gridconf 0.1.0

Lightweight config loader with imports, string interpolation, and grid-search axis expansion

Simple Hackable Configmanager

[!IMPORTANT] This is a small hackable configuration manager. It just does what you expect it to do. Load from json/yaml, importing, string interpolation, expanding axes, manual overwrite, yield mutable /immutable dictionaries whose keys you may access by attributes.

Basic Usage

from your.utils import make_model, make_optim
from yourparent.configmanager import (
    load_config, ImmutableAttributDict, AttributeDict
)

config : AttributeDict = load_config('tests/model.yaml', make_immutable=False)[0]

model = make_model(**conf.model.kwargs)
optimizer = make_optim(**conf.optim.kwargs)

# hashable (for use in jax.jitted regions)
hash(mutable_config)

Features - Imports

# config/mymodel/kwargs1.yaml
param1 : 1
param2 : "heythere buddy!"

# config/model.yaml
model:
    kwargs:
        __import__: "configs/mymodel/kwargs1.yaml"

... resolves to ...

model:
    kwargs:
        param1 : 1
        param2 : "heythere buddy!"

[!NOTE] Per default, all nested dictionaries are merged instead of overwritten. In case you want to specify custom overwrite behaviour, please take a look further down below.

Features - Axes

[!NOTE] A really nice addition is the use of expandable axes. This lets you specify a (nested) grid of axes in a single .yaml file.

model:
    kwargs:
        __axis__:
            - batchnorm: false
              use_bottleneck: false
            - batchnorm: true
              use_bottleneck: false
              additional_param: 42
            - {} # defaults

optim:    
    kwargs:
        __axis__:
            - path: 'optax.adamw'
            - path: 'your.module.optimizer'

Turned into a bunch of configurations:

"""
    this will return a list of 3 * 2 configs, as we have two
    axes which are expanded. Axis one `model.kwargs` has three entries,
    axis `optim.kwargs` has 2 entries. 
"""
configs = load_config('tests/model.yaml')

Features - String interpolation

data:
    shape: [42, 1]

seed: 0

optim:
    lr: 0.001
    info: "heytherebuddy"

model:
    kwargs:
        input_shape : "$(data.shape)" # resolves to [42, 1]
    checkpoint: "run-seed:$(seed)-lr:$(optim.lr)-checkpoint-.pickle" # resolves to "run-seed:0-lr:0.001-checkpoint.pickle"

$(optim.info) : $(optim.info) # resolves to "heytherebuddy" : "heytherebuddy"

Custom Overwrites

Sometimes, you wouldn't want dictionaries to be merged. If you're instead looking for a hard-overwrite option, here it is:

__import__: "some_default_setup.yaml"

optim:
    path: "my.module.optim"
    __overwrite__(kwargs):
        please_just_keep : "this keyword argument, instead of merging"

Instead of importing a default optimizer which would probably have the key kwargs defined, we overwrite. This resolves to something like:

model:
    ...

...

optim:
    path: "my.module.optim"
    kwargs:
        please_just_keep : "this keyword argument, instead of merging"