search

ayushjain1144 / odin

Code for the paper: "ODIN: A Single Model for 2D and 3D Segmentation" (CVPR 2024)
https://odin-seg.github.io
MIT License
113 stars 10 forks source link

Question about MAX_ITER and IMS_PER_BATCH #20

Closed mingqiJ closed 2 months ago

mingqiJ commented 2 months ago

Hi, thanks for your code. I have a question about your training scripts.

I see you set the MAX_ITER and IMS_PER_BATCH in the .sh file. If I decrease the batch size, do I need to increase the iteration? Like you set 4 IMS_PER_BATCH on 2 gpus and 10000 MAX_ITER. If I set 2 IMS_PER_BATCH on 2 gpus, do I need to double the MAX_ITER to 20000?

ayushjain1144 commented 2 months ago

Hi, I typically set them to a "high enough" value so that my training does not get killed pre-maturely; and look at the logs to early-stop when the model converges.

With half the effective batch size, it should take more iterations but most likely less than twice. In any case, maybe set it to a high enough value, and see when it converges.

mingqiJ commented 2 months ago

Hi, thanks for your explanation. Very useful.

ayushjain1144 commented 2 months ago

i am glad 🙂