Configuring machine learning experiments with Fiddle

Fiddle is the successor to gin, AFAICT, at Google for configuring ML experiments. It uses a code-based config system, where we define Python functions that change the configuration. It’s a slightly different approach. I’m not sure how to use it for things like hyperparam search, but it looks fine.

It generates a CLI via Google’s absl system. The keyword is Flag Support.

Docs are sparse; see:

• Fiddle documentation — Fiddle documentation
• Fiddle Tutorial
• Auto-Config Tutorial
• Fiddle Flag Support — Fiddle documentation

Fiddle Config Best Practices were not obvious to me at first, but this is what I have learned:

• Location: Store Fiddle configs (.py files) inside the source tree, typically in a dedicated directory (e.g., configs/), separate from core library code (e.g., src/).
• Dependency: Core business logic (ML code) should NOT import or depend on fiddle. It should accept standard Python arguments/objects.
• Integration: Use a main script/entry point to:
  1. Import the desired configuration from the configs/ module.
  2. Use fiddle.build() to materialize the configuration into Python objects.
  3. Pass these objects to your business logic functions/classes.

This should keep configuration separate from implementation for modularity, testability, and reusability. Business logic remains agnostic to the configuration system.

I wonder if this pairs well with other Google infrastructure such as their metadata store?