PAIR-code/saliency

Health v2 · maintenance only

Flags: no_releases

Adoption not part of the score

996 stars · 190 forks observed · 2026-10-22

What it is AI-extracted, prompt v1, taxonomy v1, 2026-09-07, confidence 0.95

A framework-agnostic Python library implementing state-of-the-art saliency methods for visualizing and explaining deep neural network predictions, including XRAI, Blur IG, SmoothGrad, Integrated Gradients, Grad-CAM, and more. It also provides the Performance Information Curve (PIC) metric for evaluating saliency method quality without human input.

Use cases

• generate saliency maps for a neural network's image predictions
• explain why a CNN classified an image a certain way
• visualize which pixels influenced a model's decision with SmoothGrad
• apply XRAI to produce attribution masks for object detection models
• compute Guided Integrated Gradients attributions for my model
• evaluate the quality of different saliency methods using PIC
• run Grad-CAM on a TensorFlow or PyTorch model
• compare occlusion and integrated gradients explanations

When to choose

• you need explainability or interpretability visualizations for deep neural networks on images
• you want a framework-agnostic library that works with any ML framework via a call_model_function interface
• you need access to many saliency techniques (XRAI, Blur IG, SmoothGrad, Grad-CAM, etc.) in one package
• you want a human-independent metric (PIC) to evaluate saliency maps

When to avoid

• you need explainability for text or tabular models rather than image-based neural networks
• you want production-grade model monitoring or dashboards rather than research-grade saliency code
• you need a GUI or no-code tool for generating explanations

Facets

library · maturity stable

machine-learning deep-learning image-processing data-visualization machine-learning deep-learning computer-vision image-processing python cross-platform saliency-maps explainability interpretability xai tensorflow neural-network-visualization

Member repositories

For agents

markdown · JSON · MCP: product_card(name="PAIR-code/saliency")