[WIP] Add neural DAE support #159
Conversation
|
This would need a test to build on, @ChrisRackauckas can you point me to an example how that should like? |
src/neural_de.jl
Outdated
| @@ -167,3 +167,29 @@ function (n::NeuralCDDE)(x,p=n.p) | |||
| prob = DDEProblem{false}(dudt_,x,n.hist,n.tspan,p,constant_lags = n.lags) | |||
| concrete_solve(prob,n.solver,x,p,n.args...;sensealg=TrackerAdjoint(),n.kwargs...) | |||
| end | |||
|
|
|||
| struct NeuralDAE{P,M,RE,T,S,DV,A,K} <: NeuralDELayer | |||
There was a problem hiding this comment.
This does not cover the mass matrix form. That should be added as per https://docs.juliadiffeq.org/latest/tutorials/advanced_ode_example/#Handling-Mass-Matrices-1 and the discussion in the slack thread.
|
A mass matrix one would be more useful, but this also is just incorrect. How does one specify the constraint equations? The algebraic variables? For a test DAE, use something non-stiff. The ROBER equations for a short time are. function f(out,du,u,p,t)
out[1] = - 0.04u[1] + 1e4*u[2]*u[3] - du[1]
out[2] = + 0.04u[1] - 3e7*u[2]^2 - 1e4*u[2]*u[3] - du[2]
out[3] = u[1] + u[2] + u[3] - 1.0
end
u₀ = [1.0, 0, 0]
du₀ = [-0.04, 0.04, 0.0]
tspan = (0.0,10.0)
using OrdinaryDiffEq
differential_vars = [true,true,false]
prob = DAEProblem(f,du₀,u₀,tspan,differential_vars=differential_vars)
sol = solve(prob,DABDF2())which should be differentiable with Tracker. |
abhigupta768
force-pushed
the
neuraldae
branch
from
3a0393b
to
7c90a6a
Compare
| end | ||
| dudt_(u,p,t) = f | ||
| prob = DAEProblem(dudt_,du0,x,n.tspan,p,differential_vars=n.differential_vars) | ||
| concrete_solve(prob,n.solver,x,p,n.args...;n.kwargs...) |
There was a problem hiding this comment.
default to sensealg = TrackerAdjoint()
src/neural_de.jl
Outdated
|
|
||
| function (n::NeuralDAE)(x,du0=n.du0,p=n.p) | ||
| function f(u,p,t) | ||
| vcat([n.differential_vars[i] == 1 ? n.re(p)(u[i]) : n.constraints_model(u[i],p,t) for i in 1:length(n.differential_vars)]) |
There was a problem hiding this comment.
This isn't quite correct. n.constraints_model(u,p,t) should return a vector of constraints. n.re(p)(u) returns a vector. Then n.differential_vars[i] == true is a better check for decomposing / concatenting.
test/neural_dae.jl
Outdated
| @@ -0,0 +1,22 @@ | |||
| using Flux, DiffEqFlux, OrdinaryDiffEq | |||
There was a problem hiding this comment.
this test isn't called until it's added to runtests.jl
Codecov Report
@@ Coverage Diff @@
## master #159 +/- ##
==========================================
- Coverage 75.86% 68.13% -7.73%
==========================================
Files 4 4
Lines 174 204 +30
==========================================
+ Hits 132 139 +7
- Misses 42 65 +23
Continue to review full report at Codecov.
|
abhigupta768
force-pushed
the
neuraldae
branch
from
8459b12
to
d502d98
Compare
src/neural_de.jl
Outdated
| Flux.@functor NeuralDAE | ||
|
|
||
| function (n::NeuralDAE)(x,p=n.p) | ||
| function f(du,u,p,t) |
src/neural_de.jl
Outdated
| function f(du,u,p,t) | ||
| nn_out = n.re(p)(u) | ||
| alg_out = n.constraints_model(u,p,t) | ||
| du .= vcat(nn_out,alg_out) | ||
| nothing | ||
| end |
There was a problem hiding this comment.
| function f(du,u,p,t) | |
| nn_out = n.re(p)(u) | |
| alg_out = n.constraints_model(u,p,t) | |
| du .= vcat(nn_out,alg_out) | |
| nothing | |
| end | |
| function f(u,p,t) | |
| nn_out = n.re(p)(u) | |
| alg_out = n.constraints_model(u,p,t) | |
| vcat(nn_out,alg_out) | |
| end |
@ChrisRackauckas @Vaibhavdixit02