search

mrnabati / CenterFusion

CenterFusion: Center-based Radar and Camera Fusion for 3D Object Detection
MIT License
542 stars 140 forks source link

Changing Batch Size for Training #32

Open RameshSena opened 3 years ago

RameshSena commented 3 years ago

When I try to change the --batch_size in the file train.sh to something higher than 16: that is like 32 or 64, I get an insufficient memory error. Which forces me to keep the batch size to 16. I use 2 Nvidia RTX 2080 Ti GPUs which have 11GB memory. (Total 22GB) I really need to increase the batch size as I experience fluctuations in training accuracy. Can you please suggest me a way to increase the batch size without buying GPUs with more memory. And also is there a relationship between --num_workers and --batch_size parameters?

fabrizioschiano commented 3 years ago

Hi @RameshSena , I don't think that there is a better solution than having more memory to solve your problem.

Instead, did you find a solution to the second question you had about the relationship between the name_workers and batch_size parameters? Thanks!