Metadata-Version: 2.4
Name: fast-vit-rollout
Version: 0.1.0
Summary: A robust, plug-and-play Attention Rollout explainer for Vision Transformers (timm & Hugging Face).
Author-email: Souradeep Dutta <souradeep2233@gmail.com>
Project-URL: Homepage, https://github.com/yourusername/fast-vit-rollout
Classifier: Programming Language :: Python :: 3
Classifier: License :: OSI Approved :: MIT License
Classifier: Operating System :: OS Independent
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
Requires-Python: >=3.8
Description-Content-Type: text/markdown
Requires-Dist: torch>=2.0.0
Requires-Dist: numpy
Requires-Dist: opencv-python

# Fast ViT Rollout 🚀

A robust, plug-and-play Attention Rollout extractor for Vision Transformers. Works seamlessly out of the box with `timm` and Hugging Face backbones.

### Why use this?
Most existing Attention Rollout scripts break on modern architectures. This package natively handles:
* **Flash Attention (SDPA):** Bypasses PyTorch 2.0+ fused attention issues.
* **Register Tokens:** Flawlessly parses DINOv2 and DeiT models without shape mismatch errors.
* **Auto-Detection:** Automatically hooks the correct layers without manual indexing.
* **Native Heatmaps:** Generates OpenCV-based overlays with built-in intensity colorbars.

### Installation
```bash
pip install fast-vit-rollout
