Dear author, I found that you used the Jaccard coefficient to train the mask in your paper, but I did not find it in the code.
https://github.com/Gumpest/FreeKD/blob/main/mmrazor/models/losses/freekd.py#L11
Do I need to implement it myself? And how to visualize the frequency domain mask well in the spatial domain?
The visualization effect of the mask obtained at different scales after upsampling is not ideal.
And about the code
https://github.com/Gumpest/FreeKD/blob/main/mmrazor/models/losses/freekd.py#L66
Here you named the return as cA, (cH, cV, cD). Intuitively, do they correspond to approximate, horizontal, vertical, and diagonal features?
However, from the source code, is the return result a list of features at different scales?
https://github.com/fbcotter/pytorch_wavelets/blob/master/pytorch_wavelets/dwt/transform2d.py#L68
Looking forward to your reply :3
Dear author, I found that you used the Jaccard coefficient to train the mask in your paper, but I did not find it in the code.
https://github.com/Gumpest/FreeKD/blob/main/mmrazor/models/losses/freekd.py#L11
Do I need to implement it myself? And how to visualize the frequency domain mask well in the spatial domain?
The visualization effect of the mask obtained at different scales after upsampling is not ideal.
And about the code
https://github.com/Gumpest/FreeKD/blob/main/mmrazor/models/losses/freekd.py#L66
Here you named the return as cA, (cH, cV, cD). Intuitively, do they correspond to approximate, horizontal, vertical, and diagonal features?
However, from the source code, is the return result a list of features at different scales?
https://github.com/fbcotter/pytorch_wavelets/blob/master/pytorch_wavelets/dwt/transform2d.py#L68
Looking forward to your reply :3