LRP
pnpxai.explainers.lrp
LRPBase
Bases: ZennitExplainer
Base class for LRPUniformEpsilon, LRPEpsilonGammaBox, LRPEpsilonPlus, and LRPEpsilonAlpha2Beta1 explainers.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
model
|
Module
|
The PyTorch model for which attribution is to be computed. |
required |
zennit_composite
|
Composite
|
The Composite object applies canonizers and register hooks to modules. One Composite instance may only be applied to a single module at a time. |
required |
layer
|
Optional[Union[Union[str, Module], Sequence[Union[str, Module]]]]
|
The target module to be explained |
None
|
n_classes
|
Optional[int]
|
Number of classes |
None
|
forward_arg_extractor
|
Optional[ForwardArgumentExtractor]
|
A function that extracts forward arguments from the input batch(s) where the attribution scores are assigned. |
None
|
additional_forward_arg_extractor
|
Optional[ForwardArgumentExtractor]
|
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
Bach S., Binder A., Montavon G., Klauschen F., M¨uller K.-R., and Samek. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation.
zennit_composite = zennit_composite
instance-attribute
layer = layer
instance-attribute
EXPLANATION_TYPE: ExplanationType = 'attribution'
class-attribute
instance-attribute
SUPPORTED_MODULES = []
class-attribute
instance-attribute
TUNABLES = {}
class-attribute
instance-attribute
model = model.eval()
instance-attribute
forward_arg_extractor = forward_arg_extractor
instance-attribute
additional_forward_arg_extractor = additional_forward_arg_extractor
instance-attribute
device
property
n_classes = n_classes
instance-attribute
__init__(model: Module, zennit_composite: Composite, forward_arg_extractor: Optional[ForwardArgumentExtractor] = None, additional_forward_arg_extractor: Optional[ForwardArgumentExtractor] = None, layer: Optional[TargetLayerOrListOfTargetLayers] = None, n_classes: Optional[int] = None) -> None
explainer(model) -> Union[Gradient, LayerGradient]
attribute(inputs: Union[Tensor, Tuple[Tensor]], targets: Tensor) -> Union[Tensor, Tuple[Tensor]]
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. |
__repr__()
copy()
set_kwargs(**kwargs)
get_tunables() -> Dict[str, Tuple[type, dict]]
Returns a dictionary of tunable parameters for the explainer.
Returns:
| Type | Description |
|---|---|
Dict[str, Tuple[type, dict]]
|
Dict[str, Tuple[type, dict]]: Dictionary of tunable parameters. |
__init_subclass__() -> None
LRPUniformEpsilon
Bases: LRPBase
LRPUniformEpsilon 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. |
required |
epsilon
|
Union[float, Callable[[Tensor], Tensor]]
|
The epsilon value. |
0.25
|
stabilizer
|
Union[float, Callable[[Tensor], Tensor]]
|
The stabilizer value |
1e-06
|
zennit_canonizers
|
Optional[List[Canonizer]]
|
An optional list of canonizers. Canonizers modify modules temporarily such that certain attribution rules can properly be applied. |
None
|
layer
|
Optional[Union[Union[str, Module], Sequence[Union[str, Module]]]]
|
The target module to be explained |
None
|
n_classes
|
Optional[int]
|
Number of classes |
None
|
**kwargs
|
Keyword arguments that are forwarded to the base implementation of the Explainer |
required |
SUPPORTED_MODULES = [Linear, Convolution, LSTM, RNN, Attention]
class-attribute
instance-attribute
epsilon = epsilon
instance-attribute
stabilizer = stabilizer
instance-attribute
zennit_canonizers = zennit_canonizers
instance-attribute
EXPLANATION_TYPE: ExplanationType = 'attribution'
class-attribute
instance-attribute
TUNABLES = {}
class-attribute
instance-attribute
model = model.eval()
instance-attribute
forward_arg_extractor = forward_arg_extractor
instance-attribute
additional_forward_arg_extractor = additional_forward_arg_extractor
instance-attribute
device
property
n_classes = n_classes
instance-attribute
zennit_composite = zennit_composite
instance-attribute
layer = layer
instance-attribute
__init__(model: Module, epsilon: Union[float, Callable[[Tensor], Tensor]] = 0.25, stabilizer: Union[float, Callable[[Tensor], Tensor]] = 1e-06, zennit_canonizers: Optional[List[Canonizer]] = None, forward_arg_extractor: Optional[ForwardArgumentExtractor] = None, additional_forward_arg_extractor: Optional[ForwardArgumentExtractor] = None, layer: Optional[TargetLayerOrListOfTargetLayers] = None, n_classes: Optional[int] = None) -> None
get_tunables() -> Dict[str, Tuple[type, dict]]
Provides Tunable parameters for the optimizer
Tunable parameters
epsilon (float): Value can be selected in the range of range(1e-6, 1)
__repr__()
copy()
set_kwargs(**kwargs)
attribute(inputs: Union[Tensor, Tuple[Tensor]], targets: Tensor) -> Union[Tensor, Tuple[Tensor]]
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. |
__init_subclass__() -> None
explainer(model) -> Union[Gradient, LayerGradient]
LRPEpsilonGammaBox
Bases: LRPBase
LRPEpsilonGammaBox explainer.
Supported Modules: Convolution
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
model
|
Module
|
The PyTorch model for which attribution is to be computed. |
required |
low
|
float
|
The lowest possible value for computing gamma box |
-3.0
|
high
|
float
|
The highest possible value for computing gamma box |
3.0
|
gamma
|
float
|
The gamma value for computing gamma box |
0.25
|
epsilon
|
Union[float, Callable[[Tensor], Tensor]]
|
The epsilon value. |
1e-06
|
stabilizer
|
Union[float, Callable[[Tensor], Tensor]]
|
The stabilizer value |
1e-06
|
zennit_canonizers
|
Optional[List[Canonizer]]
|
An optional list of canonizers. Canonizers modify modules temporarily such that certain attribution rules can properly be applied. |
None
|
layer
|
Optional[Union[Union[str, Module], Sequence[Union[str, Module]]]]
|
The target module to be explained |
None
|
n_classes
|
Optional[int]
|
Number of classes |
None
|
**kwargs
|
Keyword arguments that are forwarded to the base implementation of the Explainer |
required |
SUPPORTED_MODULES = [Convolution]
class-attribute
instance-attribute
low = low
instance-attribute
high = high
instance-attribute
epsilon = epsilon
instance-attribute
gamma = gamma
instance-attribute
stabilizer = stabilizer
instance-attribute
zennit_canonizers = zennit_canonizers
instance-attribute
EXPLANATION_TYPE: ExplanationType = 'attribution'
class-attribute
instance-attribute
TUNABLES = {}
class-attribute
instance-attribute
model = model.eval()
instance-attribute
forward_arg_extractor = forward_arg_extractor
instance-attribute
additional_forward_arg_extractor = additional_forward_arg_extractor
instance-attribute
device
property
n_classes = n_classes
instance-attribute
zennit_composite = zennit_composite
instance-attribute
layer = layer
instance-attribute
__init__(model: Module, low: float = -3.0, high: float = 3.0, epsilon: Union[float, Callable[[Tensor], Tensor]] = 1e-06, gamma: float = 0.25, stabilizer: Union[float, Callable[[Tensor], Tensor]] = 1e-06, zennit_canonizers: Optional[List[Canonizer]] = None, forward_arg_extractor: Optional[ForwardArgumentExtractor] = None, additional_forward_arg_extractor: Optional[ForwardArgumentExtractor] = None, layer: Optional[TargetLayerOrListOfTargetLayers] = None, n_classes: Optional[int] = None) -> None
get_tunables() -> Dict[str, Tuple[type, dict]]
Provides Tunable parameters for the optimizer
Tunable parameters
epsilon (float): Value can be selected in the range of range(1e-6, 1)
gamma (float): Value can be selected in the range of range(1e-6, 1)
__repr__()
copy()
set_kwargs(**kwargs)
attribute(inputs: Union[Tensor, Tuple[Tensor]], targets: Tensor) -> Union[Tensor, Tuple[Tensor]]
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. |
__init_subclass__() -> None
explainer(model) -> Union[Gradient, LayerGradient]
LRPEpsilonPlus
Bases: LRPBase
LRPEpsilonPlus explainer.
Supported Modules: Convolution
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
model
|
Module
|
The PyTorch model for which attribution is to be computed. |
required |
epsilon
|
Union[float, Callable[[Tensor], Tensor]]
|
The epsilon value. |
1e-06
|
stabilizer
|
Union[float, Callable[[Tensor], Tensor]]
|
The stabilizer value |
1e-06
|
zennit_canonizers
|
Optional[List[Canonizer]]
|
An optional list of canonizers. Canonizers modify modules temporarily such that certain attribution rules can properly be applied. |
None
|
layer
|
Optional[Union[Union[str, Module], Sequence[Union[str, Module]]]]
|
The target module to be explained |
None
|
n_classes
|
Optional[int]
|
Number of classes |
None
|
**kwargs
|
Keyword arguments that are forwarded to the base implementation of the Explainer |
required |
SUPPORTED_MODULES = [Convolution]
class-attribute
instance-attribute
epsilon = epsilon
instance-attribute
stabilizer = stabilizer
instance-attribute
zennit_canonizers = zennit_canonizers
instance-attribute
EXPLANATION_TYPE: ExplanationType = 'attribution'
class-attribute
instance-attribute
TUNABLES = {}
class-attribute
instance-attribute
model = model.eval()
instance-attribute
forward_arg_extractor = forward_arg_extractor
instance-attribute
additional_forward_arg_extractor = additional_forward_arg_extractor
instance-attribute
device
property
n_classes = n_classes
instance-attribute
zennit_composite = zennit_composite
instance-attribute
layer = layer
instance-attribute
__init__(model: Module, epsilon: Union[float, Callable[[Tensor], Tensor]] = 1e-06, stabilizer: Union[float, Callable[[Tensor], Tensor]] = 1e-06, zennit_canonizers: Optional[List[Canonizer]] = None, forward_arg_extractor: Optional[ForwardArgumentExtractor] = None, additional_forward_arg_extractor: Optional[ForwardArgumentExtractor] = None, layer: Optional[TargetLayerOrListOfTargetLayers] = None, n_classes: Optional[int] = None) -> None
get_tunables() -> Dict[str, Tuple[type, dict]]
Provides Tunable parameters for the optimizer
Tunable parameters
epsilon (float): Value can be selected in the range of range(1e-6, 1)
__repr__()
copy()
set_kwargs(**kwargs)
attribute(inputs: Union[Tensor, Tuple[Tensor]], targets: Tensor) -> Union[Tensor, Tuple[Tensor]]
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. |
__init_subclass__() -> None
explainer(model) -> Union[Gradient, LayerGradient]
LRPEpsilonAlpha2Beta1
Bases: LRPBase
LRPEpsilonAlpha2Beta1 explainer.
Supported Modules: Convolution
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
model
|
Module
|
The PyTorch model for which attribution is to be computed. |
required |
epsilon
|
Union[float, Callable[[Tensor], Tensor]]
|
The epsilon value. |
1e-06
|
stabilizer
|
Union[float, Callable[[Tensor], Tensor]]
|
The stabilizer value |
1e-06
|
zennit_canonizers
|
Optional[List[Canonizer]]
|
An optional list of canonizers. Canonizers modify modules temporarily such that certain attribution rules can properly be applied. |
None
|
layer
|
Optional[Union[Union[str, Module], Sequence[Union[str, Module]]]]
|
The target module to be explained |
None
|
n_classes
|
Optional[int]
|
Number of classes |
None
|
**kwargs
|
Keyword arguments that are forwarded to the base implementation of the Explainer |
required |
SUPPORTED_MODULES = [Convolution]
class-attribute
instance-attribute
epsilon = epsilon
instance-attribute
stabilizer = stabilizer
instance-attribute
zennit_canonizers = zennit_canonizers
instance-attribute
EXPLANATION_TYPE: ExplanationType = 'attribution'
class-attribute
instance-attribute
TUNABLES = {}
class-attribute
instance-attribute
model = model.eval()
instance-attribute
forward_arg_extractor = forward_arg_extractor
instance-attribute
additional_forward_arg_extractor = additional_forward_arg_extractor
instance-attribute
device
property
n_classes = n_classes
instance-attribute
zennit_composite = zennit_composite
instance-attribute
layer = layer
instance-attribute
__init__(model: Module, epsilon: Union[float, Callable[[Tensor], Tensor]] = 1e-06, stabilizer: Union[float, Callable[[Tensor], Tensor]] = 1e-06, zennit_canonizers: Optional[List[Canonizer]] = None, forward_arg_extractor: Optional[ForwardArgumentExtractor] = None, additional_forward_arg_extractor: Optional[ForwardArgumentExtractor] = None, layer: Optional[TargetLayerOrListOfTargetLayers] = None, n_classes: Optional[int] = None) -> None
get_tunables() -> Dict[str, Tuple[type, dict]]
Provides Tunable parameters for the optimizer
Tunable parameters
epsilon (float): Value can be selected in the range of range(1e-6, 1)
__repr__()
copy()
set_kwargs(**kwargs)
attribute(inputs: Union[Tensor, Tuple[Tensor]], targets: Tensor) -> Union[Tensor, Tuple[Tensor]]
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. |