Storing and Loading data in aeon¶
Getting data into the correct data structure is fundamental. This notebook describes
the data structures used in aeon and links to more complex use cases. aeon models
two abstract data types: single series and collections of series.
A single time series can be univariate (each observation is a single value) or
multivariate (each observation is a vector). We say that the length of the vector
(its dimension) is the number of channels, which in code we denote n_channels.
The length of the series is called the number of timepoints, or n_timepoints in
code. We generally store a single series in a 2D numpy array with shape
(n_channels, n_timepoints), though data can be passed the other way around in some
cases using an axis parameter. Series estimators work with a univariate series stored
as a 1D numpy array, but will internally convert to 2D.
import numpy as np
from aeon.visualisation import plot_series, plot_series_collection
# Univariate series length 100
X = np.random.random((1, 100))
print(X.shape)
plot_series(X)
(1, 100)
(<Figure size 1600x400 with 1 Axes>, <Axes: >)
# Three channel multivariate series length 200
X = np.array(
[
np.sin(np.arange(0, np.pi * 4, np.pi * 4 / 200)),
np.sin(np.arange(0, np.pi * 8, np.pi * 8 / 200)),
np.sin(np.arange(0, np.pi * 16, np.pi * 16 / 200)),
]
)
print(X.shape)
plot_series(X)
(3, 200)
(<Figure size 1600x400 with 1 Axes>, <Axes: >)
# Collection of 5 univariate series of length 50
X = np.random.random((5, 1, 50))
plot_series_collection(X)
(<Figure size 1600x400 with 1 Axes>, <Axes: >)
# Collection of 5 multivariate time series with 26 channels, length 100
X = np.random.random((5, 26, 100))
plot_series_collection(X)
(<Figure size 1600x400 with 1 Axes>, <Axes: >)
Why this shape?¶
We get asked this a lot. Packages like tensorflow assume (n_cases, n_timepoints, n_channels) rather than (n_cases, n_channels, n_timepoints). tl;dr: its a
decision we made early on because many estimators iterate over channels, and we are
not changing it now. It is simple to reshape.
Supervised collection algorithms (e.g. classifiers , regressors and some collection
transformers) require a target variable for training. In keeping with
scikit-learn, we denote the target y. For classification, y should be a numpy
array of strings or integers. For regression, y should be a numpy array of integers
or floats. The DummyClassifier used below simply predicts the majority class in
the train data. The DummyRegressor predicts the mean of the training target variable.
import numpy as np
from aeon.classification import DummyClassifier
from aeon.regression import DummyRegressor
X = np.random.random((5, 1, 10)) # Collection of five series
y1 = np.array(["run", "walk", "run", "run", "walk"])
y2 = np.array([1, 0, 1, 1, 0])
y3 = np.array([0.5, 1.11, 0.545, -0.3675, 2.662])
cls = DummyClassifier()
reg = DummyRegressor()
cls.fit(X, y1)
print(cls.predict(X))
cls.fit(X, y2)
print(cls.predict(X))
reg.fit(X, y2)
print(reg.predict(X))
reg.fit(X, y3)
print(reg.predict(X))
['run' 'run' 'run' 'run' 'run'] [1 1 1 1 1] [0.6 0.6 0.6 0.6 0.6] [0.8899 0.8899 0.8899 0.8899 0.8899]
If your collection of series are unequal length, we store them in a list of 2D numpy arrays. See the unequal length collections. aeon does not currently support single series with unequal length channels.
aeon ships with a range of datasets used in examples and testing. The provided
datasets notebook describes all these datasets.
aeon provides functions to load data directly from text files in several formats.
The data loading notebook describes the formats of our
supported files and how to load them into aeon data structures.
You can load data directly from the Time Series Machine Learning archive and the Monash time series forecasting sites. More details in the load from web notebook.