Iwin Transformer (large, ImageNet-22K→1K, 384×384)

Position-embedding-free hierarchical vision transformer. Paper: arXiv:2507.18405. All checkpoints: Iwin Transformer collection.

Usage

from transformers import AutoImageProcessor, AutoModelForImageClassification

ckpt = "cominder/iwin-large-patch4-window12-384-22kto1k"
processor = AutoImageProcessor.from_pretrained(ckpt)
model = AutoModelForImageClassification.from_pretrained(ckpt, trust_remote_code=True)
inputs = processor(images=image, return_tensors="pt")
print(model.config.id2label[model(**inputs).logits.argmax(-1).item()])

trust_remote_code=True is required because Iwin is not in the Transformers core library yet. Training and evaluation code: Cominder/Iwin-Transformer.

Reported ImageNet-1K top-1: 87.4. The snippet resizes directly to 384.

Citation

@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},
}