quantus.metrics.axiomatic.completeness module

This module contains the implementation of the Completeness metric.

final class quantus.metrics.axiomatic.completeness.Completeness(abs: bool = False, normalise: bool = True, normalise_func: Callable[[ndarray], ndarray] | None = None, normalise_func_kwargs: Dict[str, Any] | None = None, output_func: Callable | None = None, perturb_baseline: str = 'black', perturb_func: Callable | None = None, perturb_func_kwargs: Dict[str, Any] | None = None, return_aggregate: bool = False, aggregate_func: Callable | None = None, default_plot_func: Callable | None = None, disable_warnings: bool = False, display_progressbar: bool = False, **kwargs)

Bases: Metric[List[float]]

Implementation of Completeness test by Sundararajan et al., 2017, also referred to as Summation to Delta by Shrikumar et al., 2017 and Conservation by Montavon et al., 2018.

Attribution completeness asks that the total attribution is proportional to the explainable evidence at the output/ or some function of the model output. Or, that the attributions add up to the difference between the model output F at the input x and the baseline b.

Assumptions:

  • This implementation does completeness test against logits, not softmax.

References:

1) Completeness - Mukund Sundararajan et al.: “Axiomatic attribution for deep networks.” International Conference on Machine Learning. PMLR, 2017. 2) Summation to delta - Avanti Shrikumar et al.: “Learning important features through propagating activation differences.” International Conference on Machine Learning. PMLR, 2017. 3) Conservation - Grégoire Montavon et al.: “Methods for interpreting and understanding deep neural networks.” Digital Signal Processing 73 (2018): 1-15.

Attributes:

  • _name: The name of the metric.

  • _data_applicability: The data types that the metric implementation currently supports.

  • _models: The model types that this metric can work with.

  • score_direction: How to interpret the scores, whether higher/ lower values are considered better.

Attributes:
disable_warnings

A helper to avoid polluting test outputs with warnings.

display_progressbar

A helper to avoid polluting test outputs with tqdm progress bars.

get_params

List parameters of metric.

Methods

__call__(model, x_batch, y_batch[, a_batch, ...])

This implementation represents the main logic of the metric and makes the class object callable.

batch_preprocess(data_batch)

If data_batch has no a_batch, will compute explanations.

custom_batch_preprocess(*, model, x_batch, ...)

Implement this method if you need custom preprocessing of data or simply for creating/initialising additional attributes or assertions before a data_batch can be evaluated.

custom_postprocess(*, model, x_batch, ...)

Implement this method if you need custom postprocessing of results or additional attributes.

custom_preprocess(*, model, x_batch, ...)

Implement this method if you need custom preprocessing of data, model alteration or simply for creating/initialising additional attributes or assertions.

evaluate_batch(model, x_batch, y_batch, ...)

This method performs XAI evaluation on a single batch of explanations.

explain_batch(model, x_batch, y_batch)

Compute explanations, normalise and take absolute (if was configured so during metric initialization.) This method should primarily be used if you need to generate additional explanation in metrics body. It encapsulates typical for Quantus pre- and postprocessing approach. It will do few things: - call model.shape_input (if ModelInterface instance was provided) - unwrap model (if ModelInterface instance was provided) - call explain_func - expand attribution channel - (optionally) normalise a_batch - (optionally) take np.abs of a_batch.

general_preprocess(model, x_batch, y_batch, ...)

Prepares all necessary variables for evaluation.

generate_batches(data, batch_size)

Creates iterator to iterate over all batched instances in data dictionary.

interpret_scores()

Get an interpretation of the scores.

plot([plot_func, show, path_to_save])

Basic plotting functionality for Metric class.
__call__(model, x_batch: ndarray, y_batch: ndarray, a_batch: ndarray | None = None, s_batch: ndarray | None = None, channel_first: bool | None = None, explain_func: Callable | None = None, explain_func_kwargs: Dict | None = None, model_predict_kwargs: Dict | None = None, softmax: bool | None = False, device: str | None = None, batch_size: int = 64, **kwargs) → List[float]

This implementation represents the main logic of the metric and makes the class object callable. It completes instance-wise evaluation of explanations (a_batch) with respect to input data (x_batch), output labels (y_batch) and a torch or tensorflow model (model).

Calls general_preprocess() with all relevant arguments, calls () on each instance, and saves results to evaluation_scores. Calls custom_postprocess() afterwards. Finally returns evaluation_scores.

Parameters:
model: torch.nn.Module, tf.keras.Model

A torch or tensorflow model that is subject to explanation.

x_batch: np.ndarray

A np.ndarray which contains the input data that are explained.

y_batch: np.ndarray

A np.ndarray which contains the output labels that are explained.

a_batch: np.ndarray, optional

A np.ndarray which contains pre-computed attributions i.e., explanations.

s_batch: np.ndarray, optional

A np.ndarray which contains segmentation masks that matches the input.

channel_first: boolean, optional

Indicates of the image dimensions are channel first, or channel last. Inferred from the input shape if None.

explain_func: callable

Callable generating attributions.

explain_func_kwargs: dict, optional

Keyword arguments to be passed to explain_func on call.

model_predict_kwargs: dict, optional

Keyword arguments to be passed to the model’s predict method.

softmax: boolean

Indicates whether to use softmax probabilities or logits in model prediction. This is used for this __call__ only and won’t be saved as attribute. If None, self.softmax is used.

device: string

Indicated the device on which a torch.Tensor is or will be allocated: “cpu” or “gpu”.

kwargs: optional

Keyword arguments.

Returns:
evaluation_scores: list

a list of Any with the evaluation scores of the concerned batch.

Examples:

# Minimal imports. >> import quantus >> from quantus import LeNet >> import torch

# Enable GPU. >> device = torch.device(“cuda:0” if torch.cuda.is_available() else “cpu”)

# Load a pre-trained LeNet classification model (architecture at quantus/helpers/models). >> model = LeNet() >> model.load_state_dict(torch.load(“tutorials/assets/pytests/mnist_model”))

# Load MNIST datasets and make loaders. >> test_set = torchvision.datasets.MNIST(root=’./sample_data’, download=True) >> test_loader = torch.utils.data.DataLoader(test_set, batch_size=24)

# Load a batch of inputs and outputs to use for XAI evaluation. >> x_batch, y_batch = iter(test_loader).next() >> x_batch, y_batch = x_batch.cpu().numpy(), y_batch.cpu().numpy()

# Generate Saliency attributions of the test set batch of the test set. >> a_batch_saliency = Saliency(model).attribute(inputs=x_batch, target=y_batch, abs=True).sum(axis=1) >> a_batch_saliency = a_batch_saliency.cpu().numpy()

# Initialise the metric and evaluate explanations by calling the metric instance. >> metric = Metric(abs=True, normalise=False) >> scores = metric(model=model, x_batch=x_batch, y_batch=y_batch, a_batch=a_batch_saliency)

__init__(abs: bool = False, normalise: bool = True, normalise_func: Callable[[ndarray], ndarray] | None = None, normalise_func_kwargs: Dict[str, Any] | None = None, output_func: Callable | None = None, perturb_baseline: str = 'black', perturb_func: Callable | None = None, perturb_func_kwargs: Dict[str, Any] | None = None, return_aggregate: bool = False, aggregate_func: Callable | None = None, default_plot_func: Callable | None = None, disable_warnings: bool = False, display_progressbar: bool = False, **kwargs)
Parameters:
abs: boolean

Indicates whether absolute operation is applied on the attribution, default=False.

normalise: boolean

Indicates whether normalise operation is applied on the attribution, default=True.

normalise_func: callable
Attribution normalisation function applied in case normalise=True.

If normalise_func=None, the default value is used, default=normalise_by_max.

normalise_func_kwargs: dict

Keyword arguments to be passed to normalise_func on call, default={}.

output_func: callable
Function applied to the difference between the model output at the input and the

baseline before metric calculation. If output_func=None, the default value is used, default=lambda x: x.

perturb_baseline: string
Indicates the type of baseline: “mean”, “random”, “uniform”, “black” or “white”,

default=”black”.

perturb_func: callable
Input perturbation function. If None, the default value is used,

default=baseline_replacement_by_indices.

perturb_func_kwargs: dict

Keyword arguments to be passed to perturb_func, default={}.

softmax: boolean

Indicates whether to use softmax probabilities or logits in model prediction, default=False.

return_aggregate: boolean

Indicates if an aggregated score should be produced over all instances.

aggregate_func: callable

A Callable to aggregate the scores per instance to one float.

default_plot_func: callable

Callable that plots the metrics result.

disable_warnings: boolean

Indicates whether the warnings are printed, default=False.

display_progressbar: boolean

Indicates whether a tqdm-progress-bar is printed, default=False.

kwargs: optional

Keyword arguments.

data_applicability: ClassVar[Set[DataType]] = {DataType.IMAGE, DataType.TABULAR, DataType.TIMESERIES}
evaluate_batch(model: ModelInterface, x_batch: ndarray, y_batch: ndarray, a_batch: ndarray, **kwargs) → List[bool]

This method performs XAI evaluation on a single batch of explanations. For more information on the specific logic, we refer the metric’s initialisation docstring.

Parameters:
model: ModelInterface

A ModelInterface that is subject to explanation.

x_batch: np.ndarray

The input to be evaluated on a batch-basis.

y_batch: np.ndarray

The output to be evaluated on a batch-basis.

a_batch: np.ndarray

The explanation to be evaluated on a batch-basis.

kwargs:

Unused.

Returns:
scores_batch:

The evaluation results.

evaluation_category: ClassVar[EvaluationCategory] = 'Axiomatic'
model_applicability: ClassVar[Set[ModelType]] = {ModelType.TF, ModelType.TORCH}
name: ClassVar[str] = 'Completeness'
score_direction: ClassVar[ScoreDirection] = 'higher'