Overview

In modelkit, a Model is simply a subclass of modelkit.Model that implements a _predict function.

from modelkit import Model

class MyModel(Model):
    def _predict(self, item):
        ...
        # prediction code goes here
        ...
        return result

And that's it!

With this little boilerplate code, you can now call the model to get predictions, have them batched, or exposed in an API, etc.

Instantiating Models

Simple models

Very simple Model objects can be created and instantiated straight away, that is when they do not load complex assets, or rely on other Model objects for their execution. For example:

from modelkit import Model

class MyModel(Model):
    def _predict(self, item):
        return item

m = MyModel()

Complex models

In general, however, to resolve assets, dependencies, etc. modelkit models need to be instantiated using ModelLibrary with a set of models.

Models are then accessed via ModelLibrary.get("some name").

For example, to load a model that has one dependency:

from modelkit import ModelLibrary, Model

# Create two models, with names
class MyModel(Model):
    CONFIGURATIONS = {
        "a_model" : {}
    }
    def _predict(self, item):
        return item

class MyComposedModel(Model):
    CONFIGURATIONS = {
        "my_favorite_model" : {
            "model_dependencies": {"a_model"}
        }
    }
    def _predict(self, item):
        return self.model_dependencies["a_model"].predict(item)

# Create the model library
# This loads the required models (including their dependencies)
library = ModelLibrary(
    required_models=["my_favorite_model"],
    models=[MyModel, MyComposedModel]
)

# This is only a dictionary lookup
model = library.get("my_favorite_model")

The library gives you access to all the models from a single object, and deals with instantiation of the necessary objects. Furthermore, modelkit encourages you to store your models in a Python package (see Organization).