Can the network be replicated using the 2023 version of mmdetection? about fgd

[yzd-v]

yes, you can refer cls_kd for classification. This is with mmengine.

[kiven228]

Thank you for your answer. I have read your code in its entirety and have some questions. The distillation network does indeed calculate distillation losses for the global and focused features between the student model and the teacher model, but specifically, how does the student model obtain these features for prediction? For example, in using GCBLOCK to extract global features, how does the student network utilize global features on itself for prediction? I only saw the code with the ability to calculate distillation losses.Looking forward to your reply and thanks again!

[yzd-v]

After distillation, we do not need the gcblock actually. The global features are just used for distillation instead of prediction.

[kiven228]

Does it mean that the student network did not use the extracted global features during the training process? Because I have seen how to extract and distill global or attention features, but I have not found how the extracted features are transmitted to the network for use. Thank you for taking the time to answer.