hila-chefer/Transformer-MM-Explainability

Provides attention rollout and perturbation-based attribution methods to generate visual explanations across diverse transformer architectures including vision-language models (LXMERT, VisualBERT, CLIP), object detection (DETR), and pure vision transformers (ViT). The approach computes importance scores by analyzing attention weight propagation through transformer layers and measuring model sensitivity to input perturbations, supporting both unimodal and cross-modal reasoning tasks. Includes ready-to-use Colab notebooks and integrates with MMF, Hugging Face Transformers, and official model implementations for reproducible analysis.

903 stars. No commits in the last 6 months.