scaling constant.
3D numpy array
Numpy array of generated data of equal size of the input x.
For 1D time series, randomly flipping. For multivariate time series, flipping as well as axis shuffling.
aug.rotation(x)x : 3D numpy array
Numpy array of time series in format (batch, time_steps, channel).
3D numpy array
Numpy array of generated data of equal size of the input x.
A special case of 2D rotation where the pattern is spacially rotated around the center.
aug.rotation2d(x, sigma=0.2)x : 3D numpy array
Numpy array of time series in format (batch, time_steps, channel).
sigma : float
Standard deviation of the rotation amount.
3D numpy array
Numpy array of generated data of equal size of the input x.
Random permutation of segments. A random number of segments is used, up to max_segments.
aug.permutation(x, max_segments=5, seg_mode="equal")x : 3D numpy array
Numpy array of time series in format (batch, time_steps, channel).
max_segments : int
The maximum number of segments to use. The minimum number is 1.
seg_mode : str
equal uses equal sized segments and random uses randomly sized segments.
3D numpy array
Numpy array of generated data of equal size of the input x.
The magnitude of each time series is multiplied by a curve created by cubicspline with a set number of knots at random magnitudes.
aug.magnitude_warp(x, sigma=0.2, knot=4)Based on: T. T. Um et al, "Data augmentation of wearable sensor data for parkinson’s disease monitoring using convolutional neural networks," in ACM ICMI, pp. 216-220, 2017.
x : 3D numpy array
Numpy array of time series in format (batch, time_steps, channel).
sigma : float
Standard deviation of the random magnitudes.
knot : int
Number of hills/valleys.
3D numpy array
Numpy array of generated data of equal size of the input x.
Random smooth time warping.
aug.time_warp(x, sigma=0.2, knot=4)Based on: T. T. Um et al, "Data augmentation of wearable sensor data for parkinson’s disease monitoring using convolutional neural networks," in ACM ICMI, pp. 216-220, 2017.
x : 3D numpy array
Numpy array of time series in format (batch, time_steps, channel).
sigma : float
Standard deviation of the random magnitudes of the warping path.
knot : int
Number of hills/valleys on the warping path.
3D numpy array
Numpy array of generated data of equal size of the input x.
Cropping the time series by the reduce_ratio.
aug.window_slice(x, reduce_ratio=0.9)Based on: A. Le Guennec, S. Malinowski, R. Tavenard, "Data Augmentation for Time Series Classification using Convolutional Neural Networks," in ECML/PKDD Workshop on Advanced Analytics and Learning on Temporal Data, 2016.
x : 3D numpy array
Numpy array of time series in format (batch, time_steps, channel).
sigma : float
Standard deviation of the random magnitudes of the warping path.
knot : int
Number of hills/valleys on the warping path.
3D numpy array
Numpy array of generated data of equal size of the input x.
Randomly warps a window by scales.
aug.window_warp(x, window_ratio=0.1, scales=[0.5, 2.])Based on: A. Le Guennec, S. Malinowski, R. Tavenard, "Data Augmentation for Time Series Classification using Convolutional Neural Networks," in ECML/PKDD Workshop on Advanced Analytics and Learning on Temporal Data, 2016.
x : 3D numpy array
Numpy array of time series in format (batch, time_steps, channel).
window_ratio : float
Ratio of the window to the full time series.
scales : list of floats
A list ratios to warp the window by.
3D numpy array
Numpy array of generated data of equal size of the input x.
Uses SPAWNER by K. Kamycki et al.
aug.spawner(x, labels, sigma=0.05, verbose=0)Based on: K. Kamycki, T. Kapuscinski, M. Oszust, "Data Augmentation with Suboptimal Warping for Time-Series Classification," Sensors, vol. 20, no. 1, 2020.
x : 3D numpy array
Numpy array of time series in format (batch, time_steps, channel).
labels : 2D or 3D numpy array
Either list of integers or one hot of the labels.
sigma : float
Standard deviation of the jittering.
verbose : int
1 prints out a DTW matrix. 0 shows nothing.
3D numpy array
Numpy array of generated data of equal size of the input x.
Uses the Average Selected with Distance (ASD) version of DBA from G. Forestier et al.
aug.wdba(x, labels, batch_size=6, slope_constraint="symmetric", use_window=True)Based on: G. Forestier, F. Petitjean, H. A. Dau, G. I. Webb, E. Keogh, "Generating Synthetic Time Series to Augment Sparse Datasets," in IEEE ICDM, 2017.
x : 3D numpy array
Numpy array of time series in format (batch, time_steps, channel).
labels : 2D or 3D numpy array
Either list of integers or one hot of the labels.
batch_size : int
How many patterns to average.
slope_constraint : str
Slope constraint for DTW. "symmetric" or "asymmetric".
use_window : bool
Use a 10% boundary constraint window for DTW. True or False.
3D numpy array
Numpy array of generated data of equal size of the input x.
aug.random_guided_warp(x, labels, slope_constraint="symmetric", use_window=True, dtw_type="normal")Based on: B. K. Iwana, S. Uchida, "Time Series Data Augmentation for Neural Networks by Time Warping with a Discriminative Teacher," arXiv, 2020.
x : 3D numpy array
Numpy array of time series in format (batch, time_steps, channel).
labels : 2D or 3D numpy array
Either list of integers or one hot of the labels.
slope_constraint : str
Slope constraint for DTW. "symmetric" or "asymmetric".
use_window : bool
Use a 10% boundary constraint window for DTW. True or False.
dtw_type : str
Use DTW (normal) or shapeDTW (shape).
3D numpy array
Numpy array of generated data of equal size of the input x.
aug.discriminative_guided_warp(x, labels, batch_size=6, slope_constraint="symmetric", use_window=True, dtw_type="normal", use_variable_slice=True)Based on: B. K. Iwana, S. Uchida, "Time Series Data Augmentation for Neural Networks by Time Warping with a Discriminative Teacher," arXiv, 2020.
x : 3D numpy array
Numpy array of time series in format (batch, time_steps, channel).
labels : 2D or 3D numpy array
Either list of integers or one hot of the labels.
batch_size : int
How many patterns to search. np.ceil(batch_size/2.) are used as positive samples and np.floor(batch_size/2.) are used for negative samples.
slope_constraint : str
Slope constraint for DTW. "symmetric" or "asymmetric".
use_window : bool
Use a 10% boundary constraint window for DTW. True or False.
dtw_type : str
Use DTW (normal) or shapeDTW (shape).
use_variable_slice : bool
Slice by the inverse of how much the pattern is warped.
3D numpy array
Numpy array of generated data of equal size of the input x.