Before you start training a machine learning model, you have to make important choices like the number of layers in a neural network, the embedding dimension, and the details of how training should proceed. These so-called “hyperparameters” are not learned during training and can have a dramatic impact on the quality of your model. In the case of generative models, optimized hyperparameters can be the difference between low- and high-quality synthetic data. 

It is therefore essential to systematically and objectively tune your model’s hyperparameters for your particular use case. That’s why today we are excited to announce Gretel Tuner, our new config-driven tool for efficiently sweeping the hyperparameters of Gretel Synthetics models.

Introducing Gretel Tuner 🎛️

We have built Gretel Tuner directly into our Python SDK. Our implementation was designed with the following features in mind:

1. Simplicity: Running Gretel Tuner only requires a YAML configuration and a single command using our SDK's new high-level interface. 
   
   
2. Efficiency: Gretel Tuner efficiently samples the search space by leveraging Optuna’s Bayesian optimization framework to home in on an optimal model configuration.
   
   
3. Customizability: In addition to all of Gretel’s synthetic data quality scores, Gretel Tuner supports custom user-implemented optimization metrics, making it possible to find optimal model parameters for highly specific tasks. 

The third point on custom metrics is especially exciting, since this makes it possible for our users to tune Gretel Tuner (😉) to their use cases. For example, we have found optimization metrics based on downstream machine learning tasks to be very effective at tuning hyperparameters to yield models that generate synthetic data of high fidelity and utility for machine learning.  

Ready to try Gretel Tuner? Start here 👇

The best way to start sweeping hyperparameters with Gretel Tuner is to work through our introductory and advanced Google Colab tutorials:

• Gretel Tuner: Introductory Tutorial — Learn the basics of how to configure Gretel Tuner to sweep the hyperparameters of a Gretel Synthetics model.

• ‍Gretel Tuner: Advanced Tutorial — Learn advanced features of Gretel Tuner, including how to implement custom optimization metrics and how to enforce arbitrary constraints on the sampled model configurations. 

To install Gretel Tuner locally, simply add the [tuner] option to the SDK installation command:

pip install "gretel-client[tuner]"

This will install the Gretel client along with the new tuner module and its associated dependencies. 

We’re excited to see what you’ll achieve by using the Gretel Tuner alongside the Gretel Synthetic Data Platform and our fully featured SDK!