pytorch/captum

Features

• Interpretability Libraries - Explains and understands predictions from PyTorch models by attributing importance to input features, neurons, and layers.
• Integrated Gradient Attribution Methods - Implements integrated gradients to attribute predictions to input features along a baseline path.
• Concept Importance Quantification - Quantifies how user-defined high-level concepts influence model predictions using directional derivatives and activation vectors.
• Feature Attribution Methods - Computes how each input feature contributed to a prediction using gradient-based, perturbation-based, or surrogate-model methods.
• Feature Importance Attribution - Measures how each input feature contributes to a model's prediction using attribution algorithms like Integrated Gradients.
• Game Theoretic Explainability - Assigns feature importance using Shapley values from cooperative game theory, approximated via sampling or integration.
• Attribution Gradients - Computes feature importance by backpropagating model outputs through the computational graph to input gradients.
• Integrated Gradients Attribution Methods - Implements integrated gradients to attribute predictions to input features along a baseline path.
• Concept Influence Measurement - Measures how much a user-defined high-level concept, like stripes, influences a model's predictions for a specific class.
• Local Surrogate Models - Trains simpler interpretable models on perturbed samples around an input to explain local model behavior.
• Custom Attribution Algorithm Frameworks - Provides a generic framework for implementing, benchmarking, and sharing new attribution methods.
• PyTorch Libraries - An open-source library for attributing model predictions to input features, neurons, and layers using gradient-based and perturbation-based methods.
• Cross-Modal Explanation - Applies interpretability methods to vision, text, and other model types without changing the underlying architecture.
• Perturbation-Based Sampling - Measures prediction changes by systematically masking or altering input features to estimate their contribution.
• SHAP Value Computations - Approximates SHAP values by adding noise, sampling points, and averaging gradients times input-baseline differences.
• Shapley Value Calculators - Ships Shapley value calculators for game-theoretic feature attribution in model predictions.
• Attribution Algorithm Benchmarks - Compares and evaluates custom interpretability methods against built-in attribution algorithms for reliability and performance.
• Training Data Influence - Identifies which training examples most influenced a specific prediction to detect mislabeled data or debug model behavior.
• Saliency Attribution Methods - Uses raw input gradients as attribution maps to highlight influential features.
• Attribution Averaging - Reduces attribution variance by repeating methods with added Gaussian noise and combining results.
• Layer Contribution Attribution Methods - Measures how individual neurons and layers contribute to a model's final prediction.
• Neuron Importance Attribution - Measures the importance of individual neurons and layers to model predictions.
• Sparse and Dense Feature Importance - Measures importance of both categorical and numerical features in recommendation model predictions.
• Input Times Gradient Attribution Methods - Produces feature attributions by multiplying inputs by their gradients with respect to the output.
• Concept Representation Detection - Determines how well a user-defined concept is represented in a model's internal representations.
• Local Surrogate Explainers - Implements LIME-style local surrogate models to explain individual predictions.
• Attribution Quality Evaluators - Ships tools to assess attribution reliability through sensitivity and consistency tests.
• Training Data Influence Analysis - Identifies which training examples most influenced a given prediction to detect mislabeled data.
• Attribution-Based Troubleshooting - Reveals feature contributions to unexpected outputs for diagnosing model behavior.
• Concept Attribution Methods - Quantifies how user-defined high-level concepts influence model predictions.
• Local Surrogate Explainers - Provides a LIME implementation for local surrogate explanations of individual predictions.
• Layer Contribution Attribution Methods - Measures how individual neurons and layers contribute to a model's final prediction.
• Neuron Importance Attribution - Measures the importance of individual neurons and layers to model predictions.
• Adversarial Input Perturbation - Creates minimal input changes that alter a model's prediction, producing counterfactual explanations.
• Token Attribution - Attributes output tokens of large language models back to the input tokens that influenced their generation.
• Embedding Layer Contribution Analyzers - Measures how embedding layers and attention layers contribute to model predictions.
• Interpretability Algorithm Benchmarks - Compares custom interpretability methods against built-in algorithms to evaluate their performance.
• Token Attribution - Identifies which text tokens most influenced a model's prediction through attribution methods.
• Multimodal Explanations - Attributes predictions separately across different input modalities like text and image in multimodal models.
• Attribution Heatmaps - Generates heatmap overlays highlighting important image regions for model predictions.
• Adversarial Robustness Testing - Generates minimal input perturbations to test model stability and identify vulnerabilities to adversarial attacks.
• Algorithm Customization Frameworks - Provides generic base classes and hooks to implement, benchmark, and share new attribution methods.
• DeepLift Attribution Methods - Implements DeepLift to attribute prediction differences to input differences from a reference.
• DeepLiftSHAP Attribution Methods - Approximates SHAP values by averaging DeepLift attributions across multiple baseline-input pairs.
• Guided BackPropagation Attribution Methods - Produces sharper attribution maps by back-propagating only positive gradients through ReLU activations.
• Guided GradCam Attribution Methods - Combines Guided BackPropagation and GradCam for attribution maps on convolutional layers.
• Multimodal Attribution Methods - Computes separate attribution scores for text and image modalities in multimodal models.
• Attribution Heatmaps - Renders feature importance scores as heatmaps or overlays on the original input for human inspection.
• Core Frameworks - Model interpretability and understanding for neural networks.
• Explainability and Fairness - Interpretability library for PyTorch models.
• Explainable AI Libraries - Model interpretability library for PyTorch models.
• Model Interpretability - Model interpretability for PyTorch.
• 开发者工具 - Model interpretability and understanding.
• Explainability - Listed in the “Explainability” section of the The Incredible Pytorch awesome list.

Star 历史