[Gretel] Optimize config and training parameters for gretel-synthetics

[zredlined]

Problem Description

Some of the parameters in the gretel-synthetics implementation in SDGym can cause the model to fail during evaluation, and can be optimized for generating synthetic data (details below).

Expected behavior

In sdgym/synthesizers/gretel.py there are a few updates I'd recommend making:

• Add learning_rate as a parameter, set default to 0.001 as per Gretel docs.
• Add field_cluster_size as a tunable parameter
• on batcher.generate_all_batch_lines, set a default max_invalid. For larger datasets, the default value of 1000 can cause the model to unnecessarily terminate and during sampling.
• epochs can be set to a standard value (e.g. 100), no need to set epochs based on the number of columns. Early stopping and a validation set can be used to prevent overfitting.

Additional context

I'm happy to submit a PR with fixes and to compare against baseline config for tests, let me know if this would be okay. Cheers!

[katxiao]

Hi @zredlined, thanks for taking a look at the Gretel implementation in SDGym! Please feel free to submit a PR with the updates you have proposed. You can link it to this issue and we'll be happy to take a look.