search

raoyongming / DynamicViT

[NeurIPS 2021] [T-PAMI] DynamicViT: Efficient Vision Transformers with Dynamic Token Sparsification
https://dynamicvit.ivg-research.xyz/
MIT License
551 stars 69 forks source link

Dynamic token on training speedup #38

Closed ZK-Zhou closed 10 months ago

ZK-Zhou commented 10 months ago

Hi Yongming,

Thank you for your wonderful work, I would like to ask if dynamic vit will achieve speedup during training (vs vanilla ViT).

raoyongming commented 10 months ago

Hi @ZK-Zhou, thanks for your interest in our work. DynamicViT is designed for model acceleration during inference. Since we drop tokens based on their importance predicted by a well pre-trained model, our method cannot be directly used to speed up training.