TV with learneable D

Question

TV with learneable D

Closed this issue 2 years ago · 3 comments

swami1995 commented 2 years ago

Thanks for the amazing work!

I was wondering if you guys have also tried to write an implementation where D can be fed into the TV layer as a parameter and get gradients for the corresponding D matrix? (Specifically D could be parameterized using a 1xN vector or 2xNxC (N=kernel length (say 2 or 3), C=number of channels) which is what would be fed in and used for gradient computation.

And if I had to try to implement that on top of your current implementation, could you point me towards how I could go about it? Thanks!

Answer 1 · 2022-08-15T15:42:30.000Z

Hi @swami1995,

We have not tried learning the difference matrix D. The TV solver takes advantage of the specific structure of D for speedup.
It would no longer be a TV problem if D is being learned, hence, one would have to utilized a different solver.

Best,
Raymond

Answer 2 · 2022-08-29T17:36:17.000Z

This issue is stale because it has been open for 14 days with no activity.

Answer 3 · 2022-09-05T17:37:13.000Z

This issue was closed because it has been inactive for 7 days since being marked as stale.