Getting Started

Installation

Install timeserio from PyPi using pip install timeserio, or from source using pip install -e .. You will need a working TensorFlow installation - the exact steps will depend on you OS and hardware. We maintain base docker images for our own use at https://hub.docker.com/r/octoenergy/tensorflow

Core concepts

timeserio builds on keras models and scikit-learn pipelines. It introduces two core abstraction levels: multinetwork and multimodel.

Multinetwork

A multinetwork allows defining multiple named, inter-connected, parametric keras models. These could be e.g. an encoder, a decoder, a trainable encoder-decoder model, or a classification model combining the encoder with several additional layers.

A multinetwork is defined by sub-classing timeserio.keras.multinetwork.MultiNetworkBase, and creating an instance of the new class.

A multinetwork provides the usual methods such fit() and predict(), but each of theses methods accepts a model keyword argument to use a specific named model, e.g.:

multinetwork.fit(X, X, model='autoencoder')
encoding = multinetwork.predict(X, model='encoder')

Multimodel

A multimodel combines a multinetwork with multiple pipelines that describe how to obtain feature arrays for each input of each of the models defined by the multinetwork.

A complex model within a multinetwork may have multiple inputs, each requiring specific feature pre-processing or encoding:

churn_rate = multinetwork.predict([age_normalized, gender, postcode_one_hot], model='churn_model')

A multinetwork will encapsulate the pipelines and allow simple usage, e.g.

churn_rate = multimodel.predict(user_df, model='churn_model')

A multimodel is defined as an instance of timeserio.multimodel.MultiModel.

Pipelines

Pipelines are scikit-learn transformers. We mostly use pipelines that take a pandas.DataFrame as input, and return a numpy array:

X = pipeline.fit_transform(df)

A helper class, timeserio.pipeline.MultiPipeline, allows grouping multiple pipelines into a single object for use in multimodels.

Examples and Tutorials

The best way to see how the concepts above fit together is to go through some examples.

• MNIST Autoencoder and Classifier example shows how to use MultiNetworkBase to implement a neural network architecture with multiple models, training paths and callbacks, and how to make use of pre-training and model freezing.

• MovieLens Recommender Engine Example demonstrates the use of pipelines, MultiPipeline, and MultiModel to feed data from pandas to a multi-inputs neural network.

• PV Forecasting Example (Part 1) demonstrates the use of a MultiModel for working with datetime features and multiple time series.

• PV Forecasting Example (Part 2) demonstrates the use of a MultiModel and batch generators for training auto-regressive time series models from large datasets.

We are working on adding more examples and tutorials, as well as improving documentation.