Jamie Townsend

icon indicating a website link https://j-towns.github.io

icon company University of Amsterdam icon location Amsterdam, The Netherlands icon location 1 y/o post-doc in machine learning

Results 17 comments of Jamie Townsend

Switch to forward-over-reverse mode hvp for autograd backend

Hey @nkoep, I've [added](https://github.com/HIPS/autograd/pull/579) the JVP of `autograd.numpy.linalg.norm`. I'm pretty sure all of the differentiable elementwise `ufunc`s in numpy and scipy, and all of the `np.*` namespace has JVPs implemented,...

Switch to forward-over-reverse mode hvp for autograd backend

Btw if you want to submit a PR adding the 'forward over reverse' hvp implementation to [this file](https://github.com/HIPS/autograd/blob/2afc187aeb9d2d9d61117e7ffe8dc6d0f631171c/autograd/differential_operators.py) I'm happy to merge it into Autograd.

fixed-rank unitest breaks when k is equal to n

Hi, thanks for raising this. Can you post complete code to replicate the problem?

Implementation of a more generic fixed-rank manifold?

Hi there, thanks for the request. I agree we should have some options for generic low-rank matrices. We'll look into this and let you know when we have something implemented.

Implementation of a more generic fixed-rank manifold?

We're working on it now, nearly done. Should hopefully be done by the end of today. If not then tomorrow 🙂

Implementation of a more generic fixed-rank manifold?

the manifold we are currently implementing is from Manopt's fixedrankembeddedfactory.m

Implementation of a more generic fixed-rank manifold?

We've also written a simple example for using this manifold: https://github.com/pymanopt/pymanopt/blob/master/examples/low_rank_matrix_approximation.py The solvers return an array of type `_Point` for this manifold (something we will iron out when we have...

Implementation of a more generic fixed-rank manifold?

Reopening until we have sorted out Theano and Tensorflow backends.

Implementation of a more generic fixed-rank manifold?

https://github.com/pymanopt/pymanopt/commit/838cf87eec2b997fe3876c2c46ec8e95b48ee9d5 In the interests of simplicity and efficiency, we've decided that points on this manifold should be parameterised via their low rank svd. That is, they are parameterised similarly to...

Implementation of a more generic fixed-rank manifold?

There are two unimplemented methods which are necessary for those solvers. `ehess2rhess` is necessary for TrustRegions and `dist` is necessary for NelderMead and ParticleSwarm. To implement `ehess2rhess` someone needs to...