DiffEqFlux Blog Post

[ChrisRackauckas]

Two discussion points from the private channels which are still un-addressed:

• diffeq_rd is explained long before its use. Example should probably come first?
• 100 epochs of training without data is admittedly odd.

[ChrisRackauckas]

Adding in the media next.

[MikeInnes]

Maybe it's better to just write out the loop rather than using train!. Being able to use that is kind of neat but probably less clear overall.

[ChrisRackauckas]

Some media:

[ChrisRackauckas]

Needs JuliaDiff/ForwardDiff.jl#378 and a ForwardDiff tag @jrevels

[KristofferC]

That PR still fails tests.

[ChrisRackauckas]

Yes, which is a great reason to ping for help.

[ChrisRackauckas]

SciML/DelayDiffEq.jl#93 breaks the layer because of the sizing issue. Temporary fix:

function delay_lotka_volterra(du,u,h,p,t)
  x, y = u
  α, β, δ, γ = p
  du[1] = dx = (α - β*y)*h(p,t-0.1)[1]
  du[2] = dy = (δ*x - γ)*y
end
h(p,t) = ones(eltype(p),2)
prob = DDEProblem(delay_lotka_volterra,[1.0,1.0],h,(0.0,10.0),constant_lags=[0.1])
p = param([2.2, 1.0, 2.0, 0.4])
params = Flux.Params([p])
reduction2(sol) = sol[1,1:101]
function predict_fd_dde()
  diffeq_fd(p,reduction2,101,prob,MethodOfSteps(Tsit5()),saveat=0.1)
end
loss_fd_dde() = sum(abs2,x-1 for x in predict_fd_dde())
loss_fd_dde()

[ChrisRackauckas]

[jessebett]

Some first comments.

[jessebett]

More suggestions, mostly typos

[ChrisRackauckas]

Here's a Gist with the code. It may make sense to link this in the post somewhere.

https://gist.github.com/ChrisRackauckas/89bb4181bb1c82266daeeec68c7f8d0e

[ChrisRackauckas]

Following @ViralBShah 's suggestion, we put the runnable scripts in the model-zoo repo and will link to them after each animation to make it easier to follow along with the post. @MikeInnes will get that merged in a sec, add the links here, and then we are ready to release. So let's say 10:30-11? Any last comments need to get in by then, or at least highlight a previous discussion.

[fredrikekre]

The title seems to be missing from the rendered version.

[ChrisRackauckas]

It displays the info table at the top differently than the others:

https://github.com/JuliaLang/www.julialang.org/blob/master/blog/_posts/2019-01-18-fluxdiffeq.md

https://github.com/JuliaLang/www.julialang.org/blob/master/blog/_posts/2018-12-03-ml-language-compiler.md

Anyone know why?

[MikeInnes]

I just fixed it, hopefully it'll show up soon.

[jekbradbury]

The LaTeX isn’t showing up for me in mobile Safari

I think you need to include MathJax at the top of the post, like this?

[Vaibhavdixit02]

The LateX isn't working in desktop Safari as well

[ChrisRackauckas]

49ba1b2 should fix it when the post updates

[ChrisRackauckas]

we should be good. Still taking any hotfixes as they come, but it's good to post now I think

[fredrikekre]

I think 49ba1b2 broke this:

Edit: No, it is just a missing \end{align}.

[AzamatB]

In some formulas derivative sign is not typeset correctly. For example:

[oxinabox]

Probably need to switch ' for ^\prime