Fiddle Colab Series

Fiddle Colab Series#

Welcome!#

This series of colabs introduces Fiddle, a configuration library serving machine learning use cases.

Please send us feedback if you have any suggestions for improvements! Or if you built something cool and want to share it out, we’d love to link it up at the “Where to go from here” section.

General colab advice#

Most entries in this series are colab notebooks (click the blue buttons in each heading below), allowing you to run our tutorial code interactively. We encourage you to do that! Play around, change things, see what happens!

Codelab 1: Introduction / overview#

Open in colab

This colab walks through how to use Fiddle, using Flax and JAX as a motivating example. This tutorial includes an overview of the key Fiddle APIs, best practices, as well as a few tips/tricks.

Codelab 2: Basic API#

Open in colab

This colab walks through the core Fiddle abstractions and how they work. If you prefer learning in a more pedagogical than motivational style, you can start with this colab and then go through the above intro colab.

Codelab 3: Visualization / printing / codegen#

Open in colab

This colab introduces a variety of ways to visualize and print Fiddle configurations. These can be used to communicate experiments, debug errors, and make configuration more self-contained.

Codelab 4: auto_config usage#

Open in colab

This colab walks through how to use Fiddle’s “auto-config” functionality, building off of the introduction codelab above. Auto-config allows existing “glue code” functions that perform object and function wire-up to be transformed into functions that capture the wire-up as a fdl.Buildable, often with no or very minimal changes to the existing code. This minimizes the boilerplate necessary to create a Fiddle configuration.

Codelab 5: CLI flags#

The flags code lab walks through an example application that uses Fiddle and Fiddle’s support for overriding configurations with command line flags.

Codelab 6: select() and Tag APIs#

Open in colab

The select() and Tag APIs for Fiddle let users concisely change many values in a larger configuration structure, facilitating configuration codebases which are factorized into declaration of a base model, and experimental overrides.

The select() API makes it easy to set parameters across all occurrences of specific functions or classes within a config. For example:

# Set all Dropout classes to have rate 0.1.
select(root_cfg, nn.Dropout).set(rate=0.1)

Fiddle also allows values to be tagged with one or more tags, making it easy to set values that are shared in many places all at once. For example:

# Set all tagged dtypes, which may be on different functions/classes.
select(root_cfg, tag=ActivationDType).set(value=jnp.bfloat16)
Contents