registry-factory 0.1.0

Abstract codebase with utilities to register generic modules.

abstract-codebase

An abstract codebase with utilities for registering generic modules with an optional configuration setup and accreditation system.

Installation

The codebase can be installed from PyPI using pip, or your package manager of choice, with

$ pip install abstract-codebase

Dependencies

No dependencies to use the minimal Registry functionality. The configuration setup depends on yaml and hydra.

Usage

RegistryFactory

The codebase provides a way to register generic modules into a codebase. First a specific Registry is created, e.g. for deep learning models.

from abstract_codebase.registration import RegistryFactory

class ModelRegistry(RegistryFactory):
    pass

Next, any models can be added to the ModelRegistry as such.

import torch.nn as nn

@ModelRegistry.register(call_name="simple_model")
class SimpleModel(nn.Module):

    def __init__(self, layer_sizes) -> None:
        super(SimpleModel, self).__init__()
        dropout_rate = 0.25
        self.layers = nn.ModuleList()
        for layer in range(len(layer_sizes)-2):
            self.layers.append(
                nn.Sequential(
                    nn.Linear(hidden_layers[layer], hidden_layers[layer+1]),
                    nn.ReLU(),
                    nn.Dropout(p=dropout_rate)
                )
            )
        self.layers.append(
            nn.Sequential(
                nn.Linear(hidden_layers[-2], hidden_layers[-1])
            )
        )

    def forward(self, x):
        for layer in self.layers:
            x = layer(x)
        return x

Configurations

Each registered module can be accompanied with a dataclass of settings with default values.

@ModelRegistry.register(call_name="simple_model")  
@dataclass(unsafe_hash=True)
class SimpleModelArguments():

    dropout_rate = 0.25

As such, the model can rather be defined as.

@ModelRegistry.register(call_name="simple_mlp")
class SimpleModel(nn.Module):

    def __init__(self, layer_sizes, args: SimpleModelArguments) -> None:
        super(SimpleMLP, self).__init__()
        dropout_rate = args.dropout_rate
        ...

Further, the Config class can be used to read and update configurations across all modules, through yaml files or the command line interface.

Accreditation

At registration of a module, additional information can be supplied such as author, credit type and more. This information can be used to collect a summary of the accreditation required for all modules used in a given script.

@ModelRegistry.register(
    call_name="simple_model",
    author="Author name",
    credit_type=CreditType.REFERENCE,
    additional_information="Reference published work in (link)."
)
class SimpleModel(nn.Module):
    ...

TODO exemplify an accreditation summery.

ModelRegistry.get_accreditation()

Code of Conduct

Everyone interacting in the codebase, issue trackers, chat rooms, and mailing lists is expected to follow the PyPA Code of Conduct.