Golden Recipes

Great place to start

Golden Recipes are the best place to start if you are new to LEIP Design, new to Machine Learning, or if you are just looking for quick-and-easy results.

A recipe is a set of instructions defining every operation in a machine learning workflow: from the early stages of data ingestion, through training and evaluation, all the way to compilation for a specific hardware target. Golden Recipes are recipes that have been pre-validated by our team at Latent AI. They are the best of all the recipes we have experimented with.

How do we determine that a recipe is "Golden"?

We look at several factors:

1. Golden recipes have been trained on diverse sets of data and have consistently outperformed most other recipes tested.
2. Golden recipes have been optimized to produce optimal deployed artifacts: the final deployable stands out in one or more of the metrics we are benchmarking: accuracy, inference time, memory consumption, and file size.

What should I know about Golden Recipes?

1. Golden recipes are always evolving. Our team is continuously validating these recipes with more and more datasets to increase the chances that, when you pick a recipe, you have a high likelihood of getting great results.
2. We are always crunching for new Golden Recipes. At Latent AI, our developers are adding more model architectures and deployment conditions, to give you more recipe options, to further increase the likelihood that you will find just the right one for your data and deployment conditions.
3. While we pre-validated these recipes by testing their performance with diverse datasets and use cases, there is no guarantee that they will have similarly outstanding performance in your data. You should always look to train a diverse set of Golden Recipes on your data to hedge your bets and increase your chances of success.

How do I use Golden Recipes?

Golden Recipes are set up to be straighforward and automatable. They contain all the information needed to train, evaluate, and deploy your models.

The easiest way to get started is to download our database of Golden Recipes, filter for recipes that meet your deployment constraints (like hardware target, memory consumption, inference time, file size, licensing, etc), and train as many recipes as you can afford to train. Then it's a matter of picking the best perfomer. Consult the GRDB Model Training and Optimization guide for step-by-step instructions.

What if I'm a seasoned ML practitioner? Can I design my own recipe?

Absolutely! The Recipes API allows you to modify any recipe, or even start one from scratch. Consult the API Reference for more information.