search

jeanfeydy / geomloss

Geometric loss functions between point clouds, images and volumes
MIT License
600 stars 60 forks source link

Make reach depend on diameter #23

Closed AdrienCorenflos closed 4 years ago

AdrienCorenflos commented 4 years ago

Hi,

Wouldn't it be better to have the reach param depend on the diameter of the data? It's probably easier to think of saturation of the distance in terms of %diameter than it is in the direct unit of the distance. What do you think?

jeanfeydy commented 4 years ago

HI @AdrienCorenflos ,

Sorry for the late reply, and thanks for your remark! It would make sense, indeed, and this is how I always discuss entropic regularization in papers and talks. However, as far as the library is concerned, we should remember that the diameter of the (source + target) configuration typically changes between calls. In the context of optimization for geometry processing and machine learning, the current interface is thus much safer: it specifies a well-defined loss function that should (hopefully!) decrease across iterations. Changing it would leave the door open to confusions, for little benefit I think.

Best regards, Jean