mabingqi comments

Results 6 comments of


                                            mabingqi

KeyError: 'TSDSharedFCBBoxHead is not in the head registry'

I run the code, without loading checkpoint under ./TSD/ and everything is ok. Re-installing tsd maybe helpful. And the potential problem maybe that you have more than one mmdet installed...

Missing 'model.CLASSES' field in the pre-trained model for the open images dataset

I update the data download link in readme and you can find a cls-label-description.csv in challenge2019.

Missing 'model.CLASSES' field in the pre-trained model for the open images dataset

https://drive.google.com/file/d/1-CnE3akagnoyDarrDuKTl7OF9vwF_rA3/view?usp=sharing You can download with the link.

bp of shared parameters and experts

Thank you for your answer. I notice that _tutel_expert flag is used to split the parameters. But it seems that gradient of experts with _tutel_expert will also be allreduced by...

bp of shared parameters and experts

I notice the code in swin-transformer repo(https://github.com/microsoft/Swin-Transformer/blob/main/main_moe.py), which uses pytorch optimizer and ddp to train these moe models. Maybe there is something wrong. Thanks a lot.

bp of shared parameters and experts

Does it work by setting skip_allreduce as true in the scan function?