search

GATECH-EIC / ShiftAddNet

[NeurIPS 2020] ShiftAddNet: A Hardware-Inspired Deep Network
MIT License
69 stars 18 forks source link

Results on ImageNet dataset #11

Open HaoKun-Li opened 3 years ago

HaoKun-Li commented 3 years ago

Hi，is there any experimental results or code on ImageNet dataset available? Thanks!

ranery commented 1 year ago

Hi,

Thanks for asking. For this project, the Cuda implementation of the adder layer is still slow, so training on ImageNet takes several months. We did not include the final results in our paper as we only half-trained it.

That said, we manage to work on ImageNet in our following-up work. [ICML 2022] ShiftAddNAS: Hardware-Inspired Search for More Accurate and Efficient Neural Networks

In that work, we search for networks consisting of various multiplication-based or multiplication-free blocks. The results finally can go up to 83%.

We also have ongoing work that applied shift and add ideas to the Vision Transformers on ImageNet, please stay tuned.

ranery commented 1 year ago

Shift&Add on ViTs and ImageNet: https://arxiv.org/abs/2306.06446