IFORMER: INTEGRATING CONVNET AND TRANS- FORMER FOR MOBILE APPLICATION

Official PyTorch implementation of iFormer, published on ICLR 2025.

Main Results on ImageNet with Pretrained Models

Model	Params(M)	GMACs	Latency(ms)	Top-1(%)	Ckpt.	Core ML	Log
iFormer-T	2.9	0.53	0.60	74.1	300e(iFormer_t.pth)	300e(iFormer_t_224.mlpackage.zip)	300e(iFormer_t.out)
iFormer-S	6.5	1.09	0.85	78.8	300e(iFormer_s.pth)	300e(iFormer_s_224.mlpackage.zip)	300e(iFormer_s.out)
iFormer-M	8.9	1.64	1.10	80.4/81.1	300e(iFormer_m.pth)/300e distill(iFormer_m_distill.pth)	300e(iFormer_m_224.mlpackage.zip)/300e distill(iFormer_m_224_distill.mlpackage.zip)	300e(iFormer_m.out) / 300e distill(iFormer_m_distill.out)
iFormer-L	14.7	2.63	1.60	81.9/82.7	300e(iFormer_l.pth) /300e distill(iFormer_l_distill.pth)	300e(iFormer_l_224.mlpackage.zip)/300e distill(iFormer_l_224_distill.mlpackage.zip)	300e(iFormer_l.out) /300e distill(iFormer_l_distill.out)
iFormer-L2	24.5	4.50	2.30	83.9	300e distill(iFormer_l2_distill.pth)	300e distill(iFormer_l2_224_distill.mlpackage.zip)	300e distill(iFormer_l2_distill.out)
iFormer-H	99.0	15.5	-	84.8	300e(iFormer_h.pth)	300e(iFormer_h_224.mlpackage.zip)	300e(iFormer_h.out)

• iFormer-L2 is trained with distillation for 450 epochs.

Uses

Summary

You can also see iFormer github for usage.

Citation

BibTeX:

@article{zheng2025iformer,
  title={iFormer: Integrating ConvNet and Transformer for Mobile Application},
  author={Zheng, Chuanyang},
  journal={arXiv preprint arXiv:2501.15369},
  year={2025}
}

Model Card Authors

Chuanyang Zheng

Model Card Contact

chuanyang [email protected]