An ExplanationPipeline wraps a ModelPipeline with the additional processing needed to produce and evaluate attributions. It is a ConfigurableModule, registered in EXPLANATION_PIPELINES.

BaseExplanationPipeline

class BaseExplanationPipeline(ConfigurableModule[T_ExplanationPipelineConfig]):
    def __init__(self, config, model_pipeline: ModelPipeline): ...

The base class holds a reference to the ModelPipeline and provides:

• forward_func — a callable wrapping the model pipeline's forward pass in a form that attribution methods can differentiate.
• Attribution step — runs the explainer over a batch and returns BatchExplanationState (attributions + metadata).
• Metric step — computes explainability metrics on stored attributions.
• Caching — attributions and metric results are written to HDF5 via ExplanationStateCacher and MetricDataCacher.

Concrete pipeline types

ImagePipeline (image_classification)

Handles image attribution:

• Converts BaseDataInstance → image tensors via the model pipeline's transform.
• Manages baseline generation: BaselineGenerator subclasses produce the reference input (zero, Gaussian noise, blurred image) for gradient-based methods that require a baseline.
• Manages feature segmentation: FeatureSegmentor subclasses group pixels into superpixels or regions, enabling segment-level attribution rather than pixel-level.
• Returns BatchExplanationState with per-pixel (or per-segment) attribution maps.

SequencePipeline

Gradient-based token-level text attribution. Registered names:

• Converts tokenized sequences to embedding-space representations.
• Runs gradient-based methods in embedding space (for IntegratedGradients, DeepLIFT, etc.).
• Aggregates embedding-dimension attributions to token-level scores.
• Returns BatchExplanationState with per-token attribution scores.

AttnSequencePipeline

Attention-based alternative to gradient computation. Registered names:

• Extracts attention weights directly from transformer layers (no gradient computation needed).
• Supports attention rollout (multiplying attention matrices across layers) for multi-layer models.
• Returns the same BatchExplanationState format as SequencePipeline, so downstream metric computation is identical.

Baseline generators

Gradient-based attribution methods (Integrated Gradients, DeepLIFT, GradientSHAP) require a baseline — a reference input representing "no signal". BaselineGenerator subclasses in atria_insights/baseline_generators/ produce baselines appropriate for each modality:

Feature segmentors

For image attribution, raw pixel-level maps are often noisy. FeatureSegmentor subclasses group pixels into semantically meaningful regions before attribution:

• SLIC superpixel segmentation
• Grid-based segments
• Segment-anything (SAM) masks

Segment-level attribution is more robust and easier to visualize.