How can I reproduce the Cityscapes segmentation results about dist_kd HOT 5 CLOSED

Hi @luuil ,

Sorry for the delay of releasing the segmentation code. Our code is mostly based on the code in the CIRKD repo.

We will release our segmentation code this week (within 1~2 days if everything goes well).

Thanks,
Tao

from dist_kd.

Hi @luuil ,

Sorry for the delay of releasing the segmentation code. Our code is mostly based on the code in the CIRKD repo.

We will release our segmentation code this week (within 1~2 days if everything goes well).

Thanks, Tao

Hi @luuil ,

We have released the segmentation code. You can check the segmentation folder in this repo for instructions.

If you have any question about the code or our paper, please feel free to contact us.

Best,
Tao

from dist_kd.

Thank you for releasing so quickly.

By the way, I have another question.

As you mentioned in chapter Appendix-A.5-Limitations of your paper

it would be less-effective on binary classification task

And I am working on a binary mask segmentation task(that is, for a input image, the goal of the task is to segment the fore&back ground) 😢.

What do you suggest for this?

from dist_kd.

The binary classification tasks are often easily to be converged, and thus the prediction scores are very close to 0 or 1 (like one-hot labels). As a result, the teacher's logits have less relation information to be transferred than the challenging multi-task classification.

However, you can still try DIST and see if it is better than the conventional KD loss :)

from dist_kd.

Thanks for the suggestion, I'll give it a try.

from dist_kd.