Request for Stage-1 Frequency Prompt Training Code and Configuration

[zzzxj186-sys]

Hi authors,

Thank you for releasing the official implementation of FreeKD.

I am trying to reproduce Stage 1 of FreeKD and adapt the Frequency Prompt
to a new object detection dataset. However, the current repository appears
to provide the pretrained prompt checkpoint and the Stage-2 distillation
configuration, but I could not find the code/configuration used to train
the Frequency Prompt in Stage 1.

Could you please release the Stage-1 training code/config, or clarify the
following implementation details?

1. Where exactly is each MaskModule inserted in the detector?
   Is it applied to the five FPN outputs before the RetinaNet detection head?

2. During Stage 1, which parameters are trainable?

   • Frequency Prompt only;
   • Frequency Prompt and detection head;
   • or the complete teacher network?

3. What optimizer, learning rate, batch size, training epochs and LR schedule
   were used for Stage 1?

4. How is the Prompt-Dissimilarity Loss combined with the teacher detection
   loss in the released implementation?

5. The paper states T=2 imitation principles for each frequency band, while
   the released RetinaNet distillation configuration uses num_tokens=6.
   Which setting was used to train the released
   FreeKD_retinanet_r101_prompt.pth checkpoint?

6. How was the final prompt checkpoint selected, and what state_dict key
   format is expected by FreeKDLoss?

We would like to first reproduce the official COCO prompt training results
and then train a dataset-specific prompt for AI-TOD-v2.

Releasing the Stage-1 config or training wrapper would be extremely helpful
for reproducing and extending FreeKD.

Thank you very much!