Compiling error with LHSnodes

[Fanny Empacher]

Hi,

I'm quite new to nimble, but have tried to fit a few models and nimble really sped up computation time (a lot), so thank you!

Now I've run into an error that I don't understand when compiling:

Error in nimble:::nodeFunctionVector(model = model, nodeNames = LHSnodes,  : 

  In nodeFunctionVector from a case like model$doSomething(nodes[i]) where nodes may not contain well-defined node names.  Context is calculate(model = model, nodes = LHSnodes[i])

The uncompiled version runs fine (although too slow for any practical purposes). I'm hoping that this points to a very simple error that's easy to fix. If it helps I can also post my code, it's just a bit long.

[Daniel Turek]

Fanny, thanks for your use and kind words about nimble.

We could try to speculate all day, perhaps something about looping over a vector of node names (for(i in 1:numberOfNodes) .... something using nodeNames[i] ....), but without your code, and most importantly a reproducible example of the problem, it's really difficult to say.  I know you mentioned your code is long, but is it possible to provide an example demonstrating the problem?

Cheers,

Daniel

--
You received this message because you are subscribed to the Google Groups "nimble-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to nimble-users...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/nimble-users/cb0189f0-6a46-4cfd-bc49-808c065ad99cn%40googlegroups.com.

[Fanny Empacher]

Hi Daniel,

Thanks for your quick reply! I can't spot anything where I'm looping over node names but I might be wrong. Here's my (almost unabridged) code:

totalT <- 5

theta_init <- c(0.3, 0.95, 0.9, 2.0e-04, 8, 120)

phi_p_init <- theta_init[1]

phi_a_init <- theta_init[2]

alpha_init <- theta_init[3]

beta_init <- theta_init[4]

rho_init <- theta_init[5]

tau_init <- theta_init[6]

y0 <- 500

y <-  c(668,649,834,747,903)

code <- nimbleCode({

  ##### Priors 

  phi_p ~ dbeta(2.87, 1.78)

  phi_a ~ dbeta(1.6, 1.2)

  alpha ~ dbeta(2, 1.5)

  beta ~ dgamma(2, 1/(3.7e-04/2))

  rho ~ dgamma(4, 1/2.5)

  tau ~ dgamma(2.1, 1/66.67)

  ##### Initialisation

  x_upd[, ] <- matrix(nrow = totalT+1, ncol = 2)

  x_pred[, ] <- matrix(nrow = totalT, ncol = 2)

  

  x0[1:2] <- y0*c(1, 1/alpha)

  sigma2 <- y0^2/tau

  P011 <- sigma2

  P022 <- 1/(alpha^2)*((1-alpha)*y0+sigma2)

  P012 <- 1/alpha*sigma2

  P021 <- 1/alpha*sigma2

  T12 <- alpha*phi_a

  T22 <- phi_a

  x_upd[1, 1:2] <- x0[1:2]

  P_upd11[1] <- P011

  P_upd12[1] <- P012

  P_upd21[1] <- P021

  P_upd22[1] <- P022

  for(t in 1:totalT){

    real_phi_p[t] <- phi_p/(1+(beta*x_upd[t, 1])^rho)

    T21[t] <- 0.5*real_phi_p[t]

    T11[t] <-  alpha*0.5*real_phi_p[t]

    Q11[t] <- alpha*(1-alpha)*(0.5*real_phi_p[t]*x_upd[t, 1]+phi_a*x_upd[t, 2])

    Q22[t] <- 0.5*real_phi_p[t]*(1-0.5*real_phi_p[t])*x_upd[t, 1] + phi_a*(1-phi_a)*x_upd[t, 2]

    Q12[t] <- alpha*(0.5*real_phi_p[t]*(1-0.5*real_phi_p[t])*x_upd[t, 1] + phi_a*(1-phi_a)*x_upd[t, 2])

    Q21[t] <- alpha*(0.5*real_phi_p[t]*(1-0.5*real_phi_p[t])*x_upd[t, 1] + phi_a*(1-phi_a)*x_upd[t, 2])

    R_matrix[t] <- x_upd[t, 1]^2/tau

    # Prediction

    x_pred[t, 1] <- T11[t]*x_upd[t, 1] + T12*x_upd[t, 2]

    x_pred[t, 2] <- T21[t]*x_upd[t, 1] + T22*x_upd[t, 2]

    P_pred11[t] <- (T11[t]*P_upd11[t]+T12*P_upd21[t])*T11[t]+(T11[t]*P_upd12[t]+T12*P_upd22[t])*T12+Q11[t]

    P_pred12[t] <- (T11[t]*P_upd11[t]+T12*P_upd21[t])*T21[t]+(T11[t]*P_upd12[t]+T12*P_upd22[t])*T22+Q12[t]

    P_pred21[t] <- (T21[t]*P_upd11[t]+T22*P_upd21[t])*T11[t]+(T21[t]*P_upd12[t]+T22*P_upd22[t])*T12+Q21[t]

    P_pred22[t] <- (T21[t]*P_upd11[t]+T22*P_upd21[t])*T21[t]+(T21[t]*P_upd12[t]+T22*P_upd22[t])*T22+Q22[t]

    # Aid variables

    y_pred[t] <- x_pred[t, 1]

    v_pred[t] <- y[t] - y_pred[t]

    F_matrix[t] <-P_pred11[t] + R_matrix[t]

    G1[t] <- P_pred11[t]

    G2[t] <- P_pred21[t]

    F_inv[t] <- F_matrix[t]^(-1)

    K1[t] <- G1[t]*F_inv[t]

    K2[t] <- G2[t]*F_inv[t]

    # Distribution of y

    y[t] ~ dnorm(y_pred[t], var = F_matrix[t])

    # Update

    x_upd[t+1, 1] <- x_pred[t, 1] + K1[t]*v_pred[t]

    x_upd[t+1, 2] <- x_pred[t, 2] + K2[t]*v_pred[t]

    P_upd11[t+1] <- P_pred11[t] - F_inv[t]*G1[t]^2

    P_upd12[t+1] <- P_pred12[t] - F_inv[t]*G1[t]*G2[t]

    P_upd21[t+1] <- P_pred21[t] - F_inv[t]*G1[t]*G2[t]

    P_upd22[t+1] <- P_pred22[t] - F_inv[t]*G2[t]^2

  }

})

mNimble_seal2 <- nimbleModel(code,

                             dimensions = list(x_upd = c(totalT+1,2),

                                               x_pred = c(totalT,2)),

                             data = list(y=y, y0 = y0),

                             constants = list(totalT = totalT),

                             inits = list(phi_p = phi_p_init,

                                          phi_a = phi_a_init,

                                          alpha = alpha_init,

                                          beta = beta_init,

                                          rho = rho_init,

                                          tau = tau_init),

                             check = FALSE)

compNimble_seal2 <- compileNimble(mNimble_seal2)

confNimble_seal2 <- configureMCMC(mNimble_seal2,

                                  print = TRUE, useConjugacy = FALSE)

mcmcNimble_seal2 <- buildMCMC(confNimble_seal2)

# This runs but slowly

# fNimble_seal2 <- runMCMC(mcmcNimble_seal2,

#                          nburnin = 10,

#                          nchains = 2,

#                          niter = 110)

# This is where the error happens

mcmccNimble_seal2 <- compileNimble(mcmcNimble_seal2,

                                   project = mNimble_seal2,

                                   resetFunctions = TRUE,

                                   showCompilerOutput = TRUE)

[Daniel Turek]

Fanny, the issue is the inclusion of the matrix() declaration lines in your model code.  These are not necessary, and in fact are causing an issue down the line.  You should *not* include these two lines:

    x_upd[, ] <- matrix(nrow = totalT+1, ncol = 2)
    x_pred[, ] <- matrix(nrow = totalT, ncol = 2)

More troublesome, from the package development perspective, is how/why this slipped through model checking.  This should have been caught at the time of model building, and ideally issue a helpful message to correct the problem.

We'll look into that, improving the error-catching for this case, but for right now this should get your model and MCMC running.

Cheers,

Daniel

To view this discussion on the web visit https://groups.google.com/d/msgid/nimble-users/58e5be91-ad98-4d73-bcc7-dadcd78bb8fbn%40googlegroups.com.

[Fanny Empacher]

Thank you, that fixed the problem!