README.md

---
title: Iwin Transformer Image Classification
emoji: 🪟
colorFrom: blue
colorTo: pink
sdk: gradio
sdk_version: 6.26.0
app_file: app.py
short_description: Iwin Transformer for ImageNet image classification
python_version: "3.12"
startup_duration_timeout: 30m
---

# Iwin Transformer — ImageNet Classification

This Space demonstrates **Iwin Transformer**, a hierarchical vision transformer
that uses interleaved windows and depthwise separable convolution for efficient
global information exchange within a single module.

## Model

- **Paper**: [Iwin Transformer: Hierarchical Vision Transformer using Interleaved Windows](https://arxiv.org/abs/2507.18405)
- **Code**: [GitHub](https://github.com/cominder/Iwin-Transformer)
- **Weights**: [HuggingFace](https://huggingface.co/cominder/Iwin-Transformer)

The model is trained on ImageNet-1k (with 22k pretraining for the base variant)
and classifies images into 1000 ImageNet categories.

## Variants

| Variant | Params | Top-1 Acc | Checkpoint |
|---------|--------|-----------|------------|
| Tiny    | ~28M   | 81.2%     | `iwin_tiny_patch4_window7_224.pth` |
| Small   | ~50M   | 83.5%     | `iwin_small_patch4_window7_224.pth` |
| Base    | ~88M   | 85.2%     | `iwin_base_patch4_window7_224_22kto1k.pth` |

## Citation

```bibtex
@misc{huo2025iwin,
      title={Iwin Transformer: Hierarchical Vision Transformer using Interleaved Windows},
      author={Simin Huo and Ning Li},
      year={2025},
      eprint={2507.18405},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2507.18405},
}
```