Optimisation of MI loss implementation

[qiuhuaqi]

It has been noted by some collaborators that the MI implementation (here) in this library is slightly slower and less easy to tune than De Vos's implementation in TorchIR (https://github.com/BDdeVos/TorchIR/blob/main/torchir/metrics.py#L74). The TorchIR implementation also exposes the Parzen window width as a parameter for users to choose, while we automatically derive this from the choice of number of bins by setting the Gaussian kernel's FWHM as bin width. We could experiment with this as well.

This issue is to track experiments and optimisation.

[aschuh-hf]

Thanks for relaying the feedback and helping to improve this! I've added TorchIO to the list of related projects.