lucid.register_model¶
The register_model decorator registers a model function and its metadata into a JSON registry file. This is particularly useful for managing and organizing models in the lucid library.
Function Signature¶
def register_model(func: Callable) -> Callable
Parameters¶
func (Callable): The function that instantiates and returns a model class instance.
Description¶
When a function decorated with @register_model is called, the following occurs:
The registry file (JSON) is loaded from REGISTRY_PATH.
The function is invoked to instantiate the model.
The model’s name, family (class name), parameter size, and architecture are extracted.
If the model’s name already exists in the registry, the model is returned without making changes to the registry.
Otherwise, the new model’s metadata is added to the registry and saved back to the JSON file.
Metadata Captured¶
The following metadata is registered:
name (str): Name of the function that creates the model.
family (str): The class name of the instantiated model.
param_size (int): The parameter size of the model.
arch (str): The architecture’s name derived from the package structure, removing the lucid.models. prefix.
Examples¶
import lucid
import lucid.nn as nn
class SimpleModel(nn.Module):
def __init__(self):
super().__init__()
self.layer1 = nn.Linear(10, 100)
self.layer2 = nn.linear(100, 10)
@register_model
def my_simple_model(*args, **kwargs) -> SimpleModel:
return SimpleModel(*args, **kwargs)
model = my_simple_model()
This example shows a function my_simple_model that creates an instance of a SimpleModel class. When the function is called, the model’s metadata is registered into the JSON registry file.