BN Stats Adapt in cityscape -> ACDC
Closed this issue · 5 comments
Dear author:
Thanks for your code of semantic segmentation adaptation from cityscape -> ACDC. However, I can not find the code of BN Stats Adapt there. Could you tell me where is this part of code? Thanks
Hi @HaihangWu,
I am also trying to use this codebase, but somehow, I cannot understand where to put the models and the data. Could you tell me how you ran "run_base.sh", "run_cotta.sh", "run_tent.sh" ? Thanks.
Hi:
To know this, you may need to read the code. For data, you need to specify the data root in local_config/base/acdc_1024X1024_repeat. For model, I author has put it in the directory and you do not need to change the path
Dear author: Thanks for your code of semantic segmentation adaptation from cityscape -> ACDC. However, I can not find the code of BN Stats Adapt there. Could you tell me where is this part of code? Thanks
Hi, BN Stats Adapt is a simple baseline without any training.
One way to implement it is to modify the source-only baseline model. You can change model.eval() to model.train() in the baseline inference code. You also don't need to add any loss or back propagation. This will let the batch norm layer in the decoder use the target batch statistics instead of the averaged source statistics. You can check this page if you are not sure.
Thanks for your reply. For cityscape to ACDC experiments, all the results on table 5 in the paper "Continual Test-Time Domain Adaptation" are reproducible.
Thanks, and feel free to shoot me an email if you have further questions.