jaivardhankapoor/geoopt

Riemannian Adaptive Optimization Methods with pytorch optim

geoopt

Manifold aware pytorch.optim.

Unofficial implementation for “Riemannian Adaptive Optimization Methods” ICLR2019 and more.

Installation

Make sure you have pytorch>=1.1.0 installed

There are two ways to install geoopt:

1. GitHub (preferred so far) due to active development

pip install git+https://github.com/geoopt/geoopt.git

2. pypi (this might be significantly behind master branch)

pip install geoopt

The preferred way to install geoopt will change once stable project stage is achieved. Now, pypi is behind master as we actively develop and implement new features.

What is done so far

Work is in progress but you can already use this. Note that API might change in future releases.

Tensors

• geoopt.ManifoldTensor – just as torch.Tensor with additional manifold keyword argument.
• geoopt.ManifoldParameter – same as above, recognized in torch.nn.Module.parameters as correctly subclassed.

All above containers have special methods to work with them as with points on a certain manifold

• .proj_() – inplace projection on the manifold.
• .proju(u) – project vector u on the tangent space. You need to project all vectors for all methods below.
• .egrad2rgrad(u) – project gradient u on Riemannian manifold
• .inner(u, v=None) – inner product at this point for two tangent vectors at this point. The passed vectors are not projected, they are assumed to be already projected.
• .retr(u, t=1.) – retraction map following vector u for time t
• .expmap(u, t=1.) – exponential map following vector u for time t (if expmap is not available in closed form, best approximation is used)
• .transp(v, *more, u, t=1.) – transport vector v (and possibly more vectors) with direction u for time t
• .retr_transp(v, *more, u, t=1.) – transport self, vector v (and possibly more vectors) with direction u for time t (returns are plain tensors)

Manifolds

• geoopt.Euclidean – unconstrained manifold in R with Euclidean metric
• geoopt.Stiefel – Stiefel manifold on matrices A in R^{n x p} : A^t A=I, n >= p
• geoopt.Sphere - Sphere manifold ||x||=1
• geoopt.PoincareBall - Poincare ball model (wiki)

All manifolds implement methods necessary to manipulate tensors on manifolds and tangent vectors to be used in general purpose. See more in documentation.

Optimizers

• geoopt.optim.RiemannianSGD – a subclass of torch.optim.SGD with the same API
• geoopt.optim.RiemannianAdam – a subclass of torch.optim.Adam

Samplers

• geoopt.samplers.RSGLD – Riemannian Stochastic Gradient Langevin Dynamics
• geoopt.samplers.RHMC – Riemannian Hamiltonian Monte-Carlo
• geoopt.samplers.SGRHMC – Stochastic Gradient Riemannian Hamiltonian Monte-Carlo