epapoutsellis
commented
9 months ago There are two alternatives for this problem. One is for free and is from the TIGRE toolbox which we support.
One can create easily a wrapper for im_3d_denoise. Similar to FGP_TV the dual ROF is solved but not with an accelerated gradient descent.
The algorithm is basically a projected gradient descent described in here
The step-size is not fixed in every iteration but is changing. https://github.com/CERN/TIGRE/blob/dbcd848671c96a42b7c9669e75c8e7e19e55ce6d/Common/CUDA/tvdenoising.cu#L445-L446
I believe the implementation is based on ZhuChan maybe @AnderBiguri can confirm.
It works with multiple gpu and is useful for large 3D datasets.
The other option is to use cucim and dask-cuda.
In both cases, one needs to implement a __call__ method that computes the Total variation in every iterate, $$||\nabla x{k}||{2,1}$$ which can be done by the GradientOperator and MixedL21Norm.
AnderBiguri
paskino
In the FGP_TV class for the CCPi-Regularisation Toolkit the step-size (Lipschitz constant) is wrong.
At the moment is
1/26but the correct value for 3D arrays with unitary grid (voxel-size=1.0) is1./12. Because in this case the $$||\nabla|| = \sqrt{12}$$This can cause problems when Total Variation prior is used with PDHG/SPDHG/FISTA. @Letizia97 @paskino
When tolerance is 0.0 (default), for every outer iteration we run 100 iterations (default) for the proximal of TV (which is the FISTA algorithm applied on the dual-ROF problem). Since every time a smaller step-size (
1/26) is used, it is possible that we have not reached to the desired solution for every iteration, and in some cases we may observe divergence, see the figure below by BeckTeboulle. Higher step size (1/12) will give better results.Warm-starting can help but still the subproblem is not solved optimally, i.e., with the right step size.
Note: I have tried many times to make this simple change and build it locally. But failed every time. Actually, I managed to make it work with cpu but building it with gpu it was a complete failure. For the datasets that I am working, it was usually 15-20sec per iteration and after building it with this change it was 100sec per iteration.