Skip to content

VarGrad

VarGrad

Bases: SmoothGrad

VarGrad explainer.

Supported Modules: Linear, Convolution, LSTM, RNN, Attention

Parameters:

Name Type Description Default
model Module

The PyTorch model for which attribution is to be computed. noise_level (float): The noise level added during attribution. n_iter (int): The number of iterations, the input is modified. layer (Optional[Union[Union[str, Module], Sequence[Union[str, Module]]]]): The target module to be explained.

 required
forward_arg_extractor Optional[Callable[[Tuple[Tensor]], Union[Tensor, Tuple[Tensor]]]]

A function that extracts forward arguments from the input batch(s) where the attribution scores are assigned.

 None
additional_forward_arg_extractor Optional[Callable[[Tuple[Tensor]], Union[Tensor, Tuple[Tensor]]]]

A secondary function that extract additional forward arguments from the input batch(s).

None
**kwargs

Keyword arguments that are forwarded to the base implementation of the Explainer

 required
Reference

Lorenz Richter, Ayman Boustati, Nikolas Nüsken, Francisco J. R. Ruiz, Ömer Deniz Akyildiz. VarGrad: A Low-Variance Gradient Estimator for Variational Inference.

Source code in pnpxai/explainers/var_grad.py 
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
class VarGrad(SmoothGrad):
	"""
    VarGrad explainer.

    Supported Modules: `Linear`, `Convolution`, `LSTM`, `RNN`, `Attention`

    Parameters:
        model (Module): The PyTorch model for which attribution is to be computed.
		noise_level (float): The noise level added during attribution.
		n_iter (int): The number of iterations, the input is modified.
		layer (Optional[Union[Union[str, Module], Sequence[Union[str, Module]]]]): The target module to be explained.
        forward_arg_extractor: A function that extracts forward arguments from the input batch(s) where the attribution scores are assigned.
        additional_forward_arg_extractor: A secondary function that extract additional forward arguments from the input batch(s).		
        **kwargs: Keyword arguments that are forwarded to the base implementation of the Explainer

    Reference:
        Lorenz Richter, Ayman Boustati, Nikolas Nüsken, Francisco J. R. Ruiz, Ömer Deniz Akyildiz. VarGrad: A Low-Variance Gradient Estimator for Variational Inference.
    """

	SUPPORTED_MODULES = [Linear, Convolution, LSTM, RNN, Attention]

	def __init__(
		self,
		model: Module,
		noise_level: float=.1,
		n_iter: int=20,
        forward_arg_extractor: Optional[Callable[[Tuple[Tensor]], Union[Tensor, Tuple[Tensor]]]]=None,
        additional_forward_arg_extractor: Optional[Callable[[Tuple[Tensor]], Union[Tensor, Tuple[Tensor]]]]=None,
        layer: Optional[Union[Union[str, Module], Sequence[Union[str, Module]]]]=None,
        n_classes: Optional[int]=None,
	) -> None:
		super().__init__(
			model=model,
			noise_level=noise_level,
			n_iter=n_iter,
			forward_arg_extractor=forward_arg_extractor,
			additional_forward_arg_extractor=additional_forward_arg_extractor,
			layer=layer,
			n_classes=n_classes,
		)

	def attribute(
		self,
		inputs: Union[Tensor, Tuple[Tensor]],
		targets: Tensor,
	) -> Union[Tensor, Tuple[Tensor]]:
		"""
        Computes attributions for the given inputs and targets.

        Args:
            inputs (torch.Tensor): The input data.
            targets (torch.Tensor): The target labels for the inputs.

        Returns:
            torch.Tensor: The result of the explanation.
        """
		forward_args, additional_forward_args = self._extract_forward_args(inputs)
		with self.attributor() as attributor:
			avg_grads, avg_grads_sq = attributor.forward(
				forward_args,
				targets,
				additional_forward_args,
				return_squared=True,
			)
		vargrads = tuple(
			avg_grad_sq - avg_grad for avg_grad_sq, avg_grad in zip(
				format_into_tuple(avg_grads_sq),
				format_into_tuple(avg_grads),
			)
		)
		return format_out_tuple_if_single(vargrads)

attribute(inputs, targets)

Computes attributions for the given inputs and targets.

Parameters:

Name Type Description Default
inputs Tensor

The input data.

 required
targets Tensor

The target labels for the inputs.

 required

Returns:

Type Description
Union[Tensor, Tuple[Tensor]]

torch.Tensor: The result of the explanation.
Source code in pnpxai/explainers/var_grad.py 
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
	def attribute(
		self,
		inputs: Union[Tensor, Tuple[Tensor]],
		targets: Tensor,
	) -> Union[Tensor, Tuple[Tensor]]:
		"""
        Computes attributions for the given inputs and targets.

        Args:
            inputs (torch.Tensor): The input data.
            targets (torch.Tensor): The target labels for the inputs.

        Returns:
            torch.Tensor: The result of the explanation.
        """
		forward_args, additional_forward_args = self._extract_forward_args(inputs)
		with self.attributor() as attributor:
			avg_grads, avg_grads_sq = attributor.forward(
				forward_args,
				targets,
				additional_forward_args,
				return_squared=True,
			)
		vargrads = tuple(
			avg_grad_sq - avg_grad for avg_grad_sq, avg_grad in zip(
				format_into_tuple(avg_grads_sq),
				format_into_tuple(avg_grads),
			)
		)
		return format_out_tuple_if_single(vargrads)