A small library for managing deep learning models, hyper-parameters and datasets


Keywords
command-line-interface, deep-learning, hyperparameter, keras, machine-learning, python, tensorflow, tensorflow-datasets
License
Apache-2.0
Install
pip install zookeeper==1.0.5

Documentation

Zookeeper

GitHub Actions Codecov PyPI - Python Version PyPI PyPI - License Code style: black

A small library for configuring modular applications.

Installation

pip install zookeeper

Components

The fundamental building blocks of Zookeeper are components. The @component decorator is used to turn classes into components. These component classes can have configurable fields, which are declared with the Field constructor and class-level type annotations. Fields can be created with or without default values. Components can also be nested, with ComponentFields, such that child componenents can access the field values defined on their parents.

For example:

from zookeeper import component

@component
class ChildComponent:
    a: int = Field()                          # An `int` field with no default set
    b: str = Field("foo")                     # A `str` field with default value `"foo"`

@component
class ParentComponent:
    a: int = Field()                          # The same `int` field as the child
    child: ChildComponent = ComponentField()  # A nested component field, of type `ChildComponent`

After instantiation, components can be 'configured' with a configuration dictionary, containing values for a tree of nested fields. This process automatically injects the correct values into each field.

If a child sub-component declares a field which already exists in some containing ancestor component, then it will pick up the value that's set on the parent, unless a 'scoped' value is set on the child.

For example:

from zookeeper import configure

p = ParentComponent()

configure(
    p,
    {
        "a": 5,
        "child.a": 4,
    }
)

>>> 'ChildComponent' is the only concrete component class that satisfies the type
>>> of the annotated parameter 'ParentComponent.child'. Using an instance of this
>>> class by default.

print(p)

>>> ParentComponent(
>>>     a = 5,
>>>     child = ChildComponent(
>>>         a = 4,
>>>         b = "foo"
>>>     )
>>> )

Tasks and the CLI

The @task decorator is used to define Zookeeper tasks and can be applied to any class that implements an argument-less run method. Such tasks can be run through the Zookeeper CLI, with parameter values passed in through CLI arguments (configure is implicitly called).

For example:

from zookeeper import cli, task

@task
class UseChildA:
    parent: ParentComponent = ComponentField()
    def run(self):
        print(self.parent.child.a)

@task
class UseParentA(UseChildA):
    def run(self):
        print(self.parent.a)

if __name__ == "__main__":
    cli()

Running the above file then gives a nice CLI interface:

python test.py use_child_a
>>> ValueError: No configuration value found for annotated parameter 'UseChildA.parent.a' of type 'int'.

python test.py use_child_a a=5
>>> 5

python test.py use_child_a a=5 child.a=3
>>> 3

python test.py use_parent_a a=5 child.a=3
>>> 5

Using Zookeeper to define Larq or Keras experiments

See examples/larq_experiment.py for an example of how to use Zookeeper to define all the necessary components (dataset, preprocessing, and model) of a Larq experiment: training a BinaryNet on MNIST. This example can be easily adapted to other Larq or Keras models and other datasets.