kornia/tiny_vit

Pretrained weights for TinyViT, used as the encoder backbone in kornia.models.SegmentAnything (MobileSAM) and available via kornia.models.TinyViT.

TinyViT is a small Vision Transformer trained with knowledge distillation from large teacher models on ImageNet-22K. ECCV 2022.

Original repo: microsoft/Cream/TinyViT

Weights

File	Params	Pre-training	Fine-tuning
tiny_vit_5m_22k_distill.pth	5M	ImageNet-22K	—
tiny_vit_5m_22kto1k_distill.pth	5M	ImageNet-22K	ImageNet-1K 224
tiny_vit_11m_22k_distill.pth	11M	ImageNet-22K	—
tiny_vit_11m_22kto1k_distill.pth	11M	ImageNet-22K	ImageNet-1K 224
tiny_vit_21m_22k_distill.pth	21M	ImageNet-22K	—
tiny_vit_21m_22kto1k_distill.pth	21M	ImageNet-22K	ImageNet-1K 224
tiny_vit_21m_22kto1k_384_distill.pth	21M	ImageNet-22K	ImageNet-1K 384
tiny_vit_21m_22kto1k_512_distill.pth	21M	ImageNet-22K	ImageNet-1K 512

Citation

@inproceedings{wu2022tinyvit,
    title     = {{TinyViT}: Fast Pretraining Distillation for Small Vision Transformers},
    author    = {Wu, Kan and Zhang, Jinnian and Peng, Houwen and Liu, Mengchen
                 and Xiao, Bin and Fu, Jianlong and Yuan, Lu},
    booktitle = {ECCV},
    year      = {2022}
}