Synthetic data for tables and time-series, some assembly required

What it does

Gretel Synthetics is a Python library for generating artificial structured and time-series datasets using deep-learning generators. It wraps an augmented CTGAN variant called ACTGAN—adding better memory handling and column autodetection for tabular data—and a PyTorch DoppelGANger-style model for sequential readings. The library bills itself as permissive and open-source, but getting started involves hand-picking the correct versions of Torch, SDV, and TensorFlow to avoid collisions.

The interesting bit

Rather than paper over its complexity, the README openly treats the library as a collection of distinct GAN implementations and even warns that SDV’s PyTorch can conflict with Timeseries DGAN’s PyTorch. That transparency about dependency hell is almost a feature.

Key highlights

• ACTGAN extends CTGAN with column autodetection, automatic transformation, and lower memory usage for tabular synthetic data.
• Timeseries DGAN is a PyTorch DoppelGANger implementation aimed at sensor-style sequential data.
• Core dependencies—Torch, SDV, and TensorFlow—are intentionally left out of the base install to prevent version lock-in.
• A GPU setup script is provided, though it targets Ubuntu 18.04 specifically.

Caveats

• Dependency management is strictly DIY: the README warns that SDV’s PyTorch may conflict with Timeseries DGAN’s PyTorch, and you must install Torch, SDV, and TensorFlow manually.
• The GPU setup script is hard-coded for Ubuntu 18.04, so modern environments may need custom tweaks.
• The repository description advertises unstructured text generation and differential privacy, but the README documentation centers on tabular and time-series GANs without elaborating on those broader claims.

Verdict

Data scientists who need a permissive, code-first toolkit for synthetic tabular or time-series data and don’t mind wrangling Python environments will find value here. If you want a batteries-included, one-pip-install solution—or clear documentation on text generation and differential privacy—look elsewhere.