jondeuce

I'm also running into this issue. Would be good to get this merged, if it looks good to @MikeInnes

Thanks for the link, that's an interesting approach. And good point re: often not having access to model construction, I certainly agree that having a way to freeze layers post-construction...

> To permanently freeze something, I think it should be enough just to exclude all fields from `trainable` This was what I tried initially. It actually fails the tests in...

@mcabbott I actually haven't thought about whether this would be faster. I imagine you are correct that it wouldn't be with correct optimizations. In fact, probably the various partials are...

> * Higher flexibility on combining field dimensions with geometry dimensions, e.g. 3D field on a 2D geometry. Hi all, I've been using JuAFEM in my work for quite some...

Thanks for the quick reply! And no kidding, that's a happy coincidence, I'm glad it's in the works. My work generally revolves around solving the [Bloch-Torrey equation](https://en.wikipedia.org/wiki/Diffusion_MRI#Magnetization_dynamics) for MRI simulations,...

Also, this does seem to work for other factorizations, e.g. `lu`: ```julia julia> B = lu(sparse(Matrix(I,10,10))); julia> isapprox(B\@views(z[:]), z) true ``` My julia version info is following, in case it's...

They are equivalent by definition: >training throughput = (# examples processed per second) The right hand side is equal to `batch size / time per step`. Rearranging this equation gives:...

Just ran into this same issue. For the case of `Uniform`, one issue stems from [this line](https://github.com/JuliaStats/Distributions.jl/blob/master/src/univariate/continuous/uniform.jl#L154) which should use `rand(rng, T)` instead of `rand(rng)`. Another issue stems from the...

Same issue on Ubuntu 18.04.4