Training Log? about queryinst HOT 7 CLOSED

@JialianW For learning rate adjustment, please follow the Linear Scaling Rule: smaller batch size with smaller learning rate, vice versa.

Hi @JialianW.

To be precise, if you use 0.5x batchsize, then the linear scaling rule (https://arxiv.org/abs/1706.02677) recommends using 0.5x lr optimized via SGD. But QueryInst & Sparse R-CNN as well as DETR families use Adam families to optimize, therefore the linear scaling rule might be too harsh. E.g., DETR does not scale their lr even with batchsize = 64. You can refer to some issues in the DETR repo (facebookresearch/detr#48 (comment)) for details.

For QueryInst, we do use the linear scaling rule when scaling up the batchsize. But I am not sure if there exists a better principle for QueryInst & Adam families.

from queryinst.

Hi @JialianW, thanks for your interest in our work!

As for the logs, we will release them in the near future and we will let you know in the first place, please stay tuned :)

As for the warnings, it is fine and just let the bullets fly.

from queryinst.

Thanks. For the learning rate, do we need to change two times smaller when using 4 GPUs? The default learning rate is for 8 GPUs?

from queryinst.

@JialianW For learning rate adjustment, please follow the Linear Scaling Rule: smaller batch size with smaller learning rate, vice versa.

from queryinst.

Thank you! It helps a lot!

from queryinst.

This issue won't be closed until the training logs are released.

from queryinst.

QueryInst is now officially included by mmdetection library, with new checkpoints, corresponding logs, and augmented training settings.

https://github.com/open-mmlab/mmdetection/blob/master/configs/queryinst/README.md

from queryinst.