search

nimble-dev / nimbleSMC

Sequential Monte Carlo methods for NIMBLE
2 stars 2 forks source link

column names in `mv{E,}WSamples` for filters with `saveAll = FALSE` are confusing #7

Open paciorek opened 2 years ago

paciorek commented 2 years ago

when saveAll = FALSE, only a single time step is saved, namely the last time step. However the column name in the mv objects will have [1]. E.g., if the variable is 'x', it will be x[1].

Given modelValues are designed to save entire variables, not clear there is an easy way to fix this.

paciorek commented 1 year ago

Note that if each element of x is multivariate, then we'd have the situation that the indexing relates to the elements of the multivariate vector and not to time. This would be fine, unless the model were defined on, say, x[time, 2:3], in which case the 2:3 would not carry over to the modelValues naming.

At least for the scalar case, we could get around this by instantiating the modelValues based on a symbolTable rather than {vars,types,sizes}, giving sizes as numeric(0), as is what happens when setting up modelValues for MCMC.

paciorek commented 1 year ago

However, this seems to require using nimble:::symbolTable, which is not exported...

paciorek commented 1 year ago

This was my attempt at this (in BootstrapFilter.R where mv{E,}Wsamples is created if saveAll=FALSE.

        for(i in seq_along(modelSymbolObjects))
            if(is.null(dim(   modelSymbolObjects[[i]]$size))) {
                modelSymbolObjects[[i]]$size <- numeric(0)
                modelSymbolObjects[[i]]$nDim  <- 0
                } else modelSymbolObjects[[i]]$size =  as.numeric(dims[[1]])
        symTab <- nimble:::symbolTable(symbols = modelSymbolObjects)
        mvEWSamples <- modelValues(modelValuesConf(symTab))
        symTab$addSymbol(nimble:::symbolBasic(name = 'wts',type='double',size=numeric(0), nDim = 0))
        mvWSamples <- modelValues(modelValuesConf(symTab))

however, internally we do some checking for inconsistent dimensions and this warning is triggered:

Confused in varAndInds2MapParts: nDim != varAndInds.indices.size()

presumably because of a mismatch of the dimension of the latent vector compared to the modelValues representation of it.

So I'm just going to leave things as they are.