# Captum An open source PyTorch library for developers and researchers interpreting model outputs. - Findings24: - Site: - Preços: Freemium - Categorias: Ferramentas Dev IA ## O que é Captum? Captum is an open source, extensible library for model interpretability built on PyTorch. It helps developers and researchers examine which input features contribute to a model’s output and investigate how internal layers and neurons affect predictions. The library supports most types of PyTorch models with minimal modification to the original network. Its algorithms and tutorials cover vision, text, tabular data, and multimodal models, providing a practical foundation for troubleshooting predictions and researching interpretability methods. ## O que Captum pode fazer? ### Input attribution algorithms Apply Integrated Gradients, DeepLIFT, GradientShap, Occlusion, LRP, or LIME to examine how input features contribute to a selected model output. ### Layer and neuron analysis Use conductance and internal attribution methods to investigate layer and neuron importance, including the input features associated with a particular neuron. ### LLM and multimodal attribution Interpret text generation with LLM attribution functionality. For multimodal LLMs, combine ImageMaskInput, image segmentation, and FeatureAblation to examine image regions contributing to generated text. ### Concept and robustness analysis Explore concepts with Testing with Concept Activation Vectors (TCAV). Tutorials also demonstrate FGSM and PGD attacks and robustness metrics alongside attribution algorithms. ### Extensible research workflows Implement new interpretability algorithms and benchmark them against methods available in the library. Tutorials demonstrate distributed attribution using torch.distributed and DataParallel. ## Casos de uso ### Troubleshoot classification predictions Inspect influential words in sentiment predictions or image regions in vision models to investigate unexpected outputs. ### Analyze tabular regression models Evaluate feature importance and internal attribution in a house price regression network to understand how the network uses its inputs. ### Explain recommendation behavior Examine sparse and dense feature contributions in Deep Learning Recommender Models, including feature interactions involved in clicked and non-clicked ad predictions. ### Investigate question answering models Study BERT embeddings, attention layers, and token interactions to understand a particular answer prediction. ## Perguntas frequentes ### How do I install Captum? Install with \`conda install captum -c pytorch\`, the website’s recommended method, or use \`pip install captum\`. ### What knowledge do I need to get started? The tutorials assume familiarity with PyTorch and its basic features. The introductory Captum tutorial trains a neural network on Titanic survival data, then examines input, layer, and neuron importance. ### What does an Integrated Gradients workflow return? The homepage example passes input and baseline tensors to an IntegratedGradients instance, selects a target output, and returns attribution scores. With return\_convergence\_delta=True, it also returns a convergence delta. ## Informações rápidas - **Product type**: Open source model interpretability library - **Framework**: PyTorch - **Primary users**: Model developers, interpretability researchers, and application engineers - **Website**: https://captum.ai