WoojuLee24/OA-DG

The questions about DWD datasets

Closed this issue · 4 comments

startracker0 commented

A heartfelt thank you to all contributors who have contributed to advancing domain generalization techniques on the DWD dataset.

I think https://github.com/WoojuLee24/OA-DG/blob/main/configs/OA-DG/dwd/faster_rcnn_r101_dc5_1x_dwd.py is the baseline of dwd, but I'm curious about whether the learning rate is set to 0.01 or 0.001? And also I also find the lr and milestones are differ from the settings for CDSD according to https://github.com/AmingWu/Single-DGOD#training. And I am curious about the code of oamix and oadg in dwd.

Thank you very much for answering my question, and I look forward to your reply.

dazory commented

We followed the setup in "Single-Domain Generalized Object Detection in Urban Scene via Cyclic-Disentangled Self-Distillation" (See [4. Experiments]-[4.1.Experimental Setup]-[Implementation Details]). According to the paper, the author set the learning rate to 0.001 for the experiment.

Also, you can ignore the commented out 0.01 in the configuration file. This is just a comment.

startracker0 commented

Thanks a lot. I also wonder how much gpus do you use? 2 gpus?

WoojuLee24 commented

We used 1 GPU to train the model.

startracker0 commented

Thank you for your response. I have successfully reproduced your baseline of 48.1 with 4 3090s