uncbiag / SimpleClick

SimpleClick: Interactive Image Segmentation with Simple Vision Transformers (ICCV 2023)
MIT License
209 stars 32 forks source link

GPU setting #13

Closed che1007 closed 1 year ago

che1007 commented 1 year ago

Can I train your model on single GPU? If the answer is 'Yes', what hyperparameter that I need to revise

qinliuliuqin commented 1 year ago

You just need to set "--ngpus=1" for training models on the branch v1.0. I am now refactoring the code on the main branch, so don't train your models on the main branch for now.

che1007 commented 1 year ago

Thanks for your help. The last two day I had tried to train base_448_sbd several times, and I found that the training loss went to rounding 0.31 after two or three epochs. But the training loss didn't keep going down after two or three epochs. Expect the refactoring version from you.

qinliuliuqin commented 1 year ago

That's normal because we use normalized focal loss. Yes, I will refactor v1.0 soon because it's really messy.