Poor performance with registered Coco annotations

[apexx77]

Need clarification whether Adaptive Teacher is unsupervised or not.

https://github.com/facebookresearch/adaptive_teacher/blob/cba3c59cadfc9f1a3a676a82bf63d76579ab552b/adapteacher/engine/trainer.py#L646-L647 Here are you using the labels of Target dataset? @yujheli

[yujheli]

@apexx77 This line of code is for keeping the weights as default (1) for supervised loss in the source domain. It would be great to have you read through the codes starting from till the end of the function: https://github.com/facebookresearch/adaptive_teacher/blob/main/adapteacher/engine/trainer.py#L495

Then you will better understand how we process the labels for the target (removing labels in target and replacing them with pseudo labels) and train the model with source labels. Thanks!

[apexx77]

Scenario -1

• Created the same dataset directory as mentioned in the readme for CS and Foggy CS
• Started training Adaptive Teacher

Scenario-2

• Created Coco annotations for CS and Foggy CS.
• Registered the datasets in adaptive_teacher/adapteacher/data/datasets/builtin.py
• Started training Adaptive Teacher

  Expected behaviour

• Getting almost same results in both the scenarios

  Actual behaviour

• Getting the expected results in Scenario-1 but getting very low results in Scenario-2
• In the initial supervised burn up steps, Scenario-2 has very low AP as compared to Scenario-1

@yujheli

[yujheli]

Could you check if you COCO annotations are well-created? Also check if you use the correct register function for detectron2 such as setup correct meta data or so.

[apexx77]

Yes, I did bbox visualisation and trained supervised models in detectron2 as well. But getting low AP after supervised iterations in Adaptive Teacher

@yujheli