This python package should help you to create TensorFlow datasets for time-series data.
This package is available on PyPI. You install it and all of its dependencies using pip:
Suppose you have a dataset in the following form:
import numpy as np
import pandas as pd
# make things determeinisteic
np.random.seed(1)
columns=['x1', 'x2', 'x3']
periods=48 * 14
test_df=pd.DataFrame(
index=pd.date_range(
start='1/1/1992',
periods=periods,
freq='30min'
),
data=np.stack(
[
np.random.normal(0,0.5,periods),
np.random.normal(1,0.5,periods),
np.random.normal(2,0.5,periods)
],
axis=1
),
columns=columns
)
test_df.head()
x1 x2 x3 1992-01-01 00:00:00 0.812173 1.205133 1.578044 1992-01-01 00:30:00 -0.305878 1.429935 1.413295 1992-01-01 01:00:00 -0.264086 0.550658 1.602187 1992-01-01 01:30:00 -0.536484 1.159828 1.644974 1992-01-01 02:00:00 0.432704 1.159077 2.005718
The factory class WindowedTimeSeriesDatasetFactory
is used to create a TensorFlow dataset from pandas dataframes, or other data sources as we will see later.
We will use it now to create a dataset with 48
historic time-steps as the input to predict a single time-step in the future.
from tensorflow_time_series_dataset.factory import WindowedTimeSeriesDatasetFactory as Factory
factory_kwargs=dict(
history_size=48,
prediction_size=1,
history_columns=['x1', 'x2', 'x3'],
prediction_columns=['x3'],
batch_size=4,
drop_remainder=True,
)
factory=Factory(**factory_kwargs)
ds1=factory(test_df)
ds1
This returns the following TensorFlow Dataset:
<_PrefetchDataset element_spec=(TensorSpec(shape=(4, 48, 3), dtype=tf.float32, name=None), TensorSpec(shape=(4, 1, 1), dtype=tf.float32, name=None))>
We can plot the result with the utility function plot_path
:
from tensorflow_time_series_dataset.utils.visualisation import plot_patch
githubusercontent="https://raw.githubusercontent.com/MArpogaus/tensorflow_time_series_dataset/master/"
fig=plot_patch(
ds1,
figsize=(8,4),
**factory_kwargs
)
fname='.images/example1.svg'
fig.savefig(fname)
f"[[{githubusercontent}{fname}]]"
Lets now increase the prediction size to 6
half-hour time-steps.
factory_kwargs.update(dict(
prediction_size=6
))
factory=Factory(**factory_kwargs)
ds2=factory(test_df)
ds2
This returns the following TensorFlow Dataset:
<_PrefetchDataset element_spec=(TensorSpec(shape=(4, 48, 3), dtype=tf.float32, name=None), TensorSpec(shape=(4, 6, 1), dtype=tf.float32, name=None))>
Again, lets plot the results to see what changed:
fig=plot_patch(
ds2,
figsize=(8,4),
**factory_kwargs
)
fname='.images/example2.svg'
fig.savefig(fname)
f"[[{githubusercontent}{fname}]]"
Preprocessors can be used to transform the data before it is fed into the model.
A Preprocessor can be any python callable.
In this case we will be using the a class called CyclicalFeatureEncoder
to encode our one-dimensional cyclical features like the time or weekday to two-dimensional coordinates using a sine and cosine transformation as suggested in [this blogpost](https://www.kaggle.com/avanwyk/encoding-cyclical-features-for-deep-learning).
import itertools
from tensorflow_time_series_dataset.preprocessors import CyclicalFeatureEncoder
encs = {
"weekday": dict(cycl_max=6),
"dayofyear": dict(cycl_max=366, cycl_min=1),
"month": dict(cycl_max=12, cycl_min=1),
"time": dict(
cycl_max=24 * 60 - 1,
cycl_getter=lambda df, k: df.index.hour * 60 + df.index.minute,
),
}
factory_kwargs.update(dict(
meta_columns=list(itertools.chain(*[[c+'_sin', c+'_cos'] for c in encs.keys()]))
))
factory=Factory(**factory_kwargs)
for name, kwargs in encs.items():
factory.add_preprocessor(CyclicalFeatureEncoder(name, **kwargs))
ds3=factory(test_df)
ds3
This returns the following TensorFlow Dataset:
<_PrefetchDataset element_spec=((TensorSpec(shape=(4, 48, 3), dtype=tf.float32, name=None), TensorSpec(shape=(4, 1, 8), dtype=tf.float32, name=None)), TensorSpec(shape=(4, 6, 1), dtype=tf.float32, name=None))>
Again, lets plot the results to see what changed:
fig=plot_patch(
ds3,
figsize=(8,4),
**factory_kwargs
)
fname='.images/example3.svg'
fig.savefig(fname)
f"[[{githubusercontent}{fname}]]"
Any Contributions are greatly appreciated! If you have a question, an issue or would like to contribute, please read our contributing guidelines.
Distributed under the Apache License 2.0
Marcel Arpogaus - [email protected]
Project Link: https://github.com/MArpogaus/tensorflow_time_series_dataset
Parts of this work have been funded by the Federal Ministry for the Environment, Nature Conservation and Nuclear Safety due to a decision of the German Federal Parliament (AI4Grids: 67KI2012A).