Details
Presenter(s)
![Hongrui Song Headshot](https://confcats-catavault.s3.amazonaws.com/CATAVault/ieeecass/master/files/styles/cc_user_photo/s3/user-pictures/15501.jpg?h=0ab00ada&itok=5N8eXkJd)
Display Name
Hongrui Song
- Affiliation
-
AffiliationNanjing University
- Country
Abstract
We develop a unified compressed version of Vision Transformer (UCViT), whose main focus is on compressing the computation-intensive ViT model by a unified compression method, which converts the dense matrix multiplication to hardware-friendly operations with low bit-width dominated by shift and addition calculations.Meanwhile, to compensate for the accuracy degradation, we additionally introduce a small matrix with relative high-precision associated with the mechanism of multi-head attention in Transformer-based model.