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

Reproducing throughput results #31

Closed ivanke1 closed 1 year ago

ivanke1 commented 1 year ago

Hi, I am interested in your work and am trying to replicate some results, but I can't seem to observe the same throughput increase. It seems to take a similar amount of time for the original DeiT-S and DynamicViT version to evaluate on the Imagenet validation set for me. I was wondering if you had any advice for my setup, I am currently using a Google Colab notebook with a single GPU.

Thanks

raoyongming commented 1 year ago

Hi @ivanke1, thanks for your interest in our work. The throughput results reported in our paper are measured on a single 3090 GPU. We also see similar throughput improvements on other types of GPUs like 1080ti, 2080ti, and A100. I would recommend measuring throughput in a clean and controllable environment without executing other heavy tasks.