Tor Erlend Fjelde

`vi` should not work with Matrixvariate distributions

4

## Issue https://github.com/TuringLang/Bijectors.jl/issues/169 This happens because `Bijectors.Stacked` is only being compatible with 0- and 1-dimensional inputs. To "fix" this I've added a `Vec` bijector which wraps any higher-than-1-dimensional bijector with...

Zygote's compilation scales badly with the number of `~` statements

33

I'm not certain whether or not this can be considered a "Turing.jl-issue" or not, but I figured I would at least raise it as an issue here so people are...

After #1726 we can make use of alternative implementations of `AbstractVarInfo`, in particular immutable implementations. This comes with a lot of benefits, with a big one being performance improvements for...

Invalid test case for `Gibbs`?

2

I'm currently updating Turing.jl to be compatible with the new version of DynamicPPL.jl and ran into something that seems like a bug. Currently, we implement the logdensity-evaluation required for MH...

PG doesn't work multivariate distributions

7

This was discovered by @owensnick in https://discourse.julialang.org/t/avoiding-loops-in-turing/59830. ```julia n = 200 configs = [-820:180:-100, 280:180:1000] class = rand(1:2, n); pos = rand(1:5, n); x = [rand(Normal(configs[c][p], 5)) for (c, p)...

Ref: https://github.com/TuringLang/Turing.jl/pull/1567#issuecomment-817377941 This is where the failure happens: https://github.com/TuringLang/Turing.jl/blob/master/test/inference/mh.jl#L149-L152 In particular `v2` can sometimes be 0 (i.e. all samples are rejected).

Ref: https://github.com/TuringLang/Turing.jl/pull/1582 Currently, something like ```julia MH( :m => AdvancedMH.StaticProposal(Normal()), :s => AdvancedMH.RandomWalkProposal(Normal()) ) ``` won't lead to `s` being linked as one might except. The "fix" in https://github.com/TuringLang/Turing.jl/pull/1582 makes...

Allow control of sensitivity error through `solve`

1

It would be awesome if we could directly control the error on of the sensitivities through `solve`, e.g. `solve(...; adjoint_abstol=1e-6, adjoint_reltol=1e-6)`.

Methods for GPU not properly invalidated upon redefinition

1

It seems like methods aren't properly invalidated on GPU when redefined. Example (due to the async printing, it looks weird so you'll just have to take my comments as truth...

Issues with offset indices and GPU

5

Trying to implement a slightly weird "half-convolution" as follows: ```julia using NNlib, Tullio _tullio_conv(u, v) = @tullio y[i+_] := u[i+a] * v[a] function tullio_conv(u::AbstractVector, v::AbstractVector) # Left pad upad =...