SaliencyExplainer

Saliency explainer for computing gradient-based attributions.

This explainer computes saliency maps using gradients of the model output with respect to inputs. Supports both single-target and multi-target modes for both single-target and multi-target scenarios.

The saliency method computes the gradient of the output with respect to the input, providing a measure of how much each input feature contributes to the prediction. Raw gradients are returned (abs=False) to preserve sign information.

Parameters:

  • model

    (Module) –

    The PyTorch model whose output is to be explained.

  • multi_target

    (bool, default: False ) –

    Whether to use multi-target mode. When True, can compute attributions for multiple targets simultaneously. Defaults to False.

  • internal_batch_size

    (int, default: 64 ) –

    Batch size for internal computations. Defaults to 64.

  • grad_batch_size

    (int, default: 64 ) –

    Batch size for gradient computations. Defaults to 64.

Examples:

Single-target usage:

>>> import torch
>>> from torchxai.data_types import SingleTargetAcrossBatch
>>>
>>> model = torch.nn.Linear(10, 2)
>>> explainer = SaliencyExplainer(model)
>>> attributions = explainer.explain(
...     inputs=torch.randn(1, 10),
...     target=SingleTargetAcrossBatch(index=0),
... )
>>> attributions.shape   # (1, 10)

Multi-target usage:

>>> explainer_mt = SaliencyExplainer(model, multi_target=True)
>>> mt_attributions = explainer_mt.explain(
...     inputs=torch.randn(1, 10),
...     target=[SingleTargetAcrossBatch(index=0), SingleTargetAcrossBatch(index=1)],
... )
>>> len(mt_attributions), mt_attributions[0].shape   # 2, (1, 10)

Methods:

  • explain

    Compute saliency attributions for the given inputs.

Attributes:

  • model (Module) –

    The model used for attribution computation.

  • multi_target (bool) –

    Whether the explainer uses multi-target mode.

Source code in torchxai/explainers/_grad/_saliency.py
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
class SaliencyExplainer(FeatureAttributionExplainer):
    """Saliency explainer for computing gradient-based attributions.

    This explainer computes saliency maps using gradients of the model output
    with respect to inputs. Supports both single-target and multi-target modes
    for both single-target and multi-target scenarios.

    The saliency method computes the gradient of the output with respect to the input,
    providing a measure of how much each input feature contributes to the prediction.
    Raw gradients are returned (abs=False) to preserve sign information.

    Args:
        model: The PyTorch model whose output is to be explained.
        multi_target: Whether to use multi-target mode. When True, can compute
            attributions for multiple targets simultaneously. Defaults to False.
        internal_batch_size: Batch size for internal computations. Defaults to 64.
        grad_batch_size: Batch size for gradient computations. Defaults to 64.

    Examples:
        Single-target usage:
        >>> import torch
        >>> from torchxai.data_types import SingleTargetAcrossBatch
        >>>
        >>> model = torch.nn.Linear(10, 2)
        >>> explainer = SaliencyExplainer(model)
        >>> attributions = explainer.explain(
        ...     inputs=torch.randn(1, 10),
        ...     target=SingleTargetAcrossBatch(index=0),
        ... )
        >>> attributions.shape   # (1, 10)

        Multi-target usage:
        >>> explainer_mt = SaliencyExplainer(model, multi_target=True)
        >>> mt_attributions = explainer_mt.explain(
        ...     inputs=torch.randn(1, 10),
        ...     target=[SingleTargetAcrossBatch(index=0), SingleTargetAcrossBatch(index=1)],
        ... )
        >>> len(mt_attributions), mt_attributions[0].shape   # 2, (1, 10)
    """

    def _init_single_target_explanation_fn(self) -> Callable:
        """Initialize single-target saliency attribution function.

        Returns:
            Captum Saliency attribution function for single targets.
        """
        return partial(Saliency(self._model).attribute, abs=False)

    def _init_multi_target_explanation_fn(self) -> Callable:
        """Initialize multi-target saliency attribution function.

        Returns:
            MultiTargetSaliency attribution function for multiple targets.
        """
        return partial(
            MultiTargetSaliency(
                self._model, grad_batch_size=self._grad_batch_size
            ).attribute,
            abs=False,
        )

    def explain(
        self,
        inputs: TensorOrTupleOfTensorsGeneric,
        target: ExplanationTargetType | list[ExplanationTargetType],
        additional_forward_args: tuple[Any, ...] | None = None,
    ) -> TensorOrTupleOfTensorsGeneric | list[TensorOrTupleOfTensorsGeneric]:
        """Compute saliency attributions for the given inputs.

        Args:
            inputs: Input tensor(s) for attribution computation.
            target: An `ExplanationTargetType` (e.g. `SingleTargetAcrossBatch`) for single-target
                mode, or a list of them for multi-target mode.
            additional_forward_args: Additional arguments for model forward pass.

        Returns:
            Tensor in single-target mode. List of Tensors, one per target, in multi-target mode.

        Examples:
            >>> # Single tensor input (wrapped automatically)
            >>> attributions = explainer.explain(
            ...     inputs=torch.randn(2, 10), target=torch.tensor([0, 1])
            ... )
            >>>
            >>> # Multiple features (use OrderedDict)
            >>> attributions = explainer.explain(
            ...     inputs=OrderedDict(
            ...         {"feat1": torch.randn(2, 5), "feat2": torch.randn(2, 5)}
            ...     ),
            ...     target=torch.tensor([0, 1]),
            ... )
        """
        return self._default_explain(
            inputs=inputs,
            target=target,
            additional_forward_args=additional_forward_args,
        )

model property writable

model: Module

The model used for attribution computation.

multi_target property writable

multi_target: bool

Whether the explainer uses multi-target mode.

explain

explain(
    inputs: TensorOrTupleOfTensorsGeneric,
    target: ExplanationTargetType | list[ExplanationTargetType],
    additional_forward_args: tuple[Any, ...] | None = None,
) -> TensorOrTupleOfTensorsGeneric | list[TensorOrTupleOfTensorsGeneric]

Compute saliency attributions for the given inputs.

Parameters:

  • inputs

    (TensorOrTupleOfTensorsGeneric) –

    Input tensor(s) for attribution computation.

  • target

    (ExplanationTargetType | list[ExplanationTargetType]) –

    An ExplanationTargetType (e.g. SingleTargetAcrossBatch) for single-target mode, or a list of them for multi-target mode.

  • additional_forward_args

    (tuple[Any, ...] | None, default: None ) –

    Additional arguments for model forward pass.

Returns:

  • TensorOrTupleOfTensorsGeneric | list[TensorOrTupleOfTensorsGeneric]

    Tensor in single-target mode. List of Tensors, one per target, in multi-target mode.

Examples:

>>> # Single tensor input (wrapped automatically)
>>> attributions = explainer.explain(
...     inputs=torch.randn(2, 10), target=torch.tensor([0, 1])
... )
>>>
>>> # Multiple features (use OrderedDict)
>>> attributions = explainer.explain(
...     inputs=OrderedDict(
...         {"feat1": torch.randn(2, 5), "feat2": torch.randn(2, 5)}
...     ),
...     target=torch.tensor([0, 1]),
... )
Source code in torchxai/explainers/_grad/_saliency.py
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
def explain(
    self,
    inputs: TensorOrTupleOfTensorsGeneric,
    target: ExplanationTargetType | list[ExplanationTargetType],
    additional_forward_args: tuple[Any, ...] | None = None,
) -> TensorOrTupleOfTensorsGeneric | list[TensorOrTupleOfTensorsGeneric]:
    """Compute saliency attributions for the given inputs.

    Args:
        inputs: Input tensor(s) for attribution computation.
        target: An `ExplanationTargetType` (e.g. `SingleTargetAcrossBatch`) for single-target
            mode, or a list of them for multi-target mode.
        additional_forward_args: Additional arguments for model forward pass.

    Returns:
        Tensor in single-target mode. List of Tensors, one per target, in multi-target mode.

    Examples:
        >>> # Single tensor input (wrapped automatically)
        >>> attributions = explainer.explain(
        ...     inputs=torch.randn(2, 10), target=torch.tensor([0, 1])
        ... )
        >>>
        >>> # Multiple features (use OrderedDict)
        >>> attributions = explainer.explain(
        ...     inputs=OrderedDict(
        ...         {"feat1": torch.randn(2, 5), "feat2": torch.randn(2, 5)}
        ...     ),
        ...     target=torch.tensor([0, 1]),
        ... )
    """
    return self._default_explain(
        inputs=inputs,
        target=target,
        additional_forward_args=additional_forward_args,
    )