search

lorenmt / reco

The implementation of "Bootstrapping Semantic Segmentation with Regional Contrast" [ICLR 2022].
https://shikun.io/projects/regional-contrast
Other
162 stars 25 forks source link

The speed of training #11

Closed super233 closed 2 years ago

super233 commented 2 years ago

Hi, thanks for your sharing code, it's awesome work.

I'm a green hand in contrastive learning, I'm running your code with an NVIDIA Tesla T4, but it seems slowly for training and the GPU-Util can not always keep a value that is greater than 50%, is it a normal phenomenon?

It seems that the process of selecting queries and corresponding negative keys is time consuming.

lorenmt commented 2 years ago

Hello,

Yes, since ReCo samples hard pixels and accumulates losses across all semantic classes, this is a common observation. You may speed up training by sampling a lower number of queries.