Where communities thrive


  • Join over 1.5M+ people
  • Join over 100K+ communities
  • Free without limits
  • Create your own community
People
Activity
  • 04:12
    ChrisRackauckas commented #181
  • 03:56
    ChrisRackauckas commented #182
  • 03:56
    ChrisRackauckas edited #182
  • 03:30
    ChrisRackauckas commented #182
  • 03:30
    ChrisRackauckas commented #182
  • 03:13
    danielchen26 commented #509
  • 03:12
    ChrisRackauckas commented #509
  • 03:10
    danielchen26 commented #509
  • 03:02
    ChrisRackauckas commented #509
  • 02:57
    danielchen26 commented #509
  • 02:55
    ChrisRackauckas commented #509
  • 02:54
    ChrisRackauckas edited #508
  • 02:53
    ChrisRackauckas commented #508
  • 02:48
    danielchen26 commented #509
  • 02:47
    JuliaRegistrator commented #73
  • 02:47
    ChrisRackauckas commented #73
  • 02:47

    ChrisRackauckas on master

    Update Project.toml (compare)

  • 02:47

    ChrisRackauckas on nondiag

    (compare)

  • 02:46

    ChrisRackauckas on master

    fix non-diagonal noise variable… Merge pull request #76 from Jul… (compare)

  • 02:46
    ChrisRackauckas closed #76
BridgingBot
@GitterIRCbot
[slack] <mason.protter> I know that if there was better support for linear PDEs I'd have an easier time showing off julia to curious undergrads in my department. 🙂
BridgingBot
@GitterIRCbot
[slack] <chrisrackauckas> Yeah we're getting there. More of a focus on splitting methods than before, which ups the emphasis of linear when PDEs split out linear parts
BridgingBot
@GitterIRCbot
[slack] <mason.protter> Yeah, I get that you're spinning many plates!
BridgingBot
@GitterIRCbot
[slack] <Antoine Levitt> Well if you have a simple 1d problem I'd say don't bother, just use central finite differences with an explicit solver
[slack] <chrisrackauckas> depends, 1D can be hard depending on the nonlinearities
[slack] <Antoine Levitt> I thinks he just wants linear Schrödinger
[slack] <Antoine Levitt> 1d is pretty forgiving, if you just want to plot something
[slack] <Antoine Levitt> For more advanced methods I'd do a splitting with an FFT to solve the kinetic term
BridgingBot
@GitterIRCbot
[slack] <david.plankensteiner> @mason.protter Are you looking to do something like this: https://qojulia.org/documentation/examples/particle-into-barrier.html ?
BridgingBot
@GitterIRCbot
[slack] <stefanos.carlstrom> I told him to do a simple Crank–Nicolson
BridgingBot
@GitterIRCbot
[slack] <chrisrackauckas> @jessebett @Guillaume Ausset now that CNF is a thing, have you looked at FFJORD?
[slack] <chrisrackauckas> I wonder if a small change of the working CNF would give a working FFJORd
BridgingBot
@GitterIRCbot
[slack] <Guillaume Ausset> I didn't have time to look at it but I could when I have time. Want to look at why GPU is broken first
BridgingBot
@GitterIRCbot
This message was deleted
BridgingBot
@GitterIRCbot

[slack] <pihop> Hi, I ran into a problem with the coupling of VariableRateJump with SDEs again. The following mwe

using DifferentialEquations

function ff(du,u,p,t)
    if p == 0
        du .= 1.01u
    else
        du .= 2.01u
    end
end

function gg(du,u,p,t)
  du[1,1] = 0.1u[1]
  du[1,2] = 0.1u[1]
  du[2,1] = 0.2u[1]
  du[2,2] = 0.2u[2]
end

rate_switch(u,p,t) = u[1] 

function affect_switch!(integrator)
    integrator.p = 1 
end

jump_switch = VariableRateJump(rate_switch,affect_switch!)
prob = SDEProblem(ff,gg,ones(2),(0.0,1.0),0,noise_rate_prototype=zeros(2,2))
jump_prob = JumpProblem(prob, Direct(), jump_switch)
sol = solve(EnsembleProblem(jump_prob), EM(), dt = 0.01; trajectories=10)
timepoint_meanvar(sol, 0.1)

gives an error MethodError: no method matching __broadcast when calling the timepoint_meanvar. Evaluating it on the boundaries t=0 and t=1 works. The error has the cascade effect of breaking EnsembleSummary when VariableRateJumps are involved. Any ideas what is going wrong? Should this be reported to DiffEqBase or somewhere else?

BridgingBot
@GitterIRCbot
[slack] <chrisrackauckas> Make an issue here to DifferentialEquations.jl
[slack] <chrisrackauckas> this is... oh god this one might be more difficult 🙂
BridgingBot
@GitterIRCbot
[slack] <pihop> Fair enough 😄 I've opened an issue. Do you have any idea for a quick and dirty workaround that I could try for extracting the mean and variance statistics?
[slack] <chrisrackauckas> Turn it into arrays first
[slack] <chrisrackauckas> the problem is that when you are doing variable rate jumps they are arrays with a bunch of other hidden things.
[slack] <chrisrackauckas> If you just loop it would be fine.
BridgingBot
@GitterIRCbot
[slack] <chrisrackauckas> broadcast has to handle the weird things.
[slack] <pihop> I see. Makes sense, thanks 🙂
BridgingBot
@GitterIRCbot
[slack] <adam.jozefiak> @chrisrackauckas I'm fixing up the (pure) upwind operators for DiffEqOperators and I just have a question with regards to stencil lengths with respect to derivative order and approximation.
1) For interior stencils, is the the stencil length simply: derivative_order + approximation_order ?
2) For boundary stencils, such as the case where we have a forward difference near the upper boundary and it is clear that a pure upwind will not fit, what should the stencil length be? Is it still derivative_order + approximatio_order?
BridgingBot
@GitterIRCbot

[slack] <adam.jozefiak> For instance, I obtain the concretization of the second order approximation of the first derivative operator on a three interior node grid, with positive coefficients as:

[ 0.0, -1.5, 2.0, -0.5, 0.0 ]
[ 0.0, 0.0, -1.5, 2.0, -0.5 ]
[ 0.0, 0.0, -0.5, 0.0, 0.5 ]

BridgingBot
@GitterIRCbot
[slack] <chrisrackauckas> that looks possible
[slack] <chrisrackauckas> I am not sure.
[slack] <chrisrackauckas> but looks like it could work
BridgingBot
@GitterIRCbot
[slack] <xtalax> @chrisrackauckas Is the fallback definition at line 5 DiffEqBase/init.jl really needed? it's breaking \for GhostDerivativeOperator
BridgingBot
@GitterIRCbot
[slack] <chrisrackauckas> yeah, I forget for what though.
[slack] <chrisrackauckas> There's a dispatch of ldiv! on a QR that is broken on 1.1 IIRC
Ivan Borisov
@ivborissov
Hi, what is a best way to calculate gradient for a custom loss_func based on ode solution and data? Currently I do it with Calculus.finite_difference!(loss_func,x,grad,:central)
BridgingBot
@GitterIRCbot
[slack] <xtalax> it poaches ldiv!(b::Array{T,1}, A::SparseMatrix, x::Array{T,1}, causing a method error
BridgingBot
@GitterIRCbot
[slack] <chen.tianc> I am trying to use Parallel Ensemble simulation for SDE using Gillespie. But I got error julia> solve(ensemble_prob,SSAStepper(),trajectories=10) ERROR: Inappropiate solve command. The arguments do not make sense. Likely, you gave an algorithm which does not actually exist (or does not <:DiffEqBase.DEAlgorithm). if I want to randomize both initial condition and parameters, how should I write prob_func?
BridgingBot
@GitterIRCbot
[slack] <chrisrackauckas> What happens if you add in the ensemble alg?
BridgingBot
@GitterIRCbot
[slack] <chen.tianc> what do you mean add in ensemblealg? I saw sim = solve(prob,alg,ensemblealg,kwargs...) on the web and ensemblealg is optional. Also, ensembelalg only take keyword arguments (trajectories, batch_size, pmap_batch_size)
BridgingBot
@GitterIRCbot
[slack] <chrisrackauckas> yes, pass it
[slack] <chrisrackauckas> see if you are fine with that
BridgingBot
@GitterIRCbot
[slack] <chen.tianc> ERROR: UndefVarError: ensemblealg not defined
Stacktrace:
[1] top-level scope at none:0
[slack] <chen.tianc> I just tried for ode case. Am I doing something wrong?
BridgingBot
@GitterIRCbot
[slack] <chrisrackauckas> no, do it like shown in the docs
[slack] <chrisrackauckas> EnsembleSerial()
[slack] <chen.tianc> it show ERROR: type JumpProblem has no field u0 using the first prob_func and ERROR: type JumpProblem has no field f using the second prob_func
BridgingBot
@GitterIRCbot
[slack] <chen.tianc> For ODE it works, but for jump problem Gillespie it does not
[slack] <chrisrackauckas> okay, open an issue with this MWE
[slack] <chrisrackauckas> I'll get to it after some Pumas stuff.
[slack] <chen.tianc> ok, thanks