"""Module containing a template class as an interface to ML model.
Subclasses implement model interfaces for different ML frameworks such as TensorFlow, PyTorch OR Sklearn.
All model interface methods are in dice_ml.model_interfaces"""
import pickle
import numpy as np
from dice_ml.constants import ModelTypes
from dice_ml.utils.exception import SystemException
from dice_ml.utils.helpers import DataTransfomer
[docs]class BaseModel:
def __init__(self, model=None, model_path='', backend='', func=None, kw_args=None):
"""Init method
:param model: trained ML Model.
:param model_path: path to trained model.
:param backend: ML framework. For frameworks other than TensorFlow or PyTorch,
or for implementations other than standard DiCE
(https://arxiv.org/pdf/1905.07697.pdf),
provide both the module and class names as module_name.class_name.
For instance, if there is a model interface class "SklearnModel"
in module "sklearn_model.py" inside the subpackage dice_ml.model_interfaces,
then backend parameter should be "sklearn_model.SklearnModel".
:param func: function transformation required for ML model. If func is None, then func will be the identity function.
:param kw_args: Dictionary of additional keyword arguments to pass to func. DiCE's data_interface is appended to the
dictionary of kw_args, by default.
"""
self.model = model
self.model_path = model_path
self.backend = backend
# calls FunctionTransformer of scikit-learn internally
# (https://scikit-learn.org/stable/modules/generated/sklearn.preprocessing.FunctionTransformer.html)
self.transformer = DataTransfomer(func, kw_args)
[docs] def load_model(self):
if self.model_path != '':
with open(self.model_path, 'rb') as filehandle:
self.model = pickle.load(filehandle)
[docs] def get_output(self, input_instance, model_score=True):
"""returns prediction probabilities for a classifier and the predicted output for a regressor.
:returns: an array of output scores for a classifier, and a singleton
array of predicted value for a regressor.
"""
input_instance = self.transformer.transform(input_instance)
if model_score:
if self.model_type == ModelTypes.Classifier:
return self.model.predict_proba(input_instance)
else:
return self.model.predict(input_instance)
else:
return self.model.predict(input_instance)
[docs] def get_gradient(self):
raise NotImplementedError
[docs] def get_num_output_nodes(self, inp_size):
temp_input = np.transpose(np.array([np.random.uniform(0, 1) for i in range(inp_size)]).reshape(-1, 1))
return self.get_output(temp_input).shape[1]
[docs] def get_num_output_nodes2(self, input_instance):
if self.model_type == ModelTypes.Regressor:
raise SystemException('Number of output nodes not supported for regression')
return self.get_output(input_instance).shape[1]