Models

Argflow provides a number of wrappers for different models. At this point we provide support for all Keras models as well as for Bayesian classifier models, typically implemented through scikit-learn.

This is planned to be extended to conform to all ONNX (Open Neural Network Exchange) models, the option to implement adapters for all types of models is provided.

Abstract Class

To create a model adapter and ensure compatibility with the rest of Argflow, it is sufficient to provide an implementation of the following abstract class:

class Model(ABC):

    def __init__(self, model):
        super().__init__()
        self.model = model

    @abstractmethod
    def predict(self, x, *args, **kwargs):
        """
        Evaluate model on some input
        x       - input to model
        args    - some arguments, possibly a dict
        """
        pass

    @abstractmethod
    def save(self, path, *args, **kwargs):
        """
        Save model to some path

        x       - the path to save to
        """
        pass

Keras Adapter

We require adapters for compatibility with the explanation extraction functionality of Argflow, as well as for successful usage of the Argflow Hub API.

import argflow


class KerasModel(argflow.Model):

    def __init__(self, model):
        super().__init__(model)

    def predict(self, x, args=None):
        return self.model.predict(x)

    def save(self, path):
        self.model.save(path)

Bayesian Classifier Model Adapter Requirements

For creating usable adapters for bayesian classifiers, the user would need to provide to following functionality.

class BayesianClassifierModel(argflow.Model):

    def __init__(self, model):
        super().__init__(model)

    def predict(self, x, args=None):
        return self.model.predict(x)

    def predict_proba(self):
        pass

    def get_input_prior(self):
        pass

    def get_output_prior(self):
        pass

    def save(self, path):
        pass