type-registry 1.2

Package for registering types and creating systems based on config files

Python Type Registry

The type-registry library provides the ability to easily register factory methods (or classes) that can be used in a configuration file to quickly construct complex objects (or models).

Example

Lets demonstrate how to easily use this library. There are a functions and decorators, but the most important is the register decorator. By registering a type, we are telling the system that when a configuration section has a special key __factory__ to construct that section into an object.

We start by registering our PytorchExample and give it a more usable name of pytorch-model.

import torch.nn as nn

from type_registry import register

@register('pytorch-model', input_size=1024, output_size=8)
def PytorchExample(nn.Module):
    def __init__(self, input_size, hidden_size, output_size):
      ...

Now that we have the type registered with the system, in the configuration we can supplied the details around its construction (in this case we are just providing the value of the hidden_size)

model:
  __factory__: pytorch-model
  hidden_size: ...

Constructing from config

Once we have our types registered and a config defined, we need to correctly load the configuration so that it will construct the types. This is accomplished using the load_yaml and load_yaml_str functions.

from type_registery import load_yaml_str

yaml_str = get_config()
result = load_yaml_str(yaml_str)

You can also have environment variables (or custom variables) automatically expanded in the same configuration file.

model:
  __factory__: simple-model
  use_aws: $USE_AWS

Viewing the Registery

To make working with the library a little easier, there is a function that can be called which will print out the details of the registery (at the time the call takes place).

from type_registery import print_registry

if __name__ == '__main__':
  print_registery(color=True)

Setting the color value to True will have the system print out the results using shell coloring (termcolor library). You can also have the output of the print_registry call target a file instead of stdout.

from type_registery import print_registry

if __name__ == '__main__':
  with open('registry.txt', 'w') as f:
    print_registry(file=f)