initial changes for neuraldsde and neuralsde

Author: ba2tripleO

docs/src/examples/neural_sde.md

@@ -76,28 +76,24 @@ neural SDE with diagonal noise layer function:
 drift_dudt = Lux.Chain(ActivationFunction(x -> x.^3),
                        Lux.Dense(2, 50, tanh),
                        Lux.Dense(50, 2))
-p1, st1 = Lux.setup(rng, drift_dudt)
 
 diffusion_dudt = Lux.Chain(Lux.Dense(2, 2))
-p2, st2 = Lux.setup(rng, diffusion_dudt)
 
-p1 = Lux.ComponentArray(p1)
-p2 = Lux.ComponentArray(p2)
-#Component Arrays doesn't provide a name to the first ComponentVector, only subsequent ones get a name for dereferencing
-p = [p1, p2]
 
 neuralsde = NeuralDSDE(drift_dudt, diffusion_dudt, tspan, SOSRI(),
                        saveat = tsteps, reltol = 1e-1, abstol = 1e-1)
+
+p, st = Lux.setup(rng, neuralsde)
 ```
 
 Let's see what that looks like:
 
 ```@example nsde
 # Get the prediction using the correct initial condition
-prediction0, st1, st2 = neuralsde(u0,p,st1,st2)
+prediction0, st = neuralsde(u0,p,st)
 
-drift_(u, p, t) = drift_dudt(u, p[1], st1)[1]
-diffusion_(u, p, t) = diffusion_dudt(u, p[2], st2)[1]
+drift_(u, p, t) = drift_dudt(u, p.p1, st.state1)[1]
+diffusion_(u, p, t) = diffusion_dudt(u, p.p2, st.state2)[1]
 
 prob_neuralsde = SDEProblem(drift_, diffusion_, u0,(0.0f0, 1.2f0), p)
 
@@ -119,7 +115,7 @@ the data values:
 
 ```@example nsde
 function predict_neuralsde(p, u = u0)
-  return Array(neuralsde(u, p, st1, st2)[1])
+  return Array(neuralsde(u, p, st)[1])
 end
 
 function loss_neuralsde(p; n = 100)

src/neural_de.jl

@@ -121,7 +121,8 @@ Arguments:
   documentation for more details.
 
 """
-struct NeuralDSDE{M,P,RE,M2,RE2,T,A,K} <: NeuralDELayer
+# struct NeuralDSDE{M,P,RE,M2,RE2,T,A,K} <: NeuralDELayer
+struct NeuralDSDE{M,P,RE,M2,RE2,T,A,K} <: Lux.AbstractExplicitLayer
     p::P
     len::Int
     model1::M
@@ -179,19 +180,31 @@ function (n::NeuralDSDE{M})(x,p=n.p) where {M<:FastChain}
     solve(prob,n.args...;sensealg=TrackerAdjoint(),n.kwargs...)
 end
 
-function (n::NeuralDSDE{M})(x,p,st1,st2) where {M<:Lux.AbstractExplicitLayer}
-    function dudt_(u,p,t)
-      u_, st1 = n.model1(u,p[1],st1)
+function initialparameters(rng::AbstractRNG, n::NeuralDSDE)
+    p1 = Lux.initialparameters(rng, n.model1)
+    p2 = Lux.initialparameters(rng, n.model2)
+    return Lux.ComponentArray((p1 = p1, p2 = p2))
+end
+
+function initialstates(rng::AbstractRNG, n::NeuralDSDE)
+    st1 = Lux.initialstates(rng, n.model1)
+    st2 = Lux.initialstates(rng, n.model2)
+    return (state1 = st1, state2 = st2)
+end
+
+function (n::NeuralDSDE{M})(x,p,st) where {M<:Lux.AbstractExplicitLayer}
+    function dudt_(u,p,t;st=st)
+      u_, st.state1 = n.model1(u,p.p1,st.state1)
       return u_
     end
-    function g(u,p,t)
-      u_, st2 = n.model2(u,p[2],st2)
+    function g(u,p,t;st=st)
+      u_, st.state2 = n.model2(u,p.p2,st.state2)
       return u_
     end
 
     ff = SDEFunction{false}(dudt_,g,tgrad=basic_tgrad)
     prob = SDEProblem{false}(ff,g,x,n.tspan,p)
-    return solve(prob,n.args...;sensealg=TrackerAdjoint(),n.kwargs...), st1, st2
+    return solve(prob,n.args...;sensealg=TrackerAdjoint(),n.kwargs...), st
 end
 
 """
@@ -251,6 +264,15 @@ struct NeuralSDE{P,M,RE,M2,RE2,T,A,K} <: NeuralDELayer
             typeof(tspan),typeof(args),typeof(kwargs)}(
             p,length(p1),model1,re1,model2,re2,tspan,nbrown,args,kwargs)
     end
+    
+    function NeuralSDE(model1::Lux.AbstractExplicitLayer, model2::Lux.AbstractExplicitLayer,tspan,nbrown,args...;
+                        p1 = nothing, p = nothing, kwargs...)
+        re1 = nothing
+        re2 = nothing
+        new{typeof(p),typeof(model1),typeof(re1),typeof(model2),typeof(re2),
+            typeof(tspan),typeof(args),typeof(kwargs)}(
+              p,Int(1),model1,re1,model2,re2,tspan,nbrown,args,kwargs)
+    end
 end
 
 function (n::NeuralSDE)(x,p=n.p)
@@ -269,6 +291,32 @@ function (n::NeuralSDE{P,M})(x,p=n.p) where {P,M<:FastChain}
     solve(prob,n.args...;sensealg=TrackerAdjoint(),n.kwargs...)
 end
 
+function initialparameters(rng::AbstractRNG, n::NeuralSDE)
+    p1 = initialparameters(rng, n.model1)
+    p2 = initialparameters(rng, n.model2)
+    return Lux.ComponentArray((p1 = p1, p2 = p2))
+end
+function initialstates(rng::AbstractRNG, n::NeuralSDE)
+    st1 = initialstates(rng, n.model1)
+    st2 = initialstates(rng, n.model2)
+    return (state1 = st1, state2 = st2)
+end
+
+function (n::NeuralSDE{P,M})(x,p,st) where {P,M<:Lux.AbstractExplicitLayer}
+    function dudt_(u,p,t;st=st)
+        u_, st.state1 = n.model1(u,p.p1,st.state1)
+        return u_
+    end
+    function g(u,p,t;st=st)
+        u_, st.state2 = n.model2(u,p.p2,st.state2)
+        return u_
+    end
+
+    ff = SDEFunction{false}(dudt_,g,tgrad=basic_tgrad)
+    prob = SDEProblem{false}(ff,g,x,n.tspan,p,noise_rate_prototype=zeros(Float32,length(x),n.nbrown))
+    solve(prob,n.args...;sensealg=ReverseDiffAdjoint(),n.kwargs...), st
+end
+
 """
 Constructs a neural delay differential equation (neural DDE) with constant
 delays.