deepecho 0.1.3

Mixed-type multivariate time series modeling with generative adversarial networks.

An open source project from Data to AI Lab at MIT.

DeepEcho

• License: MIT
• Development Status: Pre-Alpha
• Homepage: https://github.com/sdv-dev/DeepEcho

Overview

DeepEcho is a Synthetic Data Generation Python library for mixed-type, multivariate time series. It provides:

1. Multiple models based both on classical statistical modeling of time series and the latest in Deep Learning techniques.
2. A robust benchmarking framework for evaluating these methods on multiple datasets and with multiple metrics.
3. Ability for Machine Learning researchers to submit new methods following our model and sample API and get evaluated.

Try it out now!

If you want to quickly discover DeepEcho, simply click the button below and follow the tutorials!

Join our Slack Workspace

If you want to be part of the SDV community to receive announcements of the latest releases, ask questions, suggest new features or participate in the development meetings, please join our Slack Workspace!

Install

Requirements

DeepEcho has been developed and tested on Python 3.6, 3.7 and 3.8

Also, although it is not strictly required, the usage of a virtualenv is highly recommended in order to avoid interfering with other software installed in the system where DeepEcho is run.

Install with pip

The easiest and recommended way to install DeepEcho is using pip:

pip install deepecho

This will pull and install the latest stable release from PyPi.

If you want to install from source or contribute to the project please read the Contributing Guide.

Quickstart

In this short quickstart, we show how to learn a mixed-type multivariate time series dataset and then generate synthetic data that resembles it.

We will start by loading the data and preparing the instance of our model.

from deepecho import PARModel
from deepecho.demo import load_demo

# Load demo data
data = load_demo()

# Define data types for all the columns
data_types = {
    'region': 'categorical',
    'day_of_week': 'categorical',
    'total_sales': 'continuous',
    'nb_customers': 'count',
}

model = PARModel(cuda=False)

If we want to use different settings for our model, like increasing the number of epochs or enabling CUDA, we can pass the arguments when creating the model:

model = PARModel(epochs=1024, cuda=True)

Notice that for smaller datasets like the one used on this demo, CUDA usage introduces more overhead than the gains it obtains from parallelization, so the process in this case is more efficient without CUDA, even if it is available.

Once we have created our instance, we are ready to learn the data and generate new synthetic data that resembles it:

# Learn a model from the data
model.fit(
    data=data,
    entity_columns=['store_id'],
    context_columns=['region'],
    data_types=data_types,
    sequence_index='date'
)

# Sample new data
model.sample(num_entities=5)

The output will be a table with synthetic time series data with the same properties to the demo data that we used as input.

What's next?

For more details about DeepEcho and all its possibilities and features, please check and run the tutorials.

If you want to see how we evaluate the performance and quality of our models, please have a look at the DeepEcho Benchmarking framework or Explore the obtained results

Also, please feel welcome to visit our contributing guide in order to help us developing new features or cool ideas!

Related Projects

SDV

SDV, for Synthetic Data Vault, is the end-user library for synthesizing data in development under the HDI Project. SDV allows you to easily model and sample relational datasets using DeepEcho thought a simple API. Other features include anonymization of Personal Identifiable Information (PII) and preserving relational integrity on sampled records.

CTGAN

CTGAN is a GAN based model for synthesizing tabular data. It's also developed by the MIT's Data to AI Lab and is under active development.

History

0.1.3 (2020-10-16)

This version includes several minor improvements to the PAR model and the way the sequences are generated:

• Sequences can now be generated without dropping the sequence index.
• The PAR model learns the min and max length of the sequence from the input data.
• NaN values are properly supported for both categorical and numerical columns.
• NaN values are generated for numerical columns only if there were NaNs in the input data.
• Constant columns can now be modeled.

0.1.2 (2020-09-15)

Add BasicGAN Model and additional benchmarking results.

0.1.1 (2020-08-15)

This release includes a few new features to make DeepEcho work on more types of datasets as well as to making it easier to add new datasets to the benchmarking framework.

• Add segment_size and sequence_index arguments to fit method.
• Add sequence_length as an optional argument to sample and sample_sequence methods.
• Update the Dataset storage format to add sequence_index and versioning.
• Separate the sequence assembling process in its own deepecho.sequences module.
• Add function make_dataset to create a dataset from a dataframe and just a few column names.
• Add notebook tutorial to show how to create a datasets and use them.

0.1.0 (2020-08-11)

First release.

Included Features:

• PARModel
• Demo dataset and tutorials
• Benchmarking Framework
• Support and instructions for benchmarking on a Kubernetes cluster.