ChrisRackauckas
commented
4 years ago The right thing to use is probably a closure. Here's an example of that:
using DiffEqFlux, OrdinaryDiffEq, Flux
dydt = FastChain(FastDense(3, 10), FastDense(10, 2))
dydtₚ = initial_params(dydt)
function Node(u, p, t, f_t)
dydt([u; f_t(t)], p)
end
function prediction(fₜ, θ)
prob = ODEProblem((u,p,t)->Node(u,p,t,fₜ), u0, tspan, θ)
concrete_solve(prob, Tsit5(), saveat = Δt)
end
solution = ones(2,11)
f_t = (t)->t;
u0 = ones(2)
Δt = 0.1
tspan = (0.0,1.0)
loss(dydtₚ) = sum(abs2, solution - prediction(f_t, dydtₚ))
res = DiffEqFlux.sciml_train(loss, dydtₚ, LBFGS())
solution = [ones(2,11) for i in 1:10]
signals = [(t)->-i*t for i in 1:10]
function loss_batch(θ)
n_exam = length(signals)
loss_val = 0
for i in 1:n_exam
guess = prediction(signals[i], θ)
loss_val += sum(abs2, solution[i] - guess)
end
loss_val
end
res = DiffEqFlux.sciml_train(loss_batch, dydtₚ, LBFGS())Let me know if you need anything else. Cheers.
junsebas97
I’m trying to use the DiffEqFlux package to define a machine learning model for an ODE. The model I need is a Neural ODE whose input is the initial condition and a continuos time signal.
Currently, i have been capable to train the NN with just one example with:
That works prefectly, but it uses just one signal for training. When i try to train in a batch of functions (using several signals) with:
I get this message:
I think that the bug is produced by the cycle, but i'm not sure