search

Sense-X / UniFormer

[ICLR2022] official implementation of UniFormer
Apache License 2.0
829 stars 111 forks source link

uniformer for object detection #68

Closed go-ahead-maker closed 2 years ago

go-ahead-maker commented 2 years ago

Hi, nice work!, but I have a basic question about object detection.

I found that in the mmdetection, some of the pre-defined backbones(e.g., resnet) that with BN layer define the norm_eval parameter, and this parameter will be used to decide whether to freeze the BN layer during the training process. I found that Uniformer do use the BN layer in CABLOCK, but it seems that you don't use norm_eval to freeze the BN layer. Does this mean that BN parameters are also trained during training? And I wonder if this setting has any impact on the training process and final results.

Thanks!

go-ahead-maker commented 2 years ago

ohh, I miss the NOTE in the README, but I still wonder how this setting(freeze BN or not) will affect the results since I'm a beginner. Looking forward to your reply!

Andy1621 commented 2 years ago

In my later experiments, freeze BN improve a little. I suggest you use it!

Andy1621 commented 2 years ago

As there is no more activity, I am closing the issue, don't hesitate to reopen it if necessary.