-
Hello it is unfortunately not possible to use automatic differentation with (at least) the `MvNormal` distribution.
The following code will fail at `rand(p)` due to a wrong conversion to `Float64`
`…
-
Coupling layers are not suitable for all parts of our causal flow model because the outcome variable is always one dimensional. Therefore we need a type of normalising flow architecture that is suitab…
-
Hi Smion,
I am trying to build a baseline system (GRU model) following your paper, my initial implementation generated pose which is very dynamic (arms move quickly in large scale in my upper body mo…
-
To make an actual contribution I have to design a new component of a Normalising Flow architecture that makes it a 'causal flow'. To do that I have identify an issue with the simple extension of the C…
-
tasks
- [x] Affine forward - 10 points
- [x] Stacking logprob - 20 points
-
Done when
- [x] Affine backward - 10 pts
- [x] Stacking rprob - 20 pts
-
Hi,
I'm seeing an unreasonable increase in RAM usage (of order GB) when training a normalising flow with batch norm between layers. I believe this is caused by the computation graph being extended …
-
https://github.com/asciidoctor/asciidoctor-diagram/issues/292 pointed to a rather bad regression in 2.0.3 when rerendering diagrams. I've pulled 2.0.3 from rubygems.org already. I think we should do t…
-
The first idea is to keep the CEVAE as is and augment the variational distribution q(z|x,t,y) by parameterising it as a normalising flow. This can be done with a simple flow such as planar flow or rad…
-
It would be nice to have a pSGLD optimizer as in [Le et al. (2015)](https://arxiv.org/abs/1512.07666). This implementation of MCMC is a nice drop-in replacement for a PyTorch optimizer (e.g. for Adam)…