stochman
stochman copied to clipboard
MachineLearningLifeScience
Algorithms for computations on random manifolds made easier
Hi, I was going through the source code of stochman with a colleague and we found what could likely be a bug in `stochman.curves.CubicSpline`. In essence, the code for computing...
**Is your feature related to a problem?** When plotting geodesics, calling `connecting_geodesic`, the function assumes that the graph is connected. When not connected, `networks.shortest_path` throws after some time an error...
StochMan is currently not well-enough documented. We should provide a series of example-tutorials to show how easy it is to get going.
Manifold.curve_energy should have a dim argument in the return statement, ie. line 47 ` return energy.sum()` should be `return energy.sum(dim= ... )`.
Pytorch does practically speaking not support computing Jacobians, but does have reasonable support for Jacobian-vector products. We can leverage this to e.g. provide a default implementation of Manifold.inner(). To do...
We should have a class StatisticalManifold (or something like that) where one have to provide an embed function that returns a torch distribution that has a KL. See Pulling Back...
See e.g. the paper from Fletcher's group for a simple algorithm
StochMan should have a basic set of classes for working with (at least isotropic) Brownian motion and the LAND.
The class LocalVarMetric is to specialized to be a library class. Rather this should be an example.
StochMan should support an N-dimensional manifold discretization in order to apply graph-based solvers for computing geodesics.
Metadata
Owner
Metadata
Algorithms for computations on random manifolds made easier