nimble-dev / nimbleecology Goto Github PK

We'll need to address this warning that arises for dCJS:

[Note] Detected use of function(s) that are not supported for derivative tracking in a function or method for which `buildDerivs` has been requested: is.na.

This creates an inconsistency where inputs that work fine for probability calculation errors during simulation, but there isn't a good reason simulation should behave differently.

I suggest we change the condition in dCJS to match that in rCJS. This is stricter and there's a risk it'd break something someone is doing, although I'd argue that it's too easy to misinterpret what's going on without this.

Not sure if this is just duplication or there or a reason but dDHMM and dDHMMo are registered in both zzz.R (i.e., via .onAttach) and in dDHMM.R. Perhaps the registration in the latter is needed to build the package?

In dHMM and other distributions, there is a check of whether sum(init) != 1, with an error if true.

Even if the init vector satisfies sum(init) == 1 in R, this condition can be false in C++, presumably due to floating point arithmetic issues, with possibly different orders in R vs C++ or something like that.

In such a case, which was reported by a user, it can be hard to work around. One can normalize the vector in R and still have sum(init) != 1 in C++.

How about replacing this condition with something like abs(sum(init) - 1) < 1e-6 or perhaps a stricter tolerance? This is similar to the other checks in nearby code that allow some numerical imprecision.

Once PR 1124 on nimble itself is merged in, we should be able to remove the need to set the checkNimbleFunction option to FALSE.

Hei! The example at https://rdrr.io/github/nimble-dev/nimbleEcology/man/dHMM.html throws an error: "Error in genVarInfo3() :
dimensions specified are smaller than model specification for variable 'probObs'"

Cheers
Matt

len <- 5 # length of dataset
dat <- c(1,2,1,1,2) # A vector of observations
init <- c(0.4, 0.2, 0.4) # A vector of initial state probabilities
probObs <- t(array( # A matrix of observation probabilities
  c(1, 0,
    0, 1,
    0.2, 0.8), c(2, 3)))
probTrans <- t(array( # A matrix of transition probabilities
  c(0.6, 0.3, 0.1,
    0, 0.7, 0.3,
    0, 0, 1), c(3,3)))

# Define code for a nimbleModel
nc <- nimbleCode({
  x[1:5] ~ dHMM(init[1:3], probObs = probObs[1:2,1:3],
                probTrans = probTrans[1:3, 1:3], len = 5)
  
  for (i in 1:3) {
    init[i] ~ dunif(0,1)
    
    for (j in 1:3) {
      probTrans[i,j] ~ dunif(0,1)
    }
    
    probObs[i, 1] ~ dunif(0,1)
    probObs[i, 2] <- 1 - probObs[1,i]
  }
})

# Build the model
HMM_model <- nimbleModel(nc,
                         data = list(x = dat),
                         inits = list(init = init,
                                      probObs = probObs,
                                      probTrans = probTrans))
#Error here
#defining model...
#Error in genVarInfo3() : 
#  dimensions specified are smaller than model specification for variable 'probObs'

# Calculate log probability of data from the model
HMM_model$calculate()
# Use the model for a variety of other purposes...