Test Errors? about neuralpde.jl HOT 19 CLOSED

@azev77 this happens simply because the name σ is already defined as NNlib.σ in the environment. Just change the name of the function σ (X, p, t) to something else, and it should work.
We will fix this soon.

from neuralpde.jl.

gives u(x=X0, t=tspan[1]).
How can I recover the entire function: u(x,t), u_x(x,t) etc ?

You are correct that this is what it gives. It's a very specific method for solving the terminal PDE for HJB and stochastic optimal control types of problems, and not one that can compute the full function. There are other methods which can compute the whole function (which are of course slower too) which will be added to the library.

from neuralpde.jl.

I manually downloaded the latest version w/
] add https://github.com/JuliaDiffEq/NeuralNetDiffEq.jl
and it works now.

from neuralpde.jl.

I wonder if SciML/DiffEqBase.jl#430 had an effect here. I'll test it sometime tomorrow.

from neuralpde.jl.

See if rolling back that patch fixes it.

from neuralpde.jl.

Sorry @ChrisRackauckas I don't know how to do that

from neuralpde.jl.

]add [email protected], and then ]free DiffEqBase when you're done to return it to normal.

from neuralpde.jl.

Tried, I get a new error in method definition at this line:

σ(X,p,t) = Diagonal(sigma*X) #Matrix d x d

from neuralpde.jl.

Are you not on Julia 1.3?

from neuralpde.jl.

Yes, I am using 1.3

from neuralpde.jl.

Thanks @KirillZubov @ChrisRackauckas
I'm preparing a blog post on Julia for dynamic economic models.
I believe currently

ans = solve(prob, pdealg, verbose=true, maxiters=200, trajectories=m,
                            alg=EM(), dt=dt, pabstol = 1f-6)

gives u(x=X0, t=tspan[1]).
How can I recover the entire function: u(x,t), u_x(x,t) etc ?

from neuralpde.jl.

Btw, here are other deep BSDE solvers w tf, pytorch :
https://github.com/frankhan91/DeepBSDE?files=1

https://github.com/skyhehe123/DeepBSDE-pytorch?files=1

https://github.com/vduarte/MLEcon

from neuralpde.jl.

Thanks! Is this all cleaned up? I noticed our tests are having some issues right now, but I think @KirillZubov is on it. One of the things we need to fix up is just how long to setup the CI tests for.

from neuralpde.jl.

Btw, I have NeuralNetDiffEq v1.0.1 installed, which is not the latest version.
For some reason it won't let me install the latest. Not sure if this is relevant.

from neuralpde.jl.

On Julia 1.3 with Flux 0.10?

from neuralpde.jl.

I reinstalled Julia today.
Julia 1.3.1
Flux 0.10.1

from neuralpde.jl.

Fyi, when I run the tests now I get a new error at:

ans = solve(prob, pdealg, verbose=true, maxiters=200, trajectories=m,
                            alg=EM(), dt=dt, pabstol = 1f-6)

UndefVarError: Tracker not defined
getproperty at Base.jl:13 [inlined]

from neuralpde.jl.

Interesting... Tracker doesn't exist in the repo at all.

from neuralpde.jl.

All tests are passing on master so tagging.

from neuralpde.jl.