Lotka volterra UDE not recovered as per paper

ChrisRackauckas / universal_differential_equations

Repository for the Universal Differential Equations for Scientific Machine Learning paper, describing a computational basis for high performance SciML

https://arxiv.org/abs/2001.04385

MIT License

215 stars 59 forks source link

Lotka volterra UDE not recovered as per paper #39

Open yewalenikhil65 opened 2 years ago

yewalenikhil65 commented 2 years ago

scenario_2.jl outputs in following fit for UDE.. scenario_1.jl also does does not fit (I remember in early versions, this examples were able to recover the model accurately.. what changed ?)

AlCap23 commented 2 years ago

On it.

AlCap23 commented 2 years ago

Could you provide your ]st for reference?

Also, there is a specific tag for the paper version, which might work better.

yewalenikhil65 commented 2 years ago

Which tag ?

(@v1.7) pkg> st
      Status `~/.julia/environments/v1.7/Project.toml`
  [c52e3926] Atom v0.12.36
  [6e4b80f9] BenchmarkTools v1.3.1
  [052768ef] CUDA v3.8.5
  [479239e8] Catalyst v10.8.0
  [f65535da] Convex v0.15.1
  [2445eb08] DataDrivenDiffEq v0.8.1
  [aae7a2af] DiffEqFlux v1.45.3
  [c894b116] DiffEqJump v8.3.0
  [41bf760c] DiffEqSensitivity v6.71.0
  [0c46a032] DifferentialEquations v7.1.0
  [587475ba] Flux v0.12.9
  [f6369f11] ForwardDiff v0.10.25
  [86223c79] Graphs v1.6.0
  [e5e0dc1b] Juno v0.8.4
  [bdcacae8] LoopVectorization v0.12.103
  [961ee093] ModelingToolkit v8.5.5
  [429524aa] Optim v1.6.2
  [1dea7af3] OrdinaryDiffEq v6.7.1
  [91a5bcdd] Plots v1.27.3
  [b4db0fb7] ReactionNetworkImporters v0.13.2
  [c946c3f1] SCS v1.1.1
  [e88e6eb3] Zygote v0.6.37

AlCap23 commented 2 years ago

https://github.com/ChrisRackauckas/universal_differential_equations/releases/tag/v1.0

yewalenikhil65 commented 2 years ago

https://github.com/ChrisRackauckas/universal_differential_equations/releases/tag/v1.0

Doesnt look like i can precompile after instantiate this

AlCap23 commented 2 years ago

Yeah, there is definitely something 🐟 -y going on here. I experienced a full out fail on 1.7 while creating the ContinuousDataDrivenProblem, which should not happen.

Additionally, the network performs really bad.

I will retry on 1.6 and report a little more over the course of the day.