Regularization

[RainerHeintzmann]

modified TV to be able to work without needing to specify the dimension number and whether the array is a CuArray or not.

[roflmaostc]

Thanks!

However, the CPU regularizer are more powerful than the CUDA regularizer.
Hence, sharing the same interface is misleading.

In the long term, I was hoping that Tullio.jl catches up and might perform equally well with CUDA as the Julia default functions.

[RainerHeintzmann]

Are you referring to the sum_dims and the step argument? I guess they can also be realized in the sub-index (cuda) version. The same should hold for the forward but this is anyways not an interesting choice apart from trying something out.
I think the damage that is done by throwing a hard to understand error when trying a regularizer straight away with CuArrays is far bigger.

[roflmaostc]

I agree, but it is clearly mentioned in the documentation and examples.