search

aim-uofa / AdelaiDet

AdelaiDet is an open source toolbox for multiple instance-level detection and recognition tasks.
https://git.io/AdelaiDet
Other
3.39k stars 652 forks source link

The performance of CondInst #217

Open yongliu20 opened 4 years ago

yongliu20 commented 4 years ago

Hi: I follow the instructor to train CondInst with MS_R_50_1x. But the performance of it is not good. Just as the following picture. Why? And should I adjust the code or some paramaters? Thanks a lot! image

tianzhi0549 commented 4 years ago

@workforai The default parameters should be fine. Could you please post your training log here?

yongliu20 commented 4 years ago

I'm not sure if it is this file. log.txt And does the number of gpu affect the final performance? Because I use 4 gpu to train it.

tianzhi0549 commented 4 years ago

I found you reduce the batch size. Batch size is very important to the performance. Please keep the original batch size, and use 8 GPUs to train if you can.

yongliu20 commented 4 years ago

Ok, Thanks for your help!

Syzygianinfern0 commented 4 years ago

I found you reduce the batch size. Batch size is very important to the performance. Please keep the original batch size, and use 8 GPUs to train if you can.

@tianzhi0549 I have access to a maximum of 4 GPUs for training. Is there any way I achieve good scores?

To be on the same terms, by batch size, you mean the SOLVER.IMS_PER_BATCH parameter right?

tianzhi0549 commented 4 years ago

@Syzygianinfern0 You can still train with batch size 16. If you encounter out-of-memory, please try to reduce https://github.com/aim-uofa/AdelaiDet/blob/5cd38f24c2927975df103ee0dede3081a3d4f239/configs/CondInst/Base-CondInst.yaml#L16

Syzygianinfern0 commented 4 years ago

@Syzygianinfern0 You can still train with batch size 16. If you encounter out-of-memory, please try to reduce

https://github.com/aim-uofa/AdelaiDet/blob/5cd38f24c2927975df103ee0dede3081a3d4f239/configs/CondInst/Base-CondInst.yaml#L16

Thank you for your response. I will look into it. 😄