search

layumi / Seg-Uncertainty

IJCAI2020 & IJCV2021 :city_sunrise: Unsupervised Scene Adaptation with Memory Regularization in vivo
https://arxiv.org/abs/1912.11164
MIT License
386 stars 51 forks source link

batchsize #15

Open cchenghao opened 3 years ago

cchenghao commented 3 years ago

How much performance will decrease if batchsize=1 is used in the first stage? thanks.

layumi commented 3 years ago

Hi @cchenghao It will make the result unstable. The performance will be around 44~45% from my experience.

cchenghao commented 3 years ago

Hi @cchenghao It will make the result unstable. The performance will be around 44~45% from my experience.

Is this difference too big? But I optimized your code a bit. There are many operations that can be distributed to reduce memory. Now I can train the model with batchsize=2 under 12GB. And I want to remind that: I am asking about the first stage, not the second stage. Don't other methods use 1280×720 for training under batchsize=1? Thanks for your reply.