simple-copoco 0.2.1

Effortlessly transition from YAML configuration to Python object management with the Simple Config Powered Code package.

Overview

Effortlessly transition from YAML configuration to Python object management with the Simple Config Powered Code package.

Installation

pip install simple-copoco

Why don't you just use OmegaConf + Hydra?

OmegaConf + Hydra duet is fantastic and great (also literally). So there are some reasons to search for alternatives:

1. Simplicity: The simple-copoco utility is designed to be lightweight and easy to use. If you have a project with straightforward configuration needs, simple-copoco provides an intuitive and easy to learn way to manage objects via configuration without the need for additional complexity.
2. Integration of configuration with object registration: simple-copoco offers features for object registration, allowing you to register and manage classes dynamically. If your project involves dynamic class instantiation based on configuration, simple-copoco provides a convenient way to achieve this.
3. Grid Configurations: simple-copoco includes a GridManager for handling grids of YAML configurations, which can be useful for scenarios where you need to explore multiple configuration combinations in you experiments (whatever you do).

Getting Started

Handling YAML Configurations
Consider the following YAML configuration:

berries:
  color: blue
  amount: 5
coconuts:
  amount: 1

In Python:

from simple_copoco import ConfigManager

# config path
config_path = '~/path/to/config.yaml'

# optional, possibly wide, general config with many defaults
config_template_path = '~/path/to/template.yaml'

# Option 1: template will be updated with provided config
cfg_manager = ConfigManager(config_path, config_template_path)
# Option 2: template doesn't exist
cfg_manager = ConfigManager(config_path)

my_config = cfg_manager.cfg

# access config attributes
how_many_berries_to_buy = my_config.berries.amount
what_color_of_berries_to_buy = my_config.berries.color

# unpack config attributes, like it is a dictionary
do_something_with_beries(**my_config.berries)

# dump config
cfg_manager.save_to_disk('~/where/to/save.yaml')

Managing Grids of YAML Configurations
Consider the following YAML grid:

berries:
  color: [red, green, yellow]
  amount: [5, 10, 15]
coconuts:
  amount: [1, 3]

In Python:

from simple_copoco import GridManager

# Specify paths
grid_path = '~/path/to/grid.yaml'
config_path = '~/path/to/config.yaml'
config_template_path = '~/path/to/template.yaml' # optional

# Grid manager will create all possible combinations of settings from grid.yaml
grid_manager = GridManager(grid_path, config_path, config_template_path)

# Grid manager is a generator
first_config_manager = grid_manager.next()
second_config_manager = grid_manager.next()

# Total number of config managers
total_of_config_managers = len(grid_manager)

first_config = first_config_manager.cfg

Object Registration

from torch import nn
from simple_copoco import Register, register_as, RegistrableMixin

@register_as('utility_layer')
class CustomLayer(nn.Module):
    def __init__(self):
        super(CustomLayer, self).__init__()
    ...

register = Register(nn.Module, include_parent=False)
# register.buit_ins -> Dictionary of all classes inheriting
# directly or indirectly from nn.Module that weren't registered
# with @register_as.
# register.utility_layer -> Dictionary of classes inheriting from
# nn.Module registered as "utility_layer"

# another use case, when submodules of a package are not being imported
# with their parent. It happens, when the package has complex directory tree
# and uses blank __init__.py files.
from .. import custom_layers # custom_modules has some submodules

register = Register(nn.Module, custom_layers)

@register_as('classifier_head')
class CustomHead(nn.Module):
    def __init__(self):
        super(CustomHead, self).__init__()
    ...

# You can do the same stuff with functions :)
# Registered function is stored in the register as a wrapped object,
# but can be called the normal way.
@register_as('functions')
def some_interesting_function(a, b, c):
    return a * b * c

assert register.functions['some_interesting_function'](1, 2, 3) == 6 # True

# Update the register with the current namespace
register.update_register()
# register.classifier_head -> dict of all the classes in the namespace
# inheriting from nn.Module registered as "classifier_head"

# Add a class (that has been registered after Register initialisation) to the
# register manually
register.extend(CustomHead)

class AnyClass:
    ...

def some_func():
    ...

# Register anything
register.extend('other_classes', 'AnyClass', AnyClass)
# name in the register doesn't need to the same
register.extend('functions', 'some_function', some_func)
  
# Register hint :)  
register = Register(RegistrableMixin)  
# registers anything within the namespace that inherits from RegistrableMixin
# because @register_as() injects RegistrableMixin as a parent class
# (but this inheritance will be ignored in children if passes_on_children == False).

Putting It All Together
Consider the following configuration:

core:
  type: OtherModule
  params:
    a: 1
    b: 2
head:
  type: CustomHead
  params:
    x: 1
    y: 1

In Python:

from torch.nn import Module
from simple_copoco import Register, register_as
from simple_copoco import ConfigManager


@register_as('core', passes_on_children=True)
class CustomModule(Module):
    ...

class OtherModule(CustomModule):
    ...

@register_as('head')
class CustomHead(Module):
    ...

register = Register(Module)
cfg_man = ConfigManager('path/to/some_config.yaml', 'path/to/config_template.yaml')
cfg = cfg_man.cfg

# Create dynamic code based on configuration.
# Attributes are accessible by dot notation, but you can use ** to unpack them,
# just like with dictionaries.
# It makes experimentation with different architectures and many sets of parameters a breeze.

class MemesClassifier(Module):
    
    def __init__(self, cfg, register):
        self.model_core = register.core[cfg.core.type](**cfg.core.params)
        self.classifier_head = register.head[cfg.head.type](**cfg.head.params)
        
    ...