Explainers

An explainer computes feature attributions: a score per input feature (pixel, token, or segment) indicating how much that feature contributed to a model's output.

Explainer base class

All explainer configs inherit from ExplainerConfig, which itself inherits from ModuleConfig. The registered class is the config — there is no separate nested Config class:

@EXPLAINERS.register("grad/integrated_gradients")
class IntegratedGradientsExplainerConfig(GradExplainerConfig):
    __module_path__: ClassVar[str] = "torchxai.explainers.IntegratedGradientsExplainer"
    type: Literal["grad/integrated_gradients"] = "grad/integrated_gradients"
    n_steps: int = 50

All explainer configs implement a build(model, ...) method that instantiates the underlying TorchXAI/Captum explainer, returning an object with an explain(forward_func, inputs, targets, baselines, ...) interface.

Gradient-based methods (TorchXAI wrappers)

Wraps Captum attribution methods via TorchXAI:

Explainer name	Method	Requires baseline
`grad/saliency`	Saliency (gradient magnitude)	No
`grad/input_x_gradient`	Input × Gradient	No
`grad/integrated_gradients`	Integrated Gradients	Yes
`grad/deeplift`	DeepLIFT	Yes
`grad/deeplift_shap`	DeepLIFT SHAP	Yes
`grad/gradient_shap`	GradientSHAP	Yes
`grad/guided_backprop`	Guided Backpropagation	No

Perturbation-based methods

Explainer name	Method	Notes
`perturbation/feature_ablation`	Feature Ablation	Systematic occlusion
`perturbation/occlusion`	Occlusion (sliding window)	—
`perturbation/lime`	LIME	Surrogate model
`perturbation/kernel_shap`	Kernel SHAP	SHAP with LIME framework

Perturbation methods do not require gradients — they work by masking features and observing model output changes. This makes them applicable to non-differentiable models and black-box APIs.

Attention-based methods

Explainer name	Method
`attn/raw_attention`	Raw attention weights from the last attention layer
`attn/attention_rollout`	Attention rollout across all transformer layers
`attn/attention_flow`	Attention flow (graph-based propagation)

Attention explainers are specific to transformer architectures and require a TransformersEncoderModel backbone. They do not require gradient computation through the model's input embeddings.

Baseline

Explainer name	Method
`random`	Random attribution baseline (uniform noise)

Registry and config

Explainers are selected by type in the explanation pipeline config:

explainer:
  type: "grad/integrated_gradients"
  n_steps: 50

Swapping the explainer is a one-line config change — the pipeline, dataset, and metric computation remain identical. This makes systematic comparison across explainer methods straightforward.