fob 0.0.5

a library for building configurable, repeatable experiments

fob

fob is a library for building configurable, repeatable experiments. It comes with a fob command for running and inspecting experiment pipelines, which are built from configurable ingredients.

Ingredient

An ingredient is the basic building block of a pipeline. Ingredients are configured using their __init__method, which allows unified Python and CLI APIs. They need to:

• Subclass the Ingredientclass, which allows the ingredient to be recognized (and provides many convenience methods)
• Decorate their __init__ function with @configurable, which makes the arguments passed to __init__ accessible from the self.cfg dict. This allows the arguments to be configured from the CLI and tracked for artifact provenance
• Decorate select methods with @cacheable, which saves the method's output based on (1) the method's arguments and (2) the configuration in self.cfg. When possible, decorated methods load output from disk rather than recomputing it. The @cacheable decorator takes an output type as an argument, such as @cacheable(NumpyOutput). Other valid types include PathOutput, PickleOutput, StringOutput, etc.

Command

The fob command can run arbitrary methods within a pipeline. It uses the syntax:

fob <class> [method] [method-args] [--output DIR] [--search-path DIR] [--config config-args]

where

• methodcan be a method of class or of an object in class.cfg
• method-args is a space-separated list of key=valarguments to pass to method
• --outputspecifies a directory to store @cacheable output (and to find it in)
• --search-path specifies directories to look for @cacheable output in
• config-args is a space-separated list of key=valarguments to initialize class